AI in Medical Imaging and Image Processing

Topic Information

Dear Colleagues,

In modern healthcare, the importance of computer-aided diagnosis is quickly becoming obvious, with clear benefits for the medical professionals and patients. Automatization of processes traditionally maintained by human professionals is also growing in importance. The process of image analysis can be supported by the use of networks that can carry out multilayer analyses of patterns—collectively called artificial intelligence (AI). If supported by large datasets of input data, computer networks can suggest the result with low error bias. Medical imaging focused on pattern detection is typically supported by AI algorithms. AI can be used as an important aid in three major steps of decision making in the medical imaging workflow: detection (image segmentation), recognition (assignment to the class), and result description (transformation of the result to natural language). The implementation of AI algorithms may participate in the diagnostic process standardization and markedly reduces the time needed to achieve pathology detection and description of the results. With AI support, medical specialists may work more effectively, which can improve healthcare quality. As AI has been a topic of interest for a while now, there are many approaches to and techniques for the implementation of AI based on different computing methods designed to work in various systems. The aim of this Special Issue in to present the current knowledge dedicated to the AI methods used in medical systems, with their applications in different fields of diagnostic imaging. Our goal is for this collection of works to contribute to the exchange of knowledge resulting in a better understanding of AI technical aspects and applications in modern radiology.

Dr. Karolina Nurzynska
Prof. Dr. Michał Strzelecki
Prof. Dr. Adam Piorkowski
Dr. Rafał Obuchowicz
Topic Editors

Keywords

Participating Journals

Journal Name	Impact Factor	CiteScore	Launched Year	First Decision (median)	APC
BioMed biomed	-	-	2021	20.3 Days	CHF 1000	Submit
Cancers cancers	4.5	8.0	2009	16.3 Days	CHF 2900	Submit
Diagnostics diagnostics	3.0	4.7	2011	20.5 Days	CHF 2600	Submit
Journal of Clinical Medicine jcm	3.0	5.7	2012	17.3 Days	CHF 2600	Submit
Tomography tomography	2.2	2.7	2015	23.9 Days	CHF 2400	Submit

MDPI Topics is cooperating with Preprints.org and has built a direct connection between MDPI journals and Preprints.org. Authors are encouraged to enjoy the benefits by posting a preprint at Preprints.org prior to publication:

1. Immediately share your ideas ahead of publication and establish your research priority;
2. Protect your idea from being stolen with this time-stamped preprint article;
3. Enhance the exposure and impact of your research;
4. Receive feedback from your peers in advance;
5. Have it indexed in Web of Science (Preprint Citation Index), Google Scholar, Crossref, SHARE, PrePubMed, Scilit and Europe PMC.

Related Topic

Published Papers (40 papers)

Order results

Content type Publication Date

Result details

Normal Extended Compact

Journals

All BioMed Cancers Diagnostics JCM Tomography

Show export options expand_more Show export options expand_less

Select all

Export citation of selected articles as:

Plain Text BibTeX BibTeX (without abstracts) Endnote Endnote (without abstracts) Tab-delimited RIS

Error

Oops... you haven't selected anything for export.

Ok

clear

20 pages, 3823 KiB  

Open AccessArticle

Pulmonary Fissure Segmentation in CT Images Using Image Filtering and Machine Learning

by Mikhail Fufin, Vladimir Makarov, Vadim I. Alfimov, Vladislav V. Ananev and Anna Ananeva

Tomography 2024, 10(10), 1645-1664; https://doi.org/10.3390/tomography10100121 - 9 Oct 2024

Viewed by 287

Abstract

Background: Both lung lobe segmentation and lung fissure segmentation are useful in the clinical diagnosis and evaluation of lung disease. It is often of clinical interest to quantify each lobe separately because many diseases are associated with specific lobes. Fissure segmentation is important [...] Read more.

Background: Both lung lobe segmentation and lung fissure segmentation are useful in the clinical diagnosis and evaluation of lung disease. It is often of clinical interest to quantify each lobe separately because many diseases are associated with specific lobes. Fissure segmentation is important for a significant proportion of lung lobe segmentation methods, as well as for assessing fissure completeness, since there is an increasing requirement for the quantification of fissure integrity. Methods: We propose a method for the fully automatic segmentation of pulmonary fissures on lung computed tomography (CT) based on U-Net and PAN models using a Derivative of Stick (DoS) filter for data preprocessing. Model ensembling is also used to improve prediction accuracy. Results: Our method achieved an F1 score of $0.916$ for right-lung fissures and $0.933$ for left-lung fissures, which are significantly higher than the standalone DoS results ($0.724$ and $0.666$, respectively). We also performed lung lobe segmentation using fissure segmentation. The lobe segmentation algorithm shows results close to those of state-of-the-art methods, with an average Dice score of $0.989$. Conclusions: The proposed method segments pulmonary fissures efficiently and have low memory requirements, which makes it suitable for further research in this field involving rapid experimentation. Full article

(This article belongs to the Topic AI in Medical Imaging and Image Processing)

►▼ Show Figures

Figure 1

Figure 1
<p>The segmentation pipeline overview. First, the image is masked with the lung mask and then cropped to the mask’s bounding box. Next, the image is normalized and optionally filtered by a DoS filter. Then, individual slices are fed as input to the neural network. Finally, several inverse operations are performed to transform the output and retrieve the fissure segmentation. Green and red colors indicate voxel values that depend on the object class, and black is the background.</p> Full article ">Figure 2
<p>The combined preprocessing approach pipeline. Only one image channel contains the normalized image not processed by the filter.</p> Full article ">Figure 3
<p>Model ensembling pipeline. One-hot is applied to the output of each model, and then argmax is applied to the sum of the results. Here, black-and-white images represent the standard coding of classes, where for each class a particular value is used. Here, black and white images represent standard class encoding, where a specific value is used for each class. Color images represent one-hot class encoding, each color refers to one of the channels (blue—background, red and green—fissures).</p> Full article ">Figure 4
<p>Error types shown in the right lung segmentation example: (<b>a</b>) oversegmentation; (<b>b</b>) false negatives; (<b>c</b>) false segmentation; (<b>d</b>) misclassification. Green indicates the oblique fissure and yellow indicates the horizontal fissure.</p> Full article ">Figure 5
<p>A comparison of the ensemble of neural network models and the DoS method using one of the test images. The mask obtained with DoS is highlighted in purple. <math display="inline"><semantics> <mrow> <msub> <mi>F</mi> <mrow> <mi>D</mi> <mi>o</mi> <mi>S</mi> </mrow> </msub> <mo>=</mo> <mn>0.908</mn> </mrow> </semantics></math> for the DoS result, and <math display="inline"><semantics> <mrow> <msub> <mi>F</mi> <mrow> <mi>D</mi> <mi>o</mi> <mi>S</mi> </mrow> </msub> <mo>=</mo> <mn>0.967</mn> </mrow> </semantics></math> for the ensemble result (measured on this scan alone). (<b>a</b>) A 3D model. (<b>b</b>) A sagittal slice.</p> Full article ">Figure 6
<p>Plot for the right-lung models’ cross-validated performance on the oblique fissure.</p> Full article ">Figure 7
<p>Plot for the right-lung models’ cross-validated performance on the horizontal fissure.</p> Full article ">Figure 8
<p>Plot for the left-lung models’ cross-validated performance.</p> Full article ">Figure 9
<p>Right-lung lobes on one of the test images. (<b>a</b>) Lobes in the sagittal cross-section. Green, yellow and red colors are for the upper, middle, and red lobes, respectively. (<b>b</b>) A 3D model of lobes. (<b>c</b>) A 3D model of fissures. Green and yellow colors are for the oblique and horizontal fissures, respectively.</p> Full article ">Figure 10
<p>A comparison of true and predicted lobe masks of the right lung on one of the test images. Colors have the same meanings as in the <a href="#tomography-10-00121-f009" class="html-fig">Figure 9</a>, except for purple, which denotes airways. Here, the values of <math display="inline"><semantics> <msub> <mi>F</mi> <mn>1</mn> </msub> </semantics></math> are <math display="inline"><semantics> <mrow> <mn>0.992</mn> </mrow> </semantics></math> for the upper lobe; <math display="inline"><semantics> <mrow> <mn>0.952</mn> </mrow> </semantics></math> for the middle lobe; and <math display="inline"><semantics> <mrow> <mn>0.986</mn> </mrow> </semantics></math> for the lower lobe. (<b>a</b>) A region on a 3D lobe model on which the section corresponding to (<b>b</b>) is highlighted with red. (<b>b</b>) A sagittal slice with both masks superimposed. (<b>c</b>) The 3D lobe model with masks superimposed in the area of the right lung adjacent to the mediastinum; the areas of discrepancy are circled in red. (<b>d</b>) The same area highlighted on the 3D model.</p> Full article ">

13 pages, 2246 KiB  

Open AccessArticle

Opportunistic Screening for Low Bone Mineral Density in Adults with Cystic Fibrosis Using Low-Dose Computed Tomography of the Chest with Artificial Intelligence

by Matthias Welsner, Henning Navel, Rene Hosch, Peter Rathsmann, Florian Stehling, Annie Mathew, Sivagurunathan Sutharsan, Svenja Strassburg, Dirk Westhölter, Christian Taube, Sebastian Zensen, Benedikt M. Schaarschmidt, Michael Forsting, Felix Nensa, Mathias Holtkamp, Johannes Haubold, Luca Salhöfer and Marcel Opitz

J. Clin. Med. 2024, 13(19), 5961; https://doi.org/10.3390/jcm13195961 - 7 Oct 2024

Viewed by 580

Abstract

Background: Cystic fibrosis bone disease (CFBD) is a common comorbidity in adult people with cystic fibrosis (pwCF), resulting in an increased risk of bone fractures. This study evaluated the capacity of artificial intelligence (AI)-assisted low-dose chest CT (LDCT) opportunistic screening for detecting low [...] Read more.

Background: Cystic fibrosis bone disease (CFBD) is a common comorbidity in adult people with cystic fibrosis (pwCF), resulting in an increased risk of bone fractures. This study evaluated the capacity of artificial intelligence (AI)-assisted low-dose chest CT (LDCT) opportunistic screening for detecting low bone mineral density (BMD) in adult pwCF. Methods: In this retrospective single-center study, 65 adult pwCF (mean age 30.1 ± 7.5 years) underwent dual-energy X-ray absorptiometry (DXA) of the lumbar vertebrae L1 to L4 to determine BMD and corresponding z-scores and completed LDCTs of the chest within three months as part of routine clinical care. A fully automated CT-based AI algorithm measured the attenuation values (Hounsfield units [HU]) of the thoracic vertebrae Th9–Th12 and first lumbar vertebra L1. The ability of the algorithm to diagnose CFBD was assessed using receiver operating characteristic (ROC) curves. Results: HU values of Th9 to L1 and DXA-derived BMD and the corresponding z-scores of L1 to L4 showed a strong correlation (all p < 0.05). The area under the curve (AUC) for diagnosing low BMD was highest for L1 (0.796; p = 0.001) and Th11 (0.835; p < 0.001), resulting in a specificity of 84.9% at a sensitivity level of 75%. The HU threshold values for distinguishing normal from low BMD were <197 (L1) and <212 (Th11), respectively. Conclusions: Routine LDCT of the chest with the fully automated AI-guided determination of thoracic and lumbar vertebral attenuation values is a valuable tool for predicting low BMD in adult pwCF, with the best results for Th11 and L1. However, further studies are required to define clear threshold values. Full article

(This article belongs to the Topic AI in Medical Imaging and Image Processing)

►▼ Show Figures

Figure 1

Figure 1
<p>Examples of AI-assisted determinations of attenuation values for vertebral bodies L1 to Th9 using low-dose computed tomography (LDCT). From left to right: L1–Th9. (<b>a</b>) Normal findings. The ROI placed by the AI was appropriately centered on the vertebral body to be measured. The measured HU values ranged from 223 to 364. (<b>b</b>) Low BMD: The ROI is centered correctly by the AI on the vertebral bodies for the assessment of attenuation values. This results in HU values ranging between 89 and 117. (<b>c</b>) Incorrect measurements: The ROIs for L1, Th12, Th10, and Th9 were placed within the intervertebral disc space, resulting in underscored HU measurements for L1 (89), Th12 (122), Th10 (88), and Th9 (94). Conversely, the ROI for Th11 was positioned centrally within the vertebral body, with an HU value of 255. AI, artificial intelligence; ROI, region of interest; HU, Hounsfield unit; BMD, bone mass density.</p> Full article ">Figure 2
<p>Scatter plot of CT-attenuation values and DXA-derived BMD. •, normal BMD; ▪, CFBD. CFBD, cystic fibrosis-related bone disease; HU, Hounsfield units; DXA, dual-energy X-ray absorptiometry; BMD, bone mineral density; r, correlation coefficient.</p> Full article ">Figure 3
<p>Scatter plot of CT-attenuation values and DXA-derived z-scores. •, normal BMD; ▪, CFBD CFBD, cystic fibrosis-related bone disease; HU, Hounsfield units; DXA, dual-energy X-ray absorptiometry; BMD, bone mineral density; r, correlation coefficient.</p> Full article ">Figure 4
<p>Receiver operating characteristic (ROC) curves of CT-attentuation values for predicting low BMD. AUC, area under the curve; BMD, bone mineral density; ROC, receiver operating characteristic.</p> Full article ">

16 pages, 6662 KiB  

Open AccessArticle

Fully Automated Detection of the Appendix Using U-Net Deep Learning Architecture in CT Scans

by Betül Tiryaki Baştuğ, Gürkan Güneri, Mehmet Süleyman Yıldırım, Kadir Çorbacı and Emre Dandıl

J. Clin. Med. 2024, 13(19), 5893; https://doi.org/10.3390/jcm13195893 - 2 Oct 2024

Viewed by 645

Abstract

Background: The accurate segmentation of the appendix with well-defined boundaries is critical for diagnosing conditions such as acute appendicitis. The manual identification of the appendix is time-consuming and highly dependent on the expertise of the radiologist. Method: In this study, we propose a [...] Read more.

Background: The accurate segmentation of the appendix with well-defined boundaries is critical for diagnosing conditions such as acute appendicitis. The manual identification of the appendix is time-consuming and highly dependent on the expertise of the radiologist. Method: In this study, we propose a fully automated approach to the detection of the appendix using deep learning architecture based on the U-Net with specific training parameters in CT scans. The proposed U-Net architecture is trained on an annotated original dataset of abdominal CT scans to segment the appendix efficiently and with high performance. In addition, to extend the training set, data augmentation techniques are applied for the created dataset. Results: In experimental studies, the proposed U-Net model is implemented using hyperparameter optimization and the performance of the model is evaluated using key metrics to measure diagnostic reliability. The trained U-Net model achieved the segmentation performance for the detection of the appendix in CT slices with a Dice Similarity Coefficient (DSC), Volumetric Overlap Error (VOE), Average Symmetric Surface Distance (ASSD), Hausdorff Distance 95 (HD95), Precision (PRE) and Recall (REC) of 85.94%, 23.29%, 1.24 mm, 5.43 mm, 86.83% and 86.62%, respectively. Moreover, our model outperforms other methods by leveraging the U-Net’s ability to capture spatial context through encoder–decoder structures and skip connections, providing a correct segmentation output. Conclusions: The proposed U-Net model showed reliable performance in segmenting the appendix region, with some limitations in cases where the appendix was close to other structures. These improvements highlight the potential of deep learning to significantly improve clinical outcomes in appendix detection. Full article

(This article belongs to the Topic AI in Medical Imaging and Image Processing)

►▼ Show Figures

Figure 1

Figure 1
<p>Block diagram of the proposed detection system for the fully automated detection of appendix region in CT scans.</p> Full article ">Figure 2
<p>Sample slices from the dataset indicating the annotated appendix regions with GT masks.</p> Full article ">Figure 3
<p>Proposed U-Net deep learning architecture for automated detection of appendix.</p> Full article ">Figure 4
<p>The effect of the data augmentation procedures on some CT scans and GT masks in the dataset.</p> Full article ">Figure 5
<p>(<b>a</b>) Training loss and (<b>b</b>) DSC development during test phase for U-Net, DenseNet, and Res U-Net methods.</p> Full article ">Figure 6
<p>The appendix regions successfully detected and segmented on CT slices using the proposed U-Net deep learning architecture during the experimental studies. Red: ground truth mask for appendix, yellow: U-Net segmentation for appendix.</p> Full article ">Figure 7
<p>Some examples of unsuccessful appendix detection and segmentation by the proposed U-Net model. Red: ground truth mask for appendix, yellow: U-Net segmentation for appendix.</p> Full article ">Figure 8
<p>Boxplot showing the performance metrics for appendix segmentation obtained using the proposed U-Net deep learning architecture and other models in the study.</p> Full article ">Figure 9
<p>The comparison of the appendix segmentation performances of the proposed U-Net deep learning model and other state-of-the-art DenseNet and Res U-Net architectures on the same CT slices.</p> Full article ">

14 pages, 828 KiB  

Open AccessArticle

Lightweight MRI Brain Tumor Segmentation Enhanced by Hierarchical Feature Fusion

by Lei Zhang, Rong Zhang, Zhongjie Zhu, Pei Li, Yongqiang Bai and Ming Wang

Tomography 2024, 10(10), 1577-1590; https://doi.org/10.3390/tomography10100116 - 1 Oct 2024

Viewed by 553

Abstract

Background: Existing methods for MRI brain tumor segmentation often suffer from excessive model parameters and suboptimal performance in delineating tumor boundaries. Methods: For this issue, a lightweight MRI brain tumor segmentation method, enhanced by hierarchical feature fusion (EHFF), is proposed. This method reduces [...] Read more.

Background: Existing methods for MRI brain tumor segmentation often suffer from excessive model parameters and suboptimal performance in delineating tumor boundaries. Methods: For this issue, a lightweight MRI brain tumor segmentation method, enhanced by hierarchical feature fusion (EHFF), is proposed. This method reduces model parameters while improving segmentation performance by integrating hierarchical features. Initially, a fine-grained feature adjustment network is crafted and guided by global contextual information, leading to the establishment of an adaptive feature learning (AFL) module. This module captures the global features of MRI brain tumor images through macro perception and micro focus, adjusting spatial granularity to enhance feature details and reduce computational complexity. Subsequently, a hierarchical feature weighting (HFW) module is constructed. This module extracts multi-scale refined features through multi-level weighting, enhancing the detailed features of spatial positions and alleviating the lack of attention to local position details in macro perception. Finally, a hierarchical feature retention (HFR) module is designed as a supplementary decoder. This module retains, up-samples, and fuses feature maps from each layer, thereby achieving better detail preservation and reconstruction. Results: Experimental results on the BraTS 2021 dataset demonstrate that the proposed method surpasses existing methods. Dice similarity coefficients (DSC) for the three semantic categories ET, TC, and WT are 88.57%, 91.53%, and 93.09%, respectively. Full article

(This article belongs to the Topic AI in Medical Imaging and Image Processing)

►▼ Show Figures

Figure 1

Figure 1
<p>The overall architecture of EHHF.</p> Full article ">Figure 2
<p>Multistage architecture of the encoder.</p> Full article ">Figure 3
<p>The flowchart of the AFL module.</p> Full article ">Figure 4
<p>The flowchart of the HFW module.</p> Full article ">Figure 5
<p>Visualization of 3D MRI images and brain tumor annotation.</p> Full article ">Figure 6
<p>Segmentation comparison by various methods.</p> Full article ">Figure 7
<p>The visual comparison of the performance.</p> Full article ">

attachment

Supplementary material:
Supplementary File 1 (ZIP, 158 KiB)

10 pages, 757 KiB  

Open AccessArticle

Effectiveness of an Artificial Intelligence Software for Limb Radiographic Fracture Recognition in an Emergency Department

by Guillaume Herpe, Helena Nelken, Tanguy Vendeuvre, Jeremy Guenezan, Clement Giraud, Olivier Mimoz, Antoine Feydy, Jean-Pierre Tasu and Rémy Guillevin

J. Clin. Med. 2024, 13(18), 5575; https://doi.org/10.3390/jcm13185575 - 20 Sep 2024

Viewed by 686

Abstract

Objectives: To assess the impact of an Artificial Intelligence (AI) limb bone fracture diagnosis software (AIS) on emergency department (ED) workflow and diagnostic accuracy. Materials and Methods: A retrospective study was conducted in two phases—without AIS (Period 1: 1 January 2020–30 June 2020) [...] Read more.

Objectives: To assess the impact of an Artificial Intelligence (AI) limb bone fracture diagnosis software (AIS) on emergency department (ED) workflow and diagnostic accuracy. Materials and Methods: A retrospective study was conducted in two phases—without AIS (Period 1: 1 January 2020–30 June 2020) and with AIS (Period 2: 1 January 2021–30 June 2021). Results: Among 3720 patients (1780 in Period 1; 1940 in Period 2), the discrepancy rate decreased by 17% (p = 0.04) after AIS implementation. Clinically relevant discrepancies showed no significant change (−1.8%, p = 0.99). The mean length of stay in the ED was reduced by 9 min (p = 0.03), and expert consultation rates decreased by 1% (p = 0.38). Conclusions: AIS implementation reduced the overall discrepancy rate and slightly decreased ED length of stay, although its impact on clinically relevant discrepancies remains inconclusive. Key Point: After AI software deployment, the rate of radiographic discrepancies decreased by 17% (p = 0.04) but this was not clinically relevant (−2%, p = 0.99). Length of patient stay in the emergency department decreased by 5% with AI (p = 0.03). Bone fracture AI software is effective, but its effectiveness remains to be demonstrated. Full article

(This article belongs to the Topic AI in Medical Imaging and Image Processing)

►▼ Show Figures

Figure 1

Figure 1
<p>Flowchart of the study.</p> Full article ">Figure 2
<p>Multivariate risk analysis regarding discrepancies focusing on AI deployment along with fracture location and expert opinion.</p> Full article ">

14 pages, 1218 KiB  

Open AccessArticle

Are Preoperative CT Findings Useful in Predicting the Duration of Laparoscopic Appendectomy in Pediatric Patients? A Single Center Study

by Ismail Taskent, Bunyamin Ece and Mehmet Ali Narsat

J. Clin. Med. 2024, 13(18), 5504; https://doi.org/10.3390/jcm13185504 - 18 Sep 2024

Viewed by 497

Abstract

Background/Objectives: Preoperative computed tomography (CT) imaging plays a vital role in accurately diagnosing acute appendicitis and assessing the severity of the condition, as well as the complexity of the surgical procedure. CT imaging provides detailed information on the anatomical and pathological aspects of [...] Read more.

Background/Objectives: Preoperative computed tomography (CT) imaging plays a vital role in accurately diagnosing acute appendicitis and assessing the severity of the condition, as well as the complexity of the surgical procedure. CT imaging provides detailed information on the anatomical and pathological aspects of appendicitis, allowing surgeons to anticipate technical challenges and select the most appropriate surgical approach. This retrospective study aimed to investigate the correlation between preoperative CT findings and the duration of laparoscopic appendectomy (LA) in pediatric patients. Methods: This retrospective study included 104 pediatric patients diagnosed with acute appendicitis via contrast-enhanced CT who subsequently underwent laparoscopic appendectomy (LA) between November 2021 and February 2024. CT images were meticulously reviewed by two experienced radiologists blinded to the clinical and surgical outcomes. The severity of appendicitis was evaluated using a five-point scale based on the presence of periappendiceal fat, fluid, extraluminal air, and abscesses. Results: The average operation time was 51.1 ± 21.6 min. Correlation analysis revealed significant positive associations between operation time and neutrophil count (p = 0.014), C-reactive protein levels (p = 0.002), symptom-to-operation time (p = 0.004), and appendix diameter (p = 0.017). The total CT score also showed a significant correlation with operation time (p < 0.001). Multiple regression analysis demonstrated that a symptom duration of more than 2 days (p = 0.047), time from CT to surgery (p = 0.039), and the presence of a periappendiceal abscess (p = 0.005) were independent predictors of prolonged operation time. In the perforated appendicitis group, the presence of a periappendiceal abscess on CT was significantly associated with prolonged operation time (p = 0.020). In the non-perforated group, the presence of periappendiceal fluid was significantly related to longer operation times (p = 0.026). Conclusions: In our study, preoperative CT findings, particularly the presence of a periappendiceal abscess, were significantly associated with prolonged operation times in pediatric patients undergoing laparoscopic appendectomy. Elevated CRP levels, the time between CT imaging and surgery, and a symptom duration of more than 2 days were also found to significantly impact the procedure’s duration. Full article

(This article belongs to the Topic AI in Medical Imaging and Image Processing)

►▼ Show Figures

Figure 1

Figure 1
<p>Axial CT images of a 189-month-old male patient. (<b>A</b>) The appendix (black arrow) exhibits an increased diameter and a thickened wall, indicative of inflammation. There is a notable collection of periappendiceal fluid (black arrowhead), and the surrounding fat tissue shows increased attenuation (white arrowhead), suggesting inflammatory changes. (<b>B</b>) In another section, an appendicolith (black arrow) is clearly visible. Additionally, there is increased attenuation in the periappendiceal fat tissue (white arrow), consistent with inflammatory changes associated with acute appendicitis.</p> Full article ">Figure 2
<p>The impact of a scoring system comprising five variables (0 to 5) from CT findings on average operation time (min).</p> Full article ">Figure 3
<p>LOESS curve depicting the nonlinear relationship between symptom duration and operation time.</p> Full article ">

14 pages, 898 KiB  

Open AccessArticle

Interrater Variability of ML-Based CT-FFR in Patients without Obstructive CAD before TAVR: Influence of Image Quality, Coronary Artery Calcifications, and Location of Measurement

by Robin F. Gohmann, Adrian Schug, Christian Krieghoff, Patrick Seitz, Nicolas Majunke, Maria Buske, Fyn Kaiser, Sebastian Schaudt, Katharina Renatus, Steffen Desch, Sergey Leontyev, Thilo Noack, Philipp Kiefer, Konrad Pawelka, Christian Lücke, Ahmed Abdelhafez, Sebastian Ebel, Michael A. Borger, Holger Thiele, Christoph Panknin, Mohamed Abdel-Wahab, Matthias Horn and Matthias Gutberletadd Show full author list remove Hide full author list

J. Clin. Med. 2024, 13(17), 5247; https://doi.org/10.3390/jcm13175247 - 4 Sep 2024

Viewed by 720

Abstract

Objectives: CT-derived fractional flow reserve (CT-FFR) can improve the specificity of coronary CT-angiography (cCTA) for ruling out relevant coronary artery disease (CAD) prior to transcatheter aortic valve replacement (TAVR). However, little is known about the reproducibility of CT-FFR and the influence of [...] Read more.

Objectives: CT-derived fractional flow reserve (CT-FFR) can improve the specificity of coronary CT-angiography (cCTA) for ruling out relevant coronary artery disease (CAD) prior to transcatheter aortic valve replacement (TAVR). However, little is known about the reproducibility of CT-FFR and the influence of diffuse coronary artery calcifications or segment location. The objective was to assess the reliability of machine-learning (ML)-based CT-FFR prior to TAVR in patients without obstructive CAD and to assess the influence of image quality, coronary artery calcium score (CAC), and the location of measurement within the coronary tree. Methods: Patients assessed for TAVR, without obstructive CAD on cCTA were evaluated with ML-based CT-FFR by two observers with differing experience. Differences in absolute values and categorization into hemodynamically relevant CAD (CT-FFR ≤ 0.80) were compared. Results in regard to CAD were also compared against invasive coronary angiography. The influence of segment location, image quality, and CAC was evaluated. Results: Of the screened patients, 109/388 patients did not have obstructive CAD on cCTA and were included. The median (interquartile range) difference of CT-FFR values was −0.005 (−0.09 to 0.04) (p = 0.47). Differences were smaller with high values. Recategorizations were more frequent in distal segments. Diagnostic accuracy of CT-FFR between both observers was comparable (proximal: Δ0.2%; distal: Δ0.5%) but was lower in distal segments (proximal: 98.9%/99.1%; distal: 81.1%/81.6%). Image quality and CAC had no clinically relevant influence on CT-FFR. Conclusions: ML-based CT-FFR evaluation of proximal segments was more reliable. Distal segments with CT-FFR values close to the given threshold were prone to recategorization, even if absolute differences between observers were minimal and independent of image quality or CAC. Full article

(This article belongs to the Topic AI in Medical Imaging and Image Processing)

►▼ Show Figures

Figure 1

Figure 1
<p>Diagram of the 18-segment model of the coronary tree. Diagram showing the coronary tree with its division into 18 segments according to the SCCT guidelines [<a href="#B16-jcm-13-05247" class="html-bibr">16</a>]. Shaded segments were defined as proximal in this study. L = left; LAD = left anterior descending artery; PDA = posterior descending artery; PLB = posterior-lateral branch; R = right; RCA = right coronary artery; SCCT = Society of Cardiovascular Computed Tomography.</p> Full article ">Figure 2
<p>CT-FFR values at patient level. Bland-Altman plot showing the distribution of differences in CT-FFR values between both observers. Differences between both observers are overall very small. The plot shows no indication of a systematic bias. Lower mean CT-FFR values show one outlier (CT-FFR difference of −0.45 at a CT-FFR mean value of 0.58), which has been omitted from the plot for clarity. Lower mean CT-FFR values also show a larger heterogeneity of CT-FFR differences. CT-FFR = CT-derived fractional flow reserve. Dashed horizontal lines represent the mean difference (middle line) and the upper and lower limits of agreement (±1.96 SD, upper and lower lines).</p> Full article ">

14 pages, 2364 KiB  

Open AccessArticle

Repurposing the Public BraTS Dataset for Postoperative Brain Tumour Treatment Response Monitoring

by Peter Jagd Sørensen, Claes Nøhr Ladefoged, Vibeke Andrée Larsen, Flemming Littrup Andersen, Michael Bachmann Nielsen, Hans Skovgaard Poulsen, Jonathan Frederik Carlsen and Adam Espe Hansen

Tomography 2024, 10(9), 1397-1410; https://doi.org/10.3390/tomography10090105 - 1 Sep 2024

Viewed by 815

Abstract

The Brain Tumor Segmentation (BraTS) Challenge has been a main driver of the development of deep learning (DL) algorithms and provides by far the largest publicly available expert-annotated brain tumour dataset but contains solely preoperative examinations. The aim of our study was to [...] Read more.

The Brain Tumor Segmentation (BraTS) Challenge has been a main driver of the development of deep learning (DL) algorithms and provides by far the largest publicly available expert-annotated brain tumour dataset but contains solely preoperative examinations. The aim of our study was to facilitate the use of the BraTS dataset for training DL brain tumour segmentation algorithms for a postoperative setting. To this end, we introduced an automatic conversion of the three-label BraTS annotation protocol to a two-label annotation protocol suitable for postoperative brain tumour segmentation. To assess the viability of the label conversion, we trained a DL algorithm using both the three-label and the two-label annotation protocols. We assessed the models pre- and postoperatively and compared the performance with a state-of-the-art DL method. The DL algorithm trained using the BraTS three-label annotation misclassified parts of 10 out of 41 fluid-filled resection cavities in 72 postoperative glioblastoma MRIs, whereas the two-label model showed no such inaccuracies. The tumour segmentation performance of the two-label model both pre- and postoperatively was comparable to that of a state-of-the-art algorithm for tumour volumes larger than 1 cm³. Our study enables using the BraTS dataset as a basis for the training of DL algorithms for postoperative tumour segmentation. Full article

(This article belongs to the Topic AI in Medical Imaging and Image Processing)

►▼ Show Figures

Figure 1

Figure 1
<p>Conversion of the BraTS 2021 three-label segmentation masks for harmonisation with the proposed two-label annotation protocol. <b>Panel A</b>: Anatomical sequences without annotations. T1-weighted (T1) (top left), T1-weighted after contrast infusion (T1Gd) (top right), T2-weighted (bottom left) and fluid-attenuated-inversion-recovery (FLAIR) (bottom right). <b>Panels B</b>–<b>D</b>: annotations overlaid on T1Gd (top) and FLAIR (bottom). NCR+NET = necrosis, cysts and non-enhancing tumour core; AT = active contrast-enhancing tumour; ED = oedema and infiltrated tissue; WT_HD-GLIO = automatic whole-tumour segmentation by HD-GLIO; CE = contrast-enhancing tumour; NE = non-enhancing hyperintense T2/FLAIR signal abnormalities. In essence, the conversion process labels the combination of the NCR+NET mask inside the WT_HD-GLIO region and the ED mask as NE, thereby removing necrosis and cysts while retaining non-enhancing tumour core. AT is relabelled as CE without change. Only the regions inside the NCR+NET segmentations (red) are affected by the conversion.</p> Full article ">Figure 2
<p>Representative case from the postoperative test set of segmentations performed by deep learning models trained utilising the BraTS three-label annotation protocol and our proposed two-label protocol and by radiologist as reference. <b>Panel A</b>: Anatomical sequences without annotations. T1 (top left), T1Gd (top right), T2 (bottom left) and FLAIR (bottom right). <b>Panels B</b>–<b>D</b>: Annotations overlaid on T1Gd (top) and FLAIR (bottom). NCR+NET = necrosis, cysts and non-enhancing tumour core; AT = active contrast-enhancing tumour; ED = oedema and infiltrated tissue; CE = contrast-enhancing tumour; NE = non-enhancing hyperintense T2/FLAIR signal abnormalities. The three-label model includes parts of the resection cavity (panel B) while the two-label model (panel C) and radiologist reference (panel D) do not.</p> Full article ">Figure 3
<p>Agreement in preoperative brain tumour segmentations between deep learning models and two-label converted expert annotations across 251 cases from the BraTS 2021 training set, reserved for testing. The boxplots depict the distributions of Dice similarity coefficients (DSCs). CE = contrast-enhancing tumour; NE = non-enhancing hyperintense T2/FLAIR signal abnormalities.</p> Full article ">Figure 4
<p>Agreement in postoperative brain tumour segmentations between deep learning models and our independent expert two-label annotations for 72 cases from the postoperative test set. The boxplots depict the distributions of Dice similarity coefficients (DSCs). CE = contrast-enhancing tumour; NE = non-enhancing hyperintense T2/FLAIR signal abnormalities; CE (volume &gt; 1 cm³) = CE segmentations with volumes exceeding 1 cm³.</p> Full article ">Figure A1
<p>Schematic illustration of the steps of the conversion of the BraTS 2021 three-label segmentation masks for harmonisation with the proposed two-label annotation protocol. The top panel shows the original BraTS three-label annotations and the bottom panel the resulting two-label annotations. At each step, a left illustration shows what is currently treated and how, and a right illustration shows the current result of the process. AT = active contrast-enhancing tumour; ED = oedema and infiltrated tissue; WT_HD-GLIO = automatic whole-tumour segmentation by HD-GLIO; CE = contrast-enhancing tumour; NE = non-enhancing hyperintense T2/FLAIR signal abnormalities.</p> Full article ">

attachment

Supplementary material:
Supplementary File 1 (ZIP, 1145 KiB)

14 pages, 10901 KiB  

Open AccessArticle

EpidermaQuant: Unsupervised Detection and Quantification of Epidermal Differentiation Markers on H-DAB-Stained Images of Reconstructed Human Epidermis

by Dawid Zamojski, Agnieszka Gogler, Dorota Scieglinska and Michal Marczyk

Diagnostics 2024, 14(17), 1904; https://doi.org/10.3390/diagnostics14171904 - 29 Aug 2024

Viewed by 508

Abstract

The integrity of the reconstructed human epidermis generated in vitro can be assessed using histological analyses combined with immunohistochemical staining of keratinocyte differentiation markers. Technical differences during the preparation and capture of stained images may influence the outcome of computational methods. Due to [...] Read more.

The integrity of the reconstructed human epidermis generated in vitro can be assessed using histological analyses combined with immunohistochemical staining of keratinocyte differentiation markers. Technical differences during the preparation and capture of stained images may influence the outcome of computational methods. Due to the specific nature of the analyzed material, no annotated datasets or dedicated methods are publicly available. Using a dataset with 598 unannotated images showing cross-sections of in vitro reconstructed human epidermis stained with DAB-based immunohistochemistry reaction to visualize four different keratinocyte differentiation marker proteins (filaggrin, keratin 10, Ki67, HSPA2) and counterstained with hematoxylin, we developed an unsupervised method for the detection and quantification of immunohistochemical staining. The pipeline consists of the following steps: (i) color normalization; (ii) color deconvolution; (iii) morphological operations; (iv) automatic image rotation; and (v) clustering. The most effective combination of methods includes (i) Reinhard’s normalization; (ii) Ruifrok and Johnston color-deconvolution method; (iii) proposed image-rotation method based on boundary distribution of image intensity; and (iv) k-means clustering. The results of the work should enhance the performance of quantitative analyses of protein markers in reconstructed human epidermis samples and enable the comparison of their spatial distribution between different experimental conditions. Full article

(This article belongs to the Topic AI in Medical Imaging and Image Processing)

►▼ Show Figures

Figure 1

Figure 1
<p>An example of IHC images of reconstructed human epidermis stained for: (<b>A</b>)—FLG marker; (<b>B</b>)—K10 marker; (<b>C</b>)—Ki67 marker; (<b>D</b>)—HSPA2 marker.</p> Full article ">Figure 2
<p>EpidermaQuant algorithm flowchart. The main steps of the pipeline include image color normalization, image color deconvolution, background detection, image rotation and crop, initial DAB-area detection step where the algorithm decides whether the image should be further analyzed, image segmentation, and calculation of the DAB percentage on tissue.</p> Full article ">Figure 3
<p>Comparison of normalization (<b>A</b>) and color-deconvolution (<b>B</b>) methods with the original image, depending on the selected marker using exemplary images.</p> Full article ">Figure 4
<p>Subsequent steps of background removal (<b>A</b>) and comparison of two image rotation methods for an exemplary image of K10 marker (<b>B</b>).</p> Full article ">Figure 5
<p>Example of DAB image segmentation result. Interesting regions of markers of human epidermal differentiation are segmented by the k-means method. The selected cluster outlines the DAB-stained areas (red line) on the original image with additional hematoxylin staining (blue).</p> Full article ">Figure 6
<p>The percentage of DAB-stained tissue. The graph shows the percentage of DAB occupancy on the slide in each sample by study marker group.</p> Full article ">Figure 7
<p>Exemplary results of EpidermaQuant on HPA dataset. For each analyzed marker (rows), images with the lowest and the highest estimated percentage of DAB occupancy were selected. The red line indicates the estimated area of the tissue.</p> Full article ">Figure 8
<p>Summary of the ablation study using the K10-marker images. All results are compared to the full pipeline. Panel (<b>A</b>) shows the computational time comparison for selected configurations of pipeline, while panel (<b>B</b>) shows the difference in the percentage of DAB occupancy.</p> Full article ">

14 pages, 1275 KiB  

Open AccessArticle

VerFormer: Vertebrae-Aware Transformer for Automatic Spine Segmentation from CT Images

by Xinchen Li, Yuan Hong, Yang Xu and Mu Hu

Diagnostics 2024, 14(17), 1859; https://doi.org/10.3390/diagnostics14171859 - 25 Aug 2024

Viewed by 760

Abstract

The accurate and efficient segmentation of the spine is important in the diagnosis and treatment of spine malfunctions and fractures. However, it is still challenging because of large inter-vertebra variations in shape and cross-image localization of the spine. In previous methods, convolutional neural [...] Read more.

The accurate and efficient segmentation of the spine is important in the diagnosis and treatment of spine malfunctions and fractures. However, it is still challenging because of large inter-vertebra variations in shape and cross-image localization of the spine. In previous methods, convolutional neural networks (CNNs) have been widely applied as a vision backbone to tackle this task. However, these methods are challenged in utilizing the global contextual information across the whole image for accurate spine segmentation because of the inherent locality of the convolution operation. Compared with CNNs, the Vision Transformer (ViT) has been proposed as another vision backbone with a high capacity to capture global contextual information. However, when the ViT is employed for spine segmentation, it treats all input tokens equally, including vertebrae-related tokens and non-vertebrae-related tokens. Additionally, it lacks the capability to locate regions of interest, thus lowering the accuracy of spine segmentation. To address this limitation, we propose a novel Vertebrae-aware Vision Transformer (VerFormer) for automatic spine segmentation from CT images. Our VerFormer is designed by incorporating a novel Vertebrae-aware Global (VG) block into the ViT backbone. In the VG block, the vertebrae-related global contextual information is extracted by a Vertebrae-aware Global Query (VGQ) module. Then, this information is incorporated into query tokens to highlight vertebrae-related tokens in the multi-head self-attention module. Thus, this VG block can leverage global contextual information to effectively and efficiently locate spines across the whole input, thus improving the segmentation accuracy of VerFormer. Driven by this design, the VerFormer demonstrates a solid capacity to capture more discriminative dependencies and vertebrae-related context in automatic spine segmentation. The experimental results on two spine CT segmentation tasks demonstrate the effectiveness of our VG block and the superiority of our VerFormer in spine segmentation. Compared with other popular CNN- or ViT-based segmentation models, our VerFormer shows superior segmentation accuracy and generalization. Full article

(This article belongs to the Topic AI in Medical Imaging and Image Processing)

►▼ Show Figures

Figure 1

Figure 1
<p>The overall architecture of the VerFormer. It consists of an encoder, a bottleneck, and a decoder. Both the encoder and the decoder consist of three stages, and in each stage, two successive VG blocks are used for learning feature representations. Two successive VG blocks are used in the bottleneck. A <math display="inline"><semantics> <mrow> <mn>3</mn> <mo>×</mo> <mn>3</mn> </mrow> </semantics></math> convolutional layer with a stride of 2 is used for downsampling in each stage, and a <math display="inline"><semantics> <mrow> <mn>2</mn> <mo>×</mo> <mn>2</mn> </mrow> </semantics></math> transposed convolutional layer with a stride of 2 is used for upsampling in each stage. Two <math display="inline"><semantics> <mrow> <mn>3</mn> <mo>×</mo> <mn>3</mn> </mrow> </semantics></math> convolutional layers with a stride of 2 are used for the patch partition. Two <math display="inline"><semantics> <mrow> <mn>2</mn> <mo>×</mo> <mn>2</mn> </mrow> </semantics></math> transposed convolution layers with a stride of 2 are used in the final projection.</p> Full article ">Figure 2
<p>The overall architecture of the Vertebrae-aware Global (VG) block. In the VG block, a VGQ module is utilized to extract vertebrae-aware contextual information to highlight vertebrae-related query tokens. Then, these tokens are encoded in the multi-head self-attention (MHSA) module. In the VGQ module, two parallel paths are employed to extract global vertebrae-aware contextual information, including a channel attention path and a spatial attention path. In the channel attention path, an SE layer is used by cascading an average pooling (AvgPool) layer, a fully connected (FC) layer, a ReLu activation function, and a fully connected (FC) layer.</p> Full article ">Figure 3
<p>The mechanism of the Vertebrae-aware Global Query (VGQ) module. The input image is partitioned into patches, and these patches are converted into input query tokens. The VGQ module is utilized to extract vertebrae-aware contextual information by channel and spatial attention mechanisms. Then, vertebrae-related patches or tokens are highlighted by this contextual information. Thus, the VGQ module can leverage global information to locate the spine across the whole image, improving the segmentation capabilities of our VerFormer on the spine from CT images.</p> Full article ">Figure 4
<p>The visualization of segmentation results from our VerFormer. (<b>A</b>) The visualization of segmentation results from our VerFormer on the VerSe 2019 dataset. (<b>B</b>) The visualization of segmentation results from our VerFormer on the VerSe 2020 dataset.</p> Full article ">

attachment

Supplementary material:
Supplementary File 1 (ZIP, 346 KiB)

15 pages, 1533 KiB  

Open AccessArticle

MRI T2w Radiomics-Based Machine Learning Models in Imaging Simulated Biopsy Add Diagnostic Value to PI-RADS in Predicting Prostate Cancer: A Retrospective Diagnostic Study

by Jia-Cheng Liu, Xiao-Hao Ruan, Tsun-Tsun Chun, Chi Yao, Da Huang, Hoi-Lung Wong, Chun-Ting Lai, Chiu-Fung Tsang, Sze-Ho Ho, Tsui-Lin Ng, Dan-Feng Xu and Rong Na

Cancers 2024, 16(17), 2944; https://doi.org/10.3390/cancers16172944 - 23 Aug 2024

Viewed by 582

Abstract

Background: Currently, prostate cancer (PCa) prebiopsy medical image diagnosis mainly relies on mpMRI and PI-RADS scores. However, PI-RADS has its limitations, such as inter- and intra-radiologist variability and the potential for imperceptible features. The primary objective of this study is to evaluate the [...] Read more.

Background: Currently, prostate cancer (PCa) prebiopsy medical image diagnosis mainly relies on mpMRI and PI-RADS scores. However, PI-RADS has its limitations, such as inter- and intra-radiologist variability and the potential for imperceptible features. The primary objective of this study is to evaluate the effectiveness of a machine learning model based on radiomics analysis of MRI T2-weighted (T2w) images for predicting PCa in prebiopsy cases. Method: A retrospective analysis was conducted using 820 lesions (363 cases, 457 controls) from The Cancer Imaging Archive (TCIA) Database for model development and validation. An additional 83 lesions (30 cases, 53 controls) from Hong Kong Queen Mary Hospital were used for independent external validation. The MRI T2w images were preprocessed, and radiomic features were extracted. Feature selection was performed using Cross Validation Least Angle Regression (CV-LARS). Using three different machine learning algorithms, a total of 18 prediction models and 3 shape control models were developed. The performance of the models, including the area under the curve (AUC) and diagnostic values such as sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV), were compared to the PI-RADS scoring system for both internal and external validation. Results: All the models showed significant differences compared to the shape control model (all p < 0.001, except SVM model PI-RADS+2 Features p = 0.004, SVM model PI-RADS+3 Features p = 0.002). In internal validation, the best model, based on the LR algorithm, incorporated 3 radiomic features (AUC = 0.838, sensitivity = 76.85%, specificity = 77.36%). In external validation, the LR (3 features) model outperformed PI-RADS in predictive value with AUC 0.870 vs. 0.658, sensitivity 56.67% vs. 46.67%, specificity 92.45% vs. 84.91%, PPV 80.95% vs. 63.64%, and NPV 79.03% vs. 73.77%. Conclusions: The machine learning model based on radiomics analysis of MRI T2w images, along with simulated biopsy, provides additional diagnostic value to the PI-RADS scoring system in predicting PCa. Full article

(This article belongs to the Topic AI in Medical Imaging and Image Processing)

►▼ Show Figures

Figure 1

Figure 1
<p>Workflow of models’ construction and validation. <a href="#cancers-16-02944-f001" class="html-fig">Figure 1</a> describes the workflow of this study: (1) The MRI images were obtained from two sources: The Cancer Imaging Archive (TCIA) Database Collection (75 patients, 820 lesions) and Hong Kong Queen Mary Hospital (8 patients, 83 lesions); (2) The 820 lesions from TCIA were undersampled to 718 lesions and randomly separated to training group (<span class="html-italic">n</span> = 504, 70%) and validation group (<span class="html-italic">n</span> = 214, 30%); (3) Initial regions of interest (ROIs) were manually delineated on the MRI image, as indicated by the red markers (BxTop and BxBot); (4) ROIs were refined and segmented. The green overlay highlights the segmented prostate region; (5) A set of radiomic features was screened using the LARS (Least Angle Regression) algorithm. This step identified the most relevant features for prostate cancer prediction; (6) The extracted features were used to train logistic regression, random forest, and support vector machine models; (7) Performance of models was both validated internally and externally.</p> Full article ">Figure 2
<p>Ten-fold cross-validation least angle regression for feature selection. <a href="#cancers-16-02944-f002" class="html-fig">Figure 2</a> Plots describing the 10-fold cross-validation least angle regression (cv-LARS) based feature selection process and exhibits the results. (<b>A</b>). illustrating the changes in cross-validated Mean Squared Error (MSE) with the number of steps. At the 8th step, the LARS algorithm reaches the minimum MSE. (<b>B</b>). exhibiting the solution path plot of 10-fold cv-LARS.</p> Full article ">Figure 3
<p>Comparison between best-performing models. <a href="#cancers-16-02944-f003" class="html-fig">Figure 3</a> column plots showing the area under the curve (AUC) of three best-performing models based on different algorithms. * <span class="html-italic">p</span> &lt; 0.05; ** <span class="html-italic">p</span> &lt; 0.01; n.s. indicates non-significance.</p> Full article ">Figure 4
<p>Comparison between the optimal model and PI-RADS in HQM set. <a href="#cancers-16-02944-f004" class="html-fig">Figure 4</a> column plot showing the area under the curve of the logistic regression model (3 features + PI-RADS).</p> Full article ">

15 pages, 2761 KiB  

Open AccessArticle

Preoperative OCT Characteristics Contributing to Prediction of Postoperative Visual Acuity in Eyes with Macular Hole

by Yoko Mase, Yoshitsugu Matsui, Koki Imai, Kazuya Imamura, Akiko Irie-Ota, Shinichiro Chujo, Hisashi Matsubara, Hiroharu Kawanaka and Mineo Kondo

J. Clin. Med. 2024, 13(16), 4826; https://doi.org/10.3390/jcm13164826 - 15 Aug 2024

Viewed by 657

Abstract

Objectives: To develop a machine learning logistic regression algorithm that can classify patients with an idiopathic macular hole (IMH) into those with good or poor vison at 6 months after a vitrectomy. In addition, to determine its accuracy and the contribution of the [...] Read more.

Objectives: To develop a machine learning logistic regression algorithm that can classify patients with an idiopathic macular hole (IMH) into those with good or poor vison at 6 months after a vitrectomy. In addition, to determine its accuracy and the contribution of the preoperative OCT characteristics to the algorithm. Methods: This was a single-center, cohort study. The classifier was developed using preoperative clinical information and the optical coherence tomographic (OCT) findings of 43 eyes of 43 patients who had undergone a vitrectomy. The explanatory variables were selected using a filtering method based on statistical significance and variance inflation factor (VIF) values, and the objective variable was the best-corrected visual acuity (BCVA) at 6 months postoperation. The discrimination threshold of the BCVA was the 0.15 logarithm of the minimum angle of the resolution (logMAR) units. Results: The performance of the classifier was 0.92 for accuracy, 0.73 for recall, 0.60 for precision, 0.74 for F-score, and 0.84 for the area under the curve (AUC). In logistic regression, the standard regression coefficients were 0.28 for preoperative BCVA, 0.13 for outer nuclear layer defect length (ONL_DL), −0.21 for outer plexiform layer defect length (OPL_DL) − (ONL_DL), and −0.17 for (OPL_DL)/(ONL_DL). In the IMH form, a stenosis pattern with a narrowing from the OPL to the ONL of the MH had a significant effect on the postoperative BCVA at 6 months. Conclusions: Our results indicate that (OPL_DL) − (ONL_DL) had a similar contribution to preoperative visual acuity in predicting the postoperative visual acuity. This model had a strong performance, suggesting that the preoperative visual acuity and MH characteristics in the OCT images were crucial in forecasting the postoperative visual acuity in IMH patients. Thus, it can be used to classify MH patients into groups with good or poor postoperative visual acuity, and the classification was comparable to that of previous studies using deep learning. Full article

(This article belongs to the Topic AI in Medical Imaging and Image Processing)

►▼ Show Figures

Figure 1

Figure 1
<p>Preprocessing and annotation of OCT images to extract explanatory variables. (<b>A</b>) A preoperative optical coherence tomographic (OCT) image of an eye with a macular hole (MH). (<b>B</b>) Trimmed image of the OCT image in the red frame in (<b>A</b>). (<b>C</b>) Annotation for the outer plexiform (OPL) line, external limiting membrane (ELM) line, ellipsoid zone (EZ) line, and retinal pigment epithelial (RPE) line as demarcation lines in a trimmed image. The top layer of the image was painted green then in yellow, sky-blue, and blue in that order. The intraretinal fluid (IRF) was painted red. (<b>D</b>) The Shortest Defect Length (DL) for each color layer and DL for the OPL, ELM, EZ, and basal diameter.</p> Full article ">Figure 2
<p>Fisher’s probability of correct answer tests were performed for the categorical variables extracted from the clinical information and OCT images. Kolmogorov–Smirnov tests were performed for continuous quantitative variables. Variables that had a statistically significant difference between Group A and Group B were used as candidates for the explanatory variables. Among these variables, those with a variance inflation factor values &gt; 10, i.e., with R² scores &gt; 0.9, were eliminated, and the tests were repeated until the R² score for all variables was &lt;0.9. Finally, the remaining variables were used as the explanatory variables.</p> Full article ">Figure 3
<p>The algorithm shown in the upper box was used to adjust the parameters to maximize the area under the curve (AUC). Then, as shown in the algorithm in the lower box, a logistic regression model was created with the training data by fixing the parameters at the most frequent values among those explored in the upper box, and then the model was evaluated with the test data.</p> Full article ">Figure 4
<p>The frequency of each of the constant terms of the parameters is shown.</p> Full article ">Figure 5
<p>The mean receiver operating characteristic (ROC) curve of the classifiers for Group A and Group B at 6 months postoperation. The standard deviation (SD) is equivalent to the SD of the area under the ROC curve (AUC) obtained by evaluating the model 100 times with the test data.</p> Full article ">Figure 6
<p>Box-and-whisker diagrams with outliers showing coefficients of the explanatory variables at 6 months postoperation. The maximum and minimum whisker lengths are set at the upper and lower limits of 1.5 times the interquartile range (IQR). The first quartile −1.5 × IQR is the lower limit of the whisker, and the third quartile +1.5 × IQR is the upper limit of the whisker. Values less than the lower end of the whisker or more than the upper end of the whisker are indicated by circles as outliers.</p> Full article ">Figure 7
<p>The mean receiver operating characteristic (ROC) curve of the classifiers for Group A and Group B at 6 months postoperation using only the patients’ clinical information. The standard deviation (SD) is equivalent to the SD of the area under the ROC curve (AUC) obtained by evaluating the model 100 times with the test data.</p> Full article ">

15 pages, 1070 KiB  

Open AccessArticle

Reproducibility and Repeatability in Focus: Evaluating LVEF Measurements with 3D Echocardiography by Medical Technologists

by Marc Østergaard Nielsen, Arlinda Ljoki, Bo Zerahn, Lars Thorbjørn Jensen and Bent Kristensen

Diagnostics 2024, 14(16), 1729; https://doi.org/10.3390/diagnostics14161729 - 9 Aug 2024

Viewed by 656

Abstract

Three-dimensional echocardiography (3DE) is currently the preferred method for monitoring left ventricular ejection fraction (LVEF) in cancer patients receiving potentially cardiotoxic anti-neoplastic therapy. In Denmark, however, the traditional standard for LVEF monitoring has been rooted in nuclear medicine departments utilizing equilibrium radionuclide angiography [...] Read more.

Three-dimensional echocardiography (3DE) is currently the preferred method for monitoring left ventricular ejection fraction (LVEF) in cancer patients receiving potentially cardiotoxic anti-neoplastic therapy. In Denmark, however, the traditional standard for LVEF monitoring has been rooted in nuclear medicine departments utilizing equilibrium radionuclide angiography (ERNA). Although ERNA remains a principal modality, there is an emerging trend towards the adoption of echocardiography for this purpose. Given this context, assessing the reproducibility of 3DE among non-specialized medical personnel is crucial for its clinical adoption in such departments. To assess the feasibility of 3DE for LVEF measurements by technologists, we evaluated the repeatability and reproducibility of two moderately experienced technologists. They performed 3DE on 12 volunteers over two sessions, with a collaborative review of the results from the first session before the second session. Two-way intraclass correlation values increased from 0.03 to 0.77 across the sessions. This increase in agreement was mainly due to the recognition of false low measurements. Our findings underscore the importance of incorporating reproducibility exercises in the context of 3DE, especially when operated by technologists. Additionally, routine control of the acquisitions by physicians is deemed necessary. Ensuring these hurdles are adequately managed enables the adoption of 3DE for LVEF measurements by technologists. Full article

(This article belongs to the Topic AI in Medical Imaging and Image Processing)

►▼ Show Figures

Figure 1

Figure 1
<p>Boxplots of LVEF measurements by OP1 (<b>a</b>) and OP2 (<b>b</b>), further split by sex and measurement session.</p> Full article ">Figure 2
<p>Strip-dot plots of LVEF percentages by OP1 (<b>a</b>) and OP2 (<b>b</b>). Triangles represent M1 and circles M2.</p> Full article ">Figure 3
<p>Bland–Altman plot with LOA (solid, colored lines) with confidence intervals (shaded grey areas). Triangles represent M1 and circles M2. The x-axis represents the average LVEF of OP1 and OP2, while the y-axis represents the difference.</p> Full article ">Figure A1
<p>Variability chart of the total measurement dataset generated using the “VCR” R package. Individual measurements are depicted as points, with the red plus symbol representing the mean of each replicate set. Grey horizontal bars indicate the mean of all measurements for each subject, while light-blue bars represent the means of the entire measurement session.</p> Full article ">Figure A2
<p>Paired-line plots or “spaghetti-plots”, representing the change in mean-LVEF between measurement sessions for operator 1 in panel (<b>a</b>) and operator 2 in panel (<b>b</b>).</p> Full article ">

19 pages, 5027 KiB  

Open AccessArticle

Brain Tumor Detection and Classification Using an Optimized Convolutional Neural Network

by Muhammad Aamir, Abdallah Namoun, Sehrish Munir, Nasser Aljohani, Meshari Huwaytim Alanazi, Yaser Alsahafi and Faris Alotibi

Diagnostics 2024, 14(16), 1714; https://doi.org/10.3390/diagnostics14161714 - 7 Aug 2024

Viewed by 3174

Abstract

Brain tumors are a leading cause of death globally, with numerous types varying in malignancy, and only 12% of adults diagnosed with brain cancer survive beyond five years. This research introduces a hyperparametric convolutional neural network (CNN) model to identify brain tumors, with [...] Read more.

Brain tumors are a leading cause of death globally, with numerous types varying in malignancy, and only 12% of adults diagnosed with brain cancer survive beyond five years. This research introduces a hyperparametric convolutional neural network (CNN) model to identify brain tumors, with significant practical implications. By fine-tuning the hyperparameters of the CNN model, we optimize feature extraction and systematically reduce model complexity, thereby enhancing the accuracy of brain tumor diagnosis. The critical hyperparameters include batch size, layer counts, learning rate, activation functions, pooling strategies, padding, and filter size. The hyperparameter-tuned CNN model was trained on three different brain MRI datasets available at Kaggle, producing outstanding performance scores, with an average value of 97% for accuracy, precision, recall, and F1-score. Our optimized model is effective, as demonstrated by our methodical comparisons with state-of-the-art approaches. Our hyperparameter modifications enhanced the model performance and strengthened its capacity for generalization, giving medical practitioners a more accurate and effective tool for making crucial judgments regarding brain tumor diagnosis. Our model is a significant step in the right direction toward trustworthy and accurate medical diagnosis, with practical implications for improving patient outcomes. Full article

(This article belongs to the Topic AI in Medical Imaging and Image Processing)

►▼ Show Figures

Figure 1

Figure 1
<p>Sample images from Dataset 2.</p> Full article ">Figure 2
<p>Sample images from Dataset 3.</p> Full article ">Figure 3
<p>Samples from Dataset 1.</p> Full article ">Figure 4
<p>The workflow of the proposed model.</p> Full article ">Figure 5
<p>The Confusion Matrix for Dataset 1.</p> Full article ">Figure 6
<p>The Confusion Matrix for Dataset 2.</p> Full article ">Figure 7
<p>The Confusion Matrix for Dataset 3.</p> Full article ">Figure 8
<p>Accuracy of training and validation for Dataset 1.</p> Full article ">Figure 9
<p>Training and validation loss in Dataset 1.</p> Full article ">Figure 10
<p>Accuracy of training and validation using Dataset 2.</p> Full article ">Figure 11
<p>Training and validation loss in Dataset 2.</p> Full article ">Figure 12
<p>Accuracy of validation and training using Dataset 3.</p> Full article ">Figure 13
<p>Losses in training and validation for Dataset 3.</p> Full article ">

14 pages, 2341 KiB  

Open AccessArticle

Evaluation of Machine Learning Classification Models for False-Positive Reduction in Prostate Cancer Detection Using MRI Data

by Malte Rippa, Ruben Schulze, Georgia Kenyon, Marian Himstedt, Maciej Kwiatkowski, Rainer Grobholz, Stephen Wyler, Alexander Cornelius, Sebastian Schindera and Felice Burn

Diagnostics 2024, 14(15), 1677; https://doi.org/10.3390/diagnostics14151677 - 2 Aug 2024

Viewed by 1031

Abstract

In this work, several machine learning (ML) algorithms, both classical ML and modern deep learning, were investigated for their ability to improve the performance of a pipeline for the segmentation and classification of prostate lesions using MRI data. The algorithms were used to [...] Read more.

In this work, several machine learning (ML) algorithms, both classical ML and modern deep learning, were investigated for their ability to improve the performance of a pipeline for the segmentation and classification of prostate lesions using MRI data. The algorithms were used to perform a binary classification of benign and malignant tissue visible in MRI sequences. The model choices include support vector machines (SVMs), random decision forests (RDFs), and multi-layer perceptrons (MLPs), along with radiomic features that are reduced by applying PCA or mRMR feature selection. Modern CNN-based architectures, such as ConvNeXt, ConvNet, and ResNet, were also evaluated in various setups, including transfer learning. To optimize the performance, different approaches were compared and applied to whole images, as well as gland, peripheral zone (PZ), and lesion segmentations. The contribution of this study is an investigation of several ML approaches regarding their performance in prostate cancer (PCa) diagnosis algorithms. This work delivers insights into the applicability of different approaches for this context based on an exhaustive examination. The outcome is a recommendation or preference for which machine learning model or family of models is best suited to optimize an existing pipeline when the model is applied as an upstream filter. Full article

(This article belongs to the Topic AI in Medical Imaging and Image Processing)

►▼ Show Figures

Figure 1

Figure 1
<p>Simplified setup of a sequential lesion segmentation/classification pipeline.</p> Full article ">Figure 2
<p>Pipeline evaluation mode 1: All classifiers that do not require lesion segmentations are placed at the beginning of the pipeline or after the model that predicts the required anatomy.</p> Full article ">Figure 3
<p>Pipeline evaluation mode 2: The lesion classifiers are placed at the end of the regular pipeline to produce a classification of the lesion candidates selected by the existing pipeline segmentation models.</p> Full article ">Figure 4
<p>MSE loss distribution with respect to the two classes.</p> Full article ">Figure 5
<p>Average 5-fold cross validation loss (binary cross entropy loss) for the training, validation, and test partition on the whole image with the Genesis model over 100 epochs.</p> Full article ">Figure 6
<p>Average 5-fold cross-validation AUC-ROC for the training, validation, and test partition on the whole image with the Genesis model over 100 epochs.</p> Full article ">

14 pages, 2354 KiB  

Open AccessArticle

Identifying Acute Aortic Syndrome and Thoracic Aortic Aneurysm from Chest Radiography in the Emergency Department Using Convolutional Neural Network Models

by Yang-Tse Lin, Bing-Cheng Wang and Jui-Yuan Chung

Diagnostics 2024, 14(15), 1646; https://doi.org/10.3390/diagnostics14151646 - 30 Jul 2024

Viewed by 892

Abstract

(1) Background: Identifying acute aortic syndrome (AAS) and thoracic aortic aneurysm (TAA) in busy emergency departments (EDs) is crucial due to their life-threatening nature, necessitating timely and accurate diagnosis. (2) Methods: This retrospective case-control study was conducted in the ED of three hospitals. [...] Read more.

(1) Background: Identifying acute aortic syndrome (AAS) and thoracic aortic aneurysm (TAA) in busy emergency departments (EDs) is crucial due to their life-threatening nature, necessitating timely and accurate diagnosis. (2) Methods: This retrospective case-control study was conducted in the ED of three hospitals. Adult patients visiting the ED between 1 January 2010 and 1 January 2020 with a chief complaint of chest or back pain were enrolled in the study. The collected chest radiography (CXRs) data were divided into training (80%) and testing (20%) datasets. The training dataset was trained by four different convolutional neural network (CNN) models. (3) Results: A total of 1625 patients were enrolled in this study. The InceptionV3 model achieved the highest F1 score of 0.76. (4) Conclusions: Analysis of CXRs using a CNN-based model provides a novel tool for clinicians to interpret ED patients with chest pain and suspected AAS and TAA. The integration of such imaging tools into ED could be considered in the future to enhance the diagnostic workflow for clinically fatal diseases. Full article

(This article belongs to the Topic AI in Medical Imaging and Image Processing)

►▼ Show Figures

Figure 1

Figure 1
<p>The flow chart of the study.</p> Full article ">Figure 2
<p>The images on the left (<b>A</b>,<b>B</b>) show the location of a possible aortic lesion on the CXR marked by the CNN model (VGG19), and these areas are colored by CAM to make it easy for users to interpret them. The pictures on the right (<b>C</b>,<b>D</b>) are the CTA images of the same patient showing the true location of the patient’s aortic lesion. After comparing with the CAM image, it was found that the CNN model correctly interpreted the location of the lesion on the CXRs, arrow: false lumen of aortic dissection.</p> Full article ">Figure 3
<p>Image enhancement and ROI selection of original images for better image quality and normalization before model training.</p> Full article ">Figure 4
<p>Image augmentation: original images were rotated by 3 degrees clockwise and counterclockwise.</p> Full article ">Figure 5
<p>(<b>A1</b>–<b>A3</b>) The CAM images of the original CXRs interpreted by the CNN model show that some feature extraction areas are not logically related to the aortic structure. In contrast, (<b>B1</b>–<b>B3</b>) displays the results after the ACF-generated feature extraction model processed the images. Training the CNN model with ROI-focused images shows that the model focuses more on the anatomical structures of the mediastinum, without being distracted by surrounding organs.</p> Full article ">

attachment

Supplementary material:
Supplementary File 1 (ZIP, 253 KiB)

10 pages, 4375 KiB  

Open AccessArticle

Deep Learning in Cardiothoracic Ratio Calculation and Cardiomegaly Detection

by Jakub Kufel, Iga Paszkiewicz, Szymon Kocot, Anna Lis, Piotr Dudek, Łukasz Czogalik, Michał Janik, Katarzyna Bargieł-Łączek, Wiktoria Bartnikowska, Maciej Koźlik, Maciej Cebula, Katarzyna Gruszczyńska and Zbigniew Nawrat

J. Clin. Med. 2024, 13(14), 4180; https://doi.org/10.3390/jcm13144180 - 17 Jul 2024

Viewed by 955

Abstract

Objectives: The purpose of this study is to evaluate the performance of our deep learning algorithm in calculating cardiothoracic ratio (CTR) and thus in the assessment of cardiomegaly or pericardial effusion occurrences on chest radiography (CXR). Methods: From a database of [...] Read more.

Objectives: The purpose of this study is to evaluate the performance of our deep learning algorithm in calculating cardiothoracic ratio (CTR) and thus in the assessment of cardiomegaly or pericardial effusion occurrences on chest radiography (CXR). Methods: From a database of 8000 CXRs, 13 folders with a comparable number of images were created. Then, 1020 images were chosen randomly, in proportion to the number of images in each folder. Afterward, CTR was calculated using RadiAnt Digital Imaging and Communications in Medicine (DICOM) Viewer software (2023.1). Next, heart and lung anatomical areas were marked in 3D Slicer. From these data, we trained an AI model which segmented heart and lung anatomy and determined the CTR value. Results: Our model achieved an Intersection over Union metric of 88.28% for the augmented training subset and 83.06% for the validation subset. F1-score for subsets were accordingly 90.22% and 90.67%. In the comparative analysis of artificial intelligence (AI) vs. humans, significantly lower transverse thoracic diameter (TTD) (p < 0.001), transverse cardiac diameter (TCD) (p < 0.001), and CTR (p < 0.001) values obtained using the neural network were observed. Conclusions: Results confirm that there is a significant correlation between the measurements made by human observers and the neural network. After validation in clinical conditions, our method may be used as a screening test or advisory tool when a specialist is not available, especially on Intensive Care Units (ICUs) or Emergency Departments (ERs) where time plays a key role. Full article

(This article belongs to the Topic AI in Medical Imaging and Image Processing)

►▼ Show Figures

Figure 1

Figure 1
<p>How measurements were manually determined using RadiAnt: blue vertical line—a line drawn through the spinous processes of the thoracic spine vertebrae; a—right side of the heart; b—left side of the heart; a + b = the widest dimension of the heart; c—the widest dimension of the chest.</p> Full article ">Figure 2
<p>How annotations were applied to the anatomical areas of the heart (green) and lungs (purple).</p> Full article ">Figure 3
<p>Metrics, loss, and learning rate during neural network training.</p> Full article ">Figure 4
<p>Difference in mean obtained by the humans observers and AI in the cardiothoracic ratio.</p> Full article ">Figure 5
<p>Difference between the human observers in the cardiothoracic ratio.</p> Full article ">Figure 6
<p>Correlation of the mean obtained by the human observers and AI for the cardiothoracic ratio.</p> Full article ">

11 pages, 1014 KiB  

Open AccessArticle

Optimizing GPT-4 Turbo Diagnostic Accuracy in Neuroradiology through Prompt Engineering and Confidence Thresholds

by Akihiko Wada, Toshiaki Akashi, George Shih, Akifumi Hagiwara, Mitsuo Nishizawa, Yayoi Hayakawa, Junko Kikuta, Keigo Shimoji, Katsuhiro Sano, Koji Kamagata, Atsushi Nakanishi and Shigeki Aoki

Diagnostics 2024, 14(14), 1541; https://doi.org/10.3390/diagnostics14141541 - 17 Jul 2024

Cited by 1 | Viewed by 1052

Abstract

Background and Objectives: Integrating large language models (LLMs) such as GPT-4 Turbo into diagnostic imaging faces a significant challenge, with current misdiagnosis rates ranging from 30–50%. This study evaluates how prompt engineering and confidence thresholds can improve diagnostic accuracy in neuroradiology. Methods: We [...] Read more.

Background and Objectives: Integrating large language models (LLMs) such as GPT-4 Turbo into diagnostic imaging faces a significant challenge, with current misdiagnosis rates ranging from 30–50%. This study evaluates how prompt engineering and confidence thresholds can improve diagnostic accuracy in neuroradiology. Methods: We analyze 751 neuroradiology cases from the American Journal of Neuroradiology using GPT-4 Turbo with customized prompts to improve diagnostic precision. Results: Initially, GPT-4 Turbo achieved a baseline diagnostic accuracy of 55.1%. By reformatting responses to list five diagnostic candidates and applying a 90% confidence threshold, the highest precision of the diagnosis increased to 72.9%, with the candidate list providing the correct diagnosis at 85.9%, reducing the misdiagnosis rate to 14.1%. However, this threshold reduced the number of cases that responded. Conclusions: Strategic prompt engineering and high confidence thresholds significantly reduce misdiagnoses and improve the precision of the LLM diagnostic in neuroradiology. More research is needed to optimize these approaches for broader clinical implementation, balancing accuracy and utility. Full article

(This article belongs to the Topic AI in Medical Imaging and Image Processing)

►▼ Show Figures

Figure 1

Figure 1
<p>Proportion of total cases by disease category. This pie chart shows the distribution of 751 clinical cases in various disease categories.</p> Full article ">Figure 2
<p>Evaluation of GPT-4 Turbo responses in 751 cases of neuroradiology.</p> Full article ">Figure 3
<p>Impact of confidence thresholds on diagnostic performance.</p> Full article ">

9 pages, 787 KiB  

Open AccessBrief Report

Enhanced Diagnostic Precision: Assessing Tumor Differentiation in Head and Neck Squamous Cell Carcinoma Using Multi-Slice Spiral CT Texture Analysis

by Lays Assolini Pinheiro de Oliveira, Diana Lorena Garcia Lopes, João Pedro Perez Gomes, Rafael Vinicius da Silveira, Daniel Vitor Aguiar Nozaki, Lana Ferreira Santos, Gabriela Castellano, Sérgio Lúcio Pereira de Castro Lopes and Andre Luiz Ferreira Costa

J. Clin. Med. 2024, 13(14), 4038; https://doi.org/10.3390/jcm13144038 - 10 Jul 2024

Viewed by 965

Abstract

This study explores the efficacy of texture analysis by using preoperative multi-slice spiral computed tomography (MSCT) to non-invasively determine the grade of cellular differentiation in head and neck squamous cell carcinoma (HNSCC). In a retrospective study, MSCT scans of patients with HNSCC were [...] Read more.

This study explores the efficacy of texture analysis by using preoperative multi-slice spiral computed tomography (MSCT) to non-invasively determine the grade of cellular differentiation in head and neck squamous cell carcinoma (HNSCC). In a retrospective study, MSCT scans of patients with HNSCC were analyzed and classified based on its histological grade as moderately differentiated, well-differentiated, or poorly differentiated. The location of the tumor was categorized as either in the bone or in soft tissues. Segmentation of the lesion areas was conducted, followed by texture analysis. Eleven GLCM parameters across five different distances were calculated. Median values and correlations of texture parameters were examined in relation to tumor differentiation grade by using Spearman’s correlation coefficient and Kruskal–Wallis and Dunn tests. Forty-six patients were included, predominantly female (87%), with a mean age of 66.7 years. Texture analysis revealed significant parameter correlations with histopathological grades of tumor differentiation. The study identified no significant age correlation with tumor differentiation, which underscores the potential of texture analysis as an age-independent biomarker. The strong correlations between texture parameters and histopathological grades support the integration of this technique into the clinical decision-making process. Full article

(This article belongs to the Topic AI in Medical Imaging and Image Processing)

►▼ Show Figures

Figure 1

Figure 1
<p>(<b>A</b>) Axial MSCT image shows squamous cell carcinoma located in the tonsillar pillar on the left side. (<b>B</b>) 3D VOI (red) segmentation process covering the lesion.</p> Full article ">Figure 2
<p>The groups are statistically different at 5% significance level regarding parameters followed by asterisk (*). All parameters were rescaled to be inserted into the graph in order to avoid distorting the scale for the other parameters.</p> Full article ">

27 pages, 12614 KiB  

Open AccessArticle

Attention-Based Deep Learning Approach for Breast Cancer Histopathological Image Multi-Classification

by Lama A. Aldakhil, Haifa F. Alhasson and Shuaa S. Alharbi

Diagnostics 2024, 14(13), 1402; https://doi.org/10.3390/diagnostics14131402 - 1 Jul 2024

Viewed by 1251

Abstract

Breast cancer diagnosis from histopathology images is often time consuming and prone to human error, impacting treatment and prognosis. Deep learning diagnostic methods offer the potential for improved accuracy and efficiency in breast cancer detection and classification. However, they struggle with limited data [...] Read more.

Breast cancer diagnosis from histopathology images is often time consuming and prone to human error, impacting treatment and prognosis. Deep learning diagnostic methods offer the potential for improved accuracy and efficiency in breast cancer detection and classification. However, they struggle with limited data and subtle variations within and between cancer types. Attention mechanisms provide feature refinement capabilities that have shown promise in overcoming such challenges. To this end, this paper proposes the Efficient Channel Spatial Attention Network (ECSAnet), an architecture built on EfficientNetV2 and augmented with a convolutional block attention module (CBAM) and additional fully connected layers. ECSAnet was fine-tuned using the BreakHis dataset, employing Reinhard stain normalization and image augmentation techniques to minimize overfitting and enhance generalizability. In testing, ECSAnet outperformed AlexNet, DenseNet121, EfficientNetV2-S, InceptionNetV3, ResNet50, and VGG16 in most settings, achieving accuracies of 94.2% at 40×, 92.96% at 100×, 88.41% at 200×, and 89.42% at 400× magnifications. The results highlight the effectiveness of CBAM in improving classification accuracy and the importance of stain normalization for generalizability. Full article

(This article belongs to the Topic AI in Medical Imaging and Image Processing)

►▼ Show Figures

Figure 1

Figure 1
<p>Workflow of the approach employed for breast histology image classification. Our approach begins with dataset processing steps, followed by training the ECSAnet model to extract features and obtain classification predictions, after which we move on to model evaluation and analysis steps.</p> Full article ">Figure 2
<p>Basic architecture of a CNN [<a href="#B28-diagnostics-14-01402" class="html-bibr">28</a>].</p> Full article ">Figure 3
<p>Architecture of EfficientNetV2 [<a href="#B17-diagnostics-14-01402" class="html-bibr">17</a>].</p> Full article ">Figure 4
<p>Computation processes of the attention modules: (<b>a</b>) CBAM, (<b>b</b>) CAM, and (<b>c</b>) SAM [<a href="#B18-diagnostics-14-01402" class="html-bibr">18</a>].</p> Full article ">Figure 5
<p>Architecture of ECSAnet.</p> Full article ">Figure 6
<p>Structures of MBConv and Fused-MBConv [<a href="#B17-diagnostics-14-01402" class="html-bibr">17</a>,<a href="#B30-diagnostics-14-01402" class="html-bibr">30</a>,<a href="#B31-diagnostics-14-01402" class="html-bibr">31</a>].</p> Full article ">Figure 7
<p>(<b>a</b>) Samples for benign tumor tissues; (<b>b</b>) samples for malignant tumor tissues [<a href="#B32-diagnostics-14-01402" class="html-bibr">32</a>].</p> Full article ">Figure 8
<p>Demonstration of Reinhard stain normalization on breast histopathology images. (<b>a</b>) Source images before normalization. (<b>b</b>) The target image providing the reference color distribution. (<b>c</b>) Images after applying the Reinhard method for color normalization.</p> Full article ">Figure 9
<p>Demonstration of implemented data augmentations. Column (<b>a</b>) shows the images in their original state. Column (<b>b</b>) shows the images after stain normalization. Columns (<b>c</b>,<b>d</b>) show the images after horizontal and vertical flips, respectively. Column (<b>e</b>) shows the images after random affine transformations, which include scaling, rotation, and translation. Lastly, column (<b>f</b>) shows the images after AugMix augmentations.</p> Full article ">Figure 10
<p>ECSAnet performance on the training set across magnification factors.</p> Full article ">Figure 11
<p>ECSAnet performance on the validation set across magnification factors.</p> Full article ">Figure 12
<p>Test set confusion matrices for ECSAnet across magnification factors.</p> Full article ">Figure 13
<p>Regions of interest identified by the model using Grad-CAM. (<b>a</b>) shows example images of benign classes while (<b>b</b>) shows examples for the malignant classes.</p> Full article ">Figure 14
<p>Comparison of the ECSAnet’s performance curves against variations with removed elements.</p> Full article ">Figure 15
<p>ECSAnet validation performance compared to other state-of-the-art models on the 40× magnification.</p> Full article ">Figure 16
<p>ECSAnet validation performance compared to other state-of-the-art models on the 100× magnification.</p> Full article ">Figure 17
<p>ECSAnet validation performance compared to other state-of-the-art models on the 200× magnification.</p> Full article ">Figure 18
<p>ECSAnet validation performance compared to other state-of-the-art models on the 400× magnification.</p> Full article ">Figure 19
<p>Grad-CAM visual explanations, illustrating the predictive focus areas for each class across models, descending vertically from the least accurate (AlexNet) to the most accurate (ECSAnet).</p> Full article ">

15 pages, 5298 KiB  

Open AccessArticle

Development and Validation of an Artificial Intelligence Model for Detecting Rib Fractures on Chest Radiographs

by Kaehong Lee, Sunhee Lee, Ji Soo Kwak, Heechan Park, Hoonji Oh and Jae Chul Koh

J. Clin. Med. 2024, 13(13), 3850; https://doi.org/10.3390/jcm13133850 - 30 Jun 2024

Viewed by 1008

Abstract

Background: Chest radiography is the standard method for detecting rib fractures. Our study aims to develop an artificial intelligence (AI) model that, with only a relatively small amount of training data, can identify rib fractures on chest radiographs and accurately mark their [...] Read more.

Background: Chest radiography is the standard method for detecting rib fractures. Our study aims to develop an artificial intelligence (AI) model that, with only a relatively small amount of training data, can identify rib fractures on chest radiographs and accurately mark their precise locations, thereby achieving a diagnostic accuracy comparable to that of medical professionals. Methods: For this retrospective study, we developed an AI model using 540 chest radiographs (270 normal and 270 with rib fractures) labeled for use with Detectron2 which incorporates a faster region-based convolutional neural network (R-CNN) enhanced with a feature pyramid network (FPN). The model’s ability to classify radiographs and detect rib fractures was assessed. Furthermore, we compared the model’s performance to that of 12 physicians, including six board-certified anesthesiologists and six residents, through an observer performance test. Results: Regarding the radiographic classification performance of the AI model, the sensitivity, specificity, and area under the receiver operating characteristic curve (AUROC) were 0.87, 0.83, and 0.89, respectively. In terms of rib fracture detection performance, the sensitivity, false-positive rate, and free-response receiver operating characteristic (JAFROC) figure of merit (FOM) were 0.62, 0.3, and 0.76, respectively. The AI model showed no statistically significant difference in the observer performance test compared to 11 of 12 and 10 of 12 physicians, respectively. Conclusions: We developed an AI model trained on a limited dataset that demonstrated a rib fracture classification and detection performance comparable to that of an experienced physician. Full article

(This article belongs to the Topic AI in Medical Imaging and Image Processing)

►▼ Show Figures

Figure 1

Figure 1
<p>Study flowchart for the development and evaluation of an artificial intelligence model for rib fracture detection and localization on chest radiographs. AI = Artificial Intelligence, VGG = Visual Geometry Group, Faster R-CNN = Faster Region-Convolutional Neural Network.</p> Full article ">Figure 2
<p>Annotation and preprocessing of rib fracture radiographs for AI training (<b>A</b>) Annotated rib fractures placed in the middle of the bounding box using Visual Geometry Group (VGG) annotator. (<b>B</b>) Converted region of interest as a bounding box to be prepared for artificial intelligence training.</p> Full article ">Figure 3
<p>AI-powered rib fracture detection system interface and diagnostic output (<b>A</b>) The initial interface of the program is designed for convenient execution in a clinical setting. (<b>B</b>) A user-friendly interface is designed to facilitate easy selection of a file/image to be analyzed. (<b>C</b>) The original chest radiograph before the site of the fracture is identified. (<b>D</b>) The image with the site of the rib fracture identified and marked by the AI model. AI: Artificial Intelligence.</p> Full article ">Figure 4
<p>Examples of rib fractures detected on chest radiographs using the AI model (<b>A</b>) Single rib fracture: The AI model’s correct identification of a rib fracture is demonstrated on the left lateral side, marked with a predicted box and assigned a 100% probability of fracture. (<b>B</b>) Single rib fracture in low resolution shows the AI model’s capacity to precisely mark a rib fracture location on a low-resolution image. (<b>C</b>) Multiple rib fractures: This image displays the AI model’s marking of several rib fracture sites, with each predicted box accompanied by a respective estimated probability of fracture. (<b>D</b>) Accurate and erroneous marking: This image features the AI model’s accurate marking of a rib fracture on the left side and a false-positive marking on the right side. Additionally, it includes an unintended marking of a clavicle fracture, not originally part of the training data. AI: Artificial Intelligence.</p> Full article ">Figure 5
<p>Classification performance of the AI model with machine learning training iterations. After reaching 10,000 training iterations, improvement in both the accuracy and the false-negative ratio tends to plateau.</p> Full article ">

21 pages, 5833 KiB  

Open AccessArticle

Developing the Benchmark: Establishing a Gold Standard for the Evaluation of AI Caries Diagnostics

by Julian Boldt, Matthias Schuster, Gabriel Krastl, Marc Schmitter, Jonas Pfundt, Angelika Stellzig-Eisenhauer and Felix Kunz

J. Clin. Med. 2024, 13(13), 3846; https://doi.org/10.3390/jcm13133846 - 29 Jun 2024

Viewed by 1016

Abstract

Background/Objectives: The aim of this study was to establish a histology-based gold standard for the evaluation of artificial intelligence (AI)-based caries detection systems on proximal surfaces in bitewing images. Methods: Extracted human teeth were used to simulate intraoral situations, including caries-free [...] Read more.

Background/Objectives: The aim of this study was to establish a histology-based gold standard for the evaluation of artificial intelligence (AI)-based caries detection systems on proximal surfaces in bitewing images. Methods: Extracted human teeth were used to simulate intraoral situations, including caries-free teeth, teeth with artificially created defects and teeth with natural proximal caries. All 153 simulations were radiographed from seven angles, resulting in 1071 in vitro bitewing images. Histological examination of the carious lesion depth was performed twice by an expert. A total of thirty examiners analyzed all the radiographs for caries. Results: We generated in vitro bitewing images to evaluate the performance of AI-based carious lesion detection against a histological gold standard. All examiners achieved a sensitivity of 0.565, a Matthews correlation coefficient (MCC) of 0.578 and an area under the curve (AUC) of 76.1. The histology receiver operating characteristic (ROC) curve significantly outperformed the examiners’ ROC curve (p < 0.001). All examiners distinguished induced defects from true caries in 54.6% of cases and correctly classified 99.8% of all teeth. Expert caries classification of the histological images showed a high level of agreement (intraclass correlation coefficient (ICC) = 0.993). Examiner performance varied with caries depth (p ≤ 0.008), except between E2 and E1 lesions (p = 1), while central beam eccentricity, gender, occupation and experience had no significant influence (all p ≥ 0.411). Conclusions: This study successfully established an unbiased dataset to evaluate AI-based caries detection on bitewing surfaces and compare it to human judgement, providing a standardized assessment for fair comparison between AI technologies and helping dental professionals to select reliable diagnostic tools. Full article

(This article belongs to the Topic AI in Medical Imaging and Image Processing)

►▼ Show Figures

Figure 1

Figure 1
<p>Trial profile.</p> Full article ">Figure 2
<p>Photographic and radiological documentation of all teeth.</p> Full article ">Figure 3
<p>Occlusion holder with fixed teeth simulating clinical bitewing scenarios.</p> Full article ">Figure 4
<p>Digital in vitro bitewing images. Top: color-coded setup—yellow: examination tooth, red: carious lesion, blue: adjacent tooth, green: antagonistic tooth. Below: The mesial-eccentric series shows increased superimposition as the ray path becomes increasingly eccentric in the proximal region of teeth 46 and 47. Conversely, the distal-eccentric series shows increased superimposition as the ray path becomes increasingly eccentric in the interproximal region of teeth 15 and 16.</p> Full article ">Figure 5
<p>Sample preparation steps.</p> Full article ">Figure 6
<p>Equipment for the preparation of histological specimens. (<b>A</b>): EXAKT-HISTOLUX light polymerization unit, (<b>B</b>) EXAKT precision vacuum bonding press, (<b>C</b>): Thermo Heraeus B6060 Incubator, (<b>D</b>): EXAKT band saw, (<b>E</b>): EXAKT horizontal microgrinding system, (<b>F</b>): EXAKT micrometer screw.</p> Full article ">Figure 7
<p>Histological specimen with different proximal carious lesion depths. E0 = Caries-free, E1 = Caries limited to the outer half of the enamel, E2 = Caries extending to the inner half of the enamel, D1 = Caries in the outer third of dentin, D2 = Caries in the middle third of dentin, D3 = Caries in the dentinal third close to the pulp or up to the pulp.</p> Full article ">Figure 8
<p>AUC of examiners and histology.</p> Full article ">Figure 9
<p>MCC by lesion class.</p> Full article ">Figure 10
<p>MCC by gender.</p> Full article ">Figure 11
<p>MCC by occupation.</p> Full article ">Figure 12
<p>MCC by experience.</p> Full article ">Figure 13
<p>MCC by eccentricity angle. m = mesial-eccentric, d = distal-eccentric.</p> Full article ">Figure 14
<p>Assessment of artificially induced lesions.</p> Full article ">Figure 15
<p>Tooth classification according to the FDI scheme.</p> Full article ">

19 pages, 7513 KiB  

Open AccessArticle

Augmenting Radiological Diagnostics with AI for Tuberculosis and COVID-19 Disease Detection: Deep Learning Detection of Chest Radiographs

by Manjur Kolhar, Ahmed M. Al Rajeh and Raisa Nazir Ahmed Kazi

Diagnostics 2024, 14(13), 1334; https://doi.org/10.3390/diagnostics14131334 - 24 Jun 2024

Viewed by 1012

Abstract

In this research, we introduce a network that can identify pneumonia, COVID-19, and tuberculosis using X-ray images of patients’ chests. The study emphasizes tuberculosis, COVID-19, and healthy lung conditions, discussing how advanced neural networks, like VGG16 and ResNet50, can improve the detection of [...] Read more.

In this research, we introduce a network that can identify pneumonia, COVID-19, and tuberculosis using X-ray images of patients’ chests. The study emphasizes tuberculosis, COVID-19, and healthy lung conditions, discussing how advanced neural networks, like VGG16 and ResNet50, can improve the detection of lung issues from images. To prepare the images for the model’s input requirements, we enhanced them through data augmentation techniques for training purposes. We evaluated the model’s performance by analyzing the precision, recall, and F1 scores across training, validation, and testing datasets. The results show that the ResNet50 model outperformed VGG16 with accuracy and resilience. It displayed superior ROC AUC values in both validation and test scenarios. Particularly impressive were ResNet50’s precision and recall rates, nearing 0.99 for all conditions in the test set. On the hand, VGG16 also performed well during testing—detecting tuberculosis with a precision of 0.99 and a recall of 0.93. Our study highlights the performance of our deep learning method by showcasing the effectiveness of ResNet50 over traditional approaches like VGG16. This progress utilizes methods to enhance classification accuracy by augmenting data and balancing them. This positions our approach as an advancement in using state-of-the-art deep learning applications in imaging. By enhancing the accuracy and reliability of diagnosing ailments such as COVID-19 and tuberculosis, our models have the potential to transform care and treatment strategies, highlighting their role in clinical diagnostics. Full article

(This article belongs to the Topic AI in Medical Imaging and Image Processing)

►▼ Show Figures

Figure 1

Figure 1
<p>Working model under training and validation along with data preparation and augmentation.</p> Full article ">Figure 2
<p>Training and Validation result in terms of loss and accuracy of ResNet50 model.</p> Full article ">Figure 3
<p>Train and validation confusion matrices for the ResNet50 model at the last epoch.</p> Full article ">Figure 4
<p>Receiver operating characteristic (ROC) curves for the ResNet50 model in diagnosing lung conditions, on validation dataset.</p> Full article ">Figure 5
<p>Receiver operating characteristic (ROC) curves for the ResNet50 model in diagnosing lung conditions, on test dataset.</p> Full article ">Figure 6
<p>Confusion matrix of testing optimized using augmentation techniques.</p> Full article ">Figure 7
<p>Confusion matrix of validation optimized using augmentation techniques.</p> Full article ">Figure 8
<p>Confusion matrix of test dataset optimized using augmentation techniques.</p> Full article ">Figure 9
<p>Receiver operating characteristic (ROC) curves for the VGG16 model in diagnosing lung conditions, on test dataset.</p> Full article ">Figure 10
<p>Comparison of AUC results for state-of-the-art lung disease detection models [<a href="#B9-diagnostics-14-01334" class="html-bibr">9</a>,<a href="#B10-diagnostics-14-01334" class="html-bibr">10</a>,<a href="#B11-diagnostics-14-01334" class="html-bibr">11</a>,<a href="#B12-diagnostics-14-01334" class="html-bibr">12</a>,<a href="#B13-diagnostics-14-01334" class="html-bibr">13</a>,<a href="#B14-diagnostics-14-01334" class="html-bibr">14</a>,<a href="#B15-diagnostics-14-01334" class="html-bibr">15</a>,<a href="#B16-diagnostics-14-01334" class="html-bibr">16</a>,<a href="#B17-diagnostics-14-01334" class="html-bibr">17</a>,<a href="#B18-diagnostics-14-01334" class="html-bibr">18</a>,<a href="#B19-diagnostics-14-01334" class="html-bibr">19</a>].</p> Full article ">

14 pages, 3006 KiB  

Open AccessArticle

Lung Disease Detection Using U-Net Feature Extractor Cascaded by Graph Convolutional Network

by Pshtiwan Qader Rashid and İlker Türker

Diagnostics 2024, 14(12), 1313; https://doi.org/10.3390/diagnostics14121313 - 20 Jun 2024

Viewed by 1136

Abstract

Computed tomography (CT) scans have recently emerged as a major technique for the fast diagnosis of lung diseases via image classification techniques. In this study, we propose a method for the diagnosis of COVID-19 disease with improved accuracy by utilizing graph convolutional networks [...] Read more.

Computed tomography (CT) scans have recently emerged as a major technique for the fast diagnosis of lung diseases via image classification techniques. In this study, we propose a method for the diagnosis of COVID-19 disease with improved accuracy by utilizing graph convolutional networks (GCN) at various layer formations and kernel sizes to extract features from CT scan images. We apply a U-Net model to aid in segmentation and feature extraction. In contrast with previous research retrieving deep features from convolutional filters and pooling layers, which fail to fully consider the spatial connectivity of the nodes, we employ GCNs for classification and prediction to capture spatial connectivity patterns, which provides a significant association benefit. We handle the extracted deep features to form an adjacency matrix that contains a graph structure and pass it to a GCN along with the original image graph and the largest kernel graph. We combine these graphs to form one block of the graph input and then pass it through a GCN with an additional dropout layer to avoid overfitting. Our findings show that the suggested framework, called the feature-extracted graph convolutional network (FGCN), performs better in identifying lung diseases compared to recently proposed deep learning architectures that are not based on graph representations. The proposed model also outperforms a variety of transfer learning models commonly used for medical diagnosis tasks, highlighting the abstraction potential of the graph representation over traditional methods. Full article

(This article belongs to the Topic AI in Medical Imaging and Image Processing)

►▼ Show Figures

Figure 1

Figure 1
<p>(<b>a</b>) COVID-19 and (<b>b</b>) non-COVID-19 samples [<a href="#B32-diagnostics-14-01313" class="html-bibr">32</a>].</p> Full article ">Figure 2
<p>Block diagram of the proposed method.</p> Full article ">Figure 3
<p>Feature extraction in U-Net architecture.</p> Full article ">Figure 4
<p>Directed graph example and its accompanying adjacency matrix modified from an image.</p> Full article ">Figure 5
<p>Evaluation of graph input data with a graph convolutional network.</p> Full article ">Figure 6
<p>Confusion matrices for the models tested.</p> Full article ">

10 pages, 1068 KiB  

Open AccessArticle

Breast Cancer Screening Using Inverse Modeling of Surface Temperatures and Steady-State Thermal Imaging

by Nithya Sritharan, Carlos Gutierrez, Isaac Perez-Raya, Jose-Luis Gonzalez-Hernandez, Alyssa Owens, Donnette Dabydeen, Lori Medeiros, Satish Kandlikar and Pradyumna Phatak

Cancers 2024, 16(12), 2264; https://doi.org/10.3390/cancers16122264 - 19 Jun 2024

Viewed by 952

Abstract

Cancer is characterized by increased metabolic activity and vascularity, leading to temperature changes in cancerous tissues compared to normal cells. This study focused on patients with abnormal mammogram findings or a clinical suspicion of breast cancer, exclusively those confirmed by biopsy. Utilizing an [...] Read more.

Cancer is characterized by increased metabolic activity and vascularity, leading to temperature changes in cancerous tissues compared to normal cells. This study focused on patients with abnormal mammogram findings or a clinical suspicion of breast cancer, exclusively those confirmed by biopsy. Utilizing an ultra-high sensitivity thermal camera and prone patient positioning, we measured surface temperatures integrated with an inverse modeling technique based on heat transfer principles to predict malignant breast lesions. Involving 25 breast tumors, our technique accurately predicted all tumors, with maximum errors below 5 mm in size and less than 1 cm in tumor location. Predictive efficacy was unaffected by tumor size, location, or breast density, with no aberrant predictions in the contralateral normal breast. Infrared temperature profiles and inverse modeling using both techniques successfully predicted breast cancer, highlighting its potential in breast cancer screening. Full article

(This article belongs to the Topic AI in Medical Imaging and Image Processing)

►▼ Show Figures

Figure 1

Figure 1
<p>Illustration of infrared imaging procedure of (<b>a</b>) a breast under an imaging table in the prone position and (<b>b</b>) eight positions to extract multi-view infrared images. (<b>c</b>) Example of a patient’s infrared images obtained using infrared imaging in the prone position [<a href="#B52-cancers-16-02264" class="html-bibr">52</a>].</p> Full article ">Figure 2
<p>Flowchart of the developed breast cancer detection algorithm that utilizes thermal modeling and inverse modeling with patient-specific breast models and infrared surface temperatures. The inverse model provides a final assessment that indicates the presence or absence of a tumor heat source.</p> Full article ">Figure 3
<p>All 25 breast tumors with a comparison of the actual and estimated tumor size by histological subtype. Histology subtypes included invasive ductal carcinoma (IDC), invasive lobular carcinoma (ILC), ductal carcinoma in situ (DCIS), lobular carcinoma in situ (LCIS) and atypical ductal hyperplasia (ADH).</p> Full article ">

12 pages, 1621 KiB  

Open AccessReview

Artificial Intelligence in Coronary Artery Calcium Scoring

by Afolasayo A. Aromiwura and Dinesh K. Kalra

J. Clin. Med. 2024, 13(12), 3453; https://doi.org/10.3390/jcm13123453 - 13 Jun 2024

Viewed by 852

Abstract

Cardiovascular disease (CVD), particularly coronary heart disease (CHD), is the leading cause of death in the US, with a high economic impact. Coronary artery calcium (CAC) is a known marker for CHD and a useful tool for estimating the risk of atherosclerotic cardiovascular [...] Read more.

Cardiovascular disease (CVD), particularly coronary heart disease (CHD), is the leading cause of death in the US, with a high economic impact. Coronary artery calcium (CAC) is a known marker for CHD and a useful tool for estimating the risk of atherosclerotic cardiovascular disease (ASCVD). Although CACS is recommended for informing the decision to initiate statin therapy, the current standard requires a dedicated CT protocol, which is time-intensive and contributes to radiation exposure. Non-dedicated CT protocols can be taken advantage of to visualize calcium and reduce overall cost and radiation exposure; however, they mainly provide visual estimates of coronary calcium and have disadvantages such as motion artifacts. Artificial intelligence is a growing field involving software that independently performs human-level tasks, and is well suited for improving CACS efficiency and repurposing non-dedicated CT for calcium scoring. We present a review of the current studies on automated CACS across various CT protocols and discuss consideration points in clinical application and some barriers to implementation. Full article

(This article belongs to the Topic AI in Medical Imaging and Image Processing)

►▼ Show Figures

Graphical abstract

Graphical abstract
Full article ">Figure 1
<p>Overview of Artificial Intelligence in Coronary Artery Calcium Scoring. Automatic cardiac segmentation and coronary labeling are hallmarks of proposed models. Improvement of data diversity, data repositories, external validation, image artifacts, and metric standardization will enhance the proposed models. CACS automation optimizes workflow, reduces cost, and allows redistribution of resources in image post-processing labs.</p> Full article ">Figure 2
<p>Comparison of Manual and Automatic CAC Identification. Image adapted from Eng et al. [<a href="#B25-jcm-13-03453" class="html-bibr">25</a>] Manual and automatic CACS comparison in two patients. The automatic model excludes calcifications in the descending aorta and some in the aortic root. (<b>a</b>) False negative calcium analysis by the automatic model in the right coronary artery (red arrow). False positive aortic root calcification is present (blue circle). (<b>b</b>) Comparison of automatic and manual coronary calcifications in another patient.</p> Full article ">

attachment

Supplementary material:
Supplementary File 1 (ZIP, 473 KiB)

18 pages, 3096 KiB  

Open AccessArticle

Causal Forest Machine Learning Analysis of Parkinson’s Disease in Resting-State Functional Magnetic Resonance Imaging

by Gabriel Solana-Lavalle, Michael D. Cusimano, Thomas Steeves, Roberto Rosas-Romero and Pascal N. Tyrrell

Tomography 2024, 10(6), 894-911; https://doi.org/10.3390/tomography10060068 - 6 Jun 2024

Viewed by 1294

Abstract

In recent years, Artificial Intelligence has been used to assist healthcare professionals in detecting and diagnosing neurodegenerative diseases. In this study, we propose a methodology to analyze functional Magnetic Resonance Imaging signals and perform classification between Parkinson’s disease patients and healthy participants using [...] Read more.

In recent years, Artificial Intelligence has been used to assist healthcare professionals in detecting and diagnosing neurodegenerative diseases. In this study, we propose a methodology to analyze functional Magnetic Resonance Imaging signals and perform classification between Parkinson’s disease patients and healthy participants using Machine Learning algorithms. In addition, the proposed approach provides insights into the brain regions affected by the disease. The functional Magnetic Resonance Imaging from the PPMI and 1000-FCP datasets were pre-processed to extract time series from 200 brain regions per participant, resulting in 11,600 features. Causal Forest and Wrapper Feature Subset Selection algorithms were used for dimensionality reduction, resulting in a subset of features based on their heterogeneity and association with the disease. We utilized Logistic Regression and XGBoost algorithms to perform PD detection, achieving 97.6% accuracy, 97.5% F₁ score, 97.9% precision, and 97.7%recall by analyzing sets with fewer than 300 features in a population including men and women. Finally, Multiple Correspondence Analysis was employed to visualize the relationships between brain regions and each group (women with Parkinson, female controls, men with Parkinson, male controls). Associations between the Unified Parkinson’s Disease Rating Scale questionnaire results and affected brain regions in different groups were also obtained to show another use case of the methodology. This work proposes a methodology to (1) classify patients and controls with Machine Learning and Causal Forest algorithm and (2) visualize associations between brain regions and groups, providing high-accuracy classification and enhanced interpretability of the correlation between specific brain regions and the disease across different groups. Full article

(This article belongs to the Topic AI in Medical Imaging and Image Processing)

►▼ Show Figures

Figure 1

Figure 1
<p>Methodology for PD detection and subsequent interpreting analysis of PD with rs-fMRI.</p> Full article ">Figure 2
<p>PD detection performance in three populations: female, male, and mixed. For each population, PD detection was implemented using two classifiers, XGBoost and logistic regression (LR), using nested cross-validation. Four metrics were used to measure detection performance: accuracy, F1 score, precision and recall.</p> Full article ">Figure 3
<p>Bubble plot that shows brain region indexes (<span class="html-italic">X</span>-axis), region average rankings (<span class="html-italic">Y</span>-axis), and number brain features (bubble size and color intensity) for female PD patients (lower panel) and female controls (upper panel). The parcellation of the ROIs shown in the figure come from the Schaefer 2018 atlas [<a href="#B26-tomography-10-00068" class="html-bibr">26</a>].</p> Full article ">Figure 4
<p>Bubble plot that shows brain region indexes (<span class="html-italic">X</span>-axis), region average rankings (<span class="html-italic">Y</span>-axis), and number of region features (bubble size and color intensity), for male PD patients (lower panel) and male controls (upper panel). The parcellation of the regions shown in this figure comes from the Schaefer 2018 atlas [<a href="#B26-tomography-10-00068" class="html-bibr">26</a>].</p> Full article ">Figure 5
<p>Associations between brain regions and groups. Each point represents an association between a group and the regions. The two most important Principal Dimensions are shown, which account for 83.78% of the total inertia. Each axis accounts for a different amount of inertia, where the weight of the <span class="html-italic">X</span>-axis is higher than the one of the <span class="html-italic">Y</span>-axis. The parcellation of the regions shown in this figure comes from the Schaefer 2018 atlas [<a href="#B26-tomography-10-00068" class="html-bibr">26</a>]. The ROIs are shown as red areas inside the yellow dashed box.</p> Full article ">Figure 6
<p>MCA analysis shows the associations between the scores of the motor section of the UPDRS test and the most relevant ROI resulting from running CF with the data from the patients with PD from the PPMI. Four groups were studied, including Males and Females with positive results from the third section UPDRS and Males and Females with negative results from the motor test. The most relevant regions are shown in the brain illustrations. The two Principal Dimensions with the highest inertia are shown, which account for 89.51% of the total inertia. Since each axis accounts for different amount of inertia, the weight of the <span class="html-italic">X</span>-axis is bigger than the one of the <span class="html-italic">Y</span>-axis. The ROIs are shown as red areas inside the yellow dashed box.</p> Full article ">

21 pages, 10826 KiB  

Open AccessArticle

Breast Cancer Diagnosis Method Based on Cross-Mammogram Four-View Interactive Learning

by Xuesong Wen, Jianjun Li and Liyuan Yang

Tomography 2024, 10(6), 848-868; https://doi.org/10.3390/tomography10060065 - 1 Jun 2024

Viewed by 810

Abstract

Computer-aided diagnosis systems play a crucial role in the diagnosis and early detection of breast cancer. However, most current methods focus primarily on the dual-view analysis of a single breast, thereby neglecting the potentially valuable information between bilateral mammograms. In this paper, we [...] Read more.

Computer-aided diagnosis systems play a crucial role in the diagnosis and early detection of breast cancer. However, most current methods focus primarily on the dual-view analysis of a single breast, thereby neglecting the potentially valuable information between bilateral mammograms. In this paper, we propose a Four-View Correlation and Contrastive Joint Learning Network (FV-Net) for the classification of bilateral mammogram images. Specifically, FV-Net focuses on extracting and matching features across the four views of bilateral mammograms while maximizing both their similarities and dissimilarities. Through the Cross-Mammogram Dual-Pathway Attention Module, feature matching between bilateral mammogram views is achieved, capturing the consistency and complementary features across mammograms and effectively reducing feature misalignment. In the reconstituted feature maps derived from bilateral mammograms, the Bilateral-Mammogram Contrastive Joint Learning module performs associative contrastive learning on positive and negative sample pairs within each local region. This aims to maximize the correlation between similar local features and enhance the differentiation between dissimilar features across the bilateral mammogram representations. Our experimental results on a test set comprising 20% of the combined Mini-DDSM and Vindr-mamo datasets, as well as on the INbreast dataset, show that our model exhibits superior performance in breast cancer classification compared to competing methods. Full article

(This article belongs to the Topic AI in Medical Imaging and Image Processing)

►▼ Show Figures

Figure 1

Figure 1
<p>Examples from the training and test datasets of mammography images. (<b>a</b>,<b>b</b>) Vindr-mamo mammography images; (<b>c</b>,<b>d</b>) Mini-DDSM mammography images; (<b>e</b>,<b>f</b>) INbreast mammography images.</p> Full article ">Figure 2
<p>The data preprocessing procedure. (<b>a</b>) The original image; (<b>b</b>) the image processed with BRD; (<b>c</b>) the image processed with BRD and cropping; (<b>d</b>) the image processed with BRD, cropping, CLAHE, and truncated normalization.</p> Full article ">Figure 3
<p>(<b>a</b>) The left CC (cranio-caudal) view; (<b>b</b>) The left MLO (mediolateral oblique) view; (<b>c</b>) The right CC view; (<b>d</b>) The right MLO view. The areas within the red bounding boxes represent the lesion regions of the mammograms.</p> Full article ">Figure 4
<p>Our proposed FV-Net framework utilizes a shared modified truncated EfficientNet-b0 for feature extraction and employs the CMDPA to reshape the fea ture maps, along with the BMCJL module for four-view local feature joint contrastive learning.</p> Full article ">Figure 5
<p>Detailed architecture of Cross-Mammogram Dual-Pathway Attention Module, (heads) # L × C’ represents numerical units with the total number of heads and a size of L × C’, and the others are similarly defined.</p> Full article ">Figure 6
<p>ROC curves graph from ablation study on the public INbreast dataset.</p> Full article ">Figure 7
<p>A comparison of visualization results for the most suspicious malignant tumor areas in paired CC and MLO views of the same case predicted by different method groups. (<b>a</b>) Raw image; (<b>b</b>) Preprocessed Baseline; (<b>c</b>) Preprocessed Baseline Model with BMCJL; (<b>d</b>) FV-Net.</p> Full article ">Figure 8
<p>The ROC curves comparison of the classification performance of FV-Net and eight competing methods on the INbreast dataset.</p> Full article ">Figure 9
<p>The comparison of ROC curves for the classification performance of FV-Net and eight competing methods on a 20% independent test set.</p> Full article ">Figure 10
<p>A comparison of the confusion matrices for our method against other methods on the INbreast dataset. (<b>a</b>) ResNet50; (<b>b</b>) ResNet101; (<b>c</b>) EfficientNet-b0; (<b>d</b>) ConvNext_small; (<b>e</b>) EfficientNet-b3; (<b>f</b>) Reg- Net_x_1_6gf; (<b>g</b>) MobileNet-v2; (<b>h</b>) ResNext101_32x4d; (<b>i</b>) Ours.</p> Full article ">Figure 11
<p>Visual comparison of the most suspicious malignant lesion areas predicted by different models. (<b>a</b>) Raw image; (<b>b</b>) ResNet50; (<b>c</b>) ResNext101_32x4d; (<b>d</b>) EfficientNet-b0; (<b>e</b>) ConvNext_small; (<b>f</b>) Efficient- Net-b3; (<b>g</b>) RegNet_x_1_6gf; (<b>h</b>) MobileNet-v2; (<b>i</b>) ResNet101; (<b>j</b>) Ours.</p> Full article ">

17 pages, 17399 KiB  

Open AccessArticle

Evaluating Cellularity Estimation Methods: Comparing AI Counting with Pathologists’ Visual Estimates

by Tomoharu Kiyuna, Eric Cosatto, Kanako C. Hatanaka, Tomoyuki Yokose, Koji Tsuta, Noriko Motoi, Keishi Makita, Ai Shimizu, Toshiya Shinohara, Akira Suzuki, Emi Takakuwa, Yasunari Takakuwa, Takahiro Tsuji, Mitsuhiro Tsujiwaki, Mitsuru Yanai, Sayaka Yuzawa, Maki Ogura and Yutaka Hatanaka

Diagnostics 2024, 14(11), 1115; https://doi.org/10.3390/diagnostics14111115 - 28 May 2024

Viewed by 1077

Abstract

The development of next-generation sequencing (NGS) has enabled the discovery of cancer-specific driver gene alternations, making precision medicine possible. However, accurate genetic testing requires a sufficient amount of tumor cells in the specimen. The evaluation of tumor content ratio (TCR) from hematoxylin and [...] Read more.

The development of next-generation sequencing (NGS) has enabled the discovery of cancer-specific driver gene alternations, making precision medicine possible. However, accurate genetic testing requires a sufficient amount of tumor cells in the specimen. The evaluation of tumor content ratio (TCR) from hematoxylin and eosin (H&E)-stained images has been found to vary between pathologists, making it an important challenge to obtain an accurate TCR. In this study, three pathologists exhaustively labeled all cells in 41 regions from 41 lung cancer cases as either tumor, non-tumor or indistinguishable, thus establishing a “gold standard” TCR. We then compared the accuracy of the TCR estimated by 13 pathologists based on visual assessment and the TCR calculated by an AI model that we have developed. It is a compact and fast model that follows a fully convolutional neural network architecture and produces cell detection maps which can be efficiently post-processed to obtain tumor and non-tumor cell counts from which TCR is calculated. Its raw cell detection accuracy is 92% while its classification accuracy is 84%. The results show that the error between the gold standard TCR and the AI calculation was significantly smaller than that between the gold standard TCR and the pathologist’s visual assessment ($p<0.05$). Additionally, the robustness of AI models across institutions is a key issue and we demonstrate that the variation in AI was smaller than that in the average of pathologists when evaluated by institution. These findings suggest that the accuracy of tumor cellularity assessments in clinical workflows is significantly improved by the introduction of robust AI models, leading to more efficient genetic testing and ultimately to better patient outcomes. Full article

(This article belongs to the Topic AI in Medical Imaging and Image Processing)

►▼ Show Figures

Figure 1

Figure 1
<p>Region of interest (ROI) selection over WSI slide. In (<b>A</b>), a tissue is selected on the WSI slide. In (<b>B</b>), the tissue is tiled and one tile is selected as ROI. In (<b>C</b>), the cells of the ROI are exhaustively labeled.</p> Full article ">Figure 2
<p>The left panel shows the overall workflow of the TCR system. An RGB image patch is the input to the model, which outputs two density maps. Annotations marking the centroid of cells (also shown as green and red overlay dots on the input image) are used to generate the two target density maps (by drawing Gaussian peaks at the location of annotated cells) which are used to train the model by back-propagating the difference (loss) between them and the maps generated by the model. The right panel shows a graphical representation of the architecture of the U-Net fully convolutional model. The numbers illustrate a 572 × 572 input image. For an 800 × 800 input image and two output maps, the total number of parameters is 28,942,850 and the total size of the model for inference (not counting training gradients) is about 3 GB.</p> Full article ">Figure 3
<p>Example of data augmentation. The left image is the original, while the right image is generated by applying a gamma correction of 2.4.</p> Full article ">Figure 4
<p>The training curve shows the cross-entropy loss as a function of the epoch. The training loss is shown in red, and the blue curve shows the validation loss.</p> Full article ">Figure 5
<p>Example of tumor detection using AI. (<b>A</b>) Gold standard label. Yellow indicates tumor cells, cyan indicates non-tumor cells, and green indicates indistinguishable cells. (<b>B</b>) Predictions by the AI model. Yellow indicates tumor cells and cyan indicates non-tumor cells.</p> Full article ">Figure 6
<p>Box and whisker plot of tumor content ratio (TCR) estimated visually by 13 pathologists (yellow) together with predicted TCR by AI (solid blue line). The red solid line represent the gold standard TCR. Boxes show medians, quartiles, and outliers for each group. The boxes extend from the 25th to the 75th percentile, and the whiskers extend from the minimum to the maximum TCR within a group.</p> Full article ">Figure 7
<p>Comparison of mean absolute errors of pathologist and AI predictions to gold standard TCR.</p> Full article ">Figure 8
<p>Examples of ROIs with small and large errors of pathologist’s estimation (<b>A</b>,<b>B</b>) and AI’s prediction (<b>C</b>,<b>D</b>). Pathologists’ estimations ranged from 75 to 90% for <math display="inline"><semantics> <mrow> <mi>T</mi> <mi>C</mi> <msub> <mi>R</mi> <mrow> <mi>G</mi> <mi>S</mi> </mrow> </msub> <mo>=</mo> <mn>82.5</mn> <mo>%</mo> </mrow> </semantics></math> with a mean absolute error of 5.0% (<b>A</b>), and ranged from 20 to 80% for <math display="inline"><semantics> <mrow> <mi>T</mi> <mi>C</mi> <msub> <mi>R</mi> <mrow> <mi>G</mi> <mi>S</mi> </mrow> </msub> <mo>=</mo> <mn>82.6</mn> <mo>%</mo> </mrow> </semantics></math> with a mean absolute error of 29.3% (<b>B</b>). AI’s prediction was 64.3% for <math display="inline"><semantics> <mrow> <mi>T</mi> <mi>C</mi> <msub> <mi>R</mi> <mrow> <mi>G</mi> <mi>S</mi> </mrow> </msub> <mo>=</mo> <mn>98.4</mn> <mo>%</mo> </mrow> </semantics></math> with an absolute error of 0.2% (<b>C</b>) and 33.0% for <math display="inline"><semantics> <mrow> <mi>T</mi> <mi>C</mi> <msub> <mi>R</mi> <mrow> <mi>G</mi> <mi>S</mi> </mrow> </msub> <mo>=</mo> <mn>66.3</mn> <mo>%</mo> </mrow> </semantics></math> with an absolute error of 33.3% (<b>D</b>).</p> Full article ">Figure 9
<p>Scatter plot of the mean absolute error of the pathologist’s estimation and the AI’s prediction.</p> Full article ">Figure 10
<p>Visualization of mean absolute error of each pathologist and AI against the gold standard per site. The length of the line above the bar represents the standard deviation.</p> Full article ">

11 pages, 880 KiB  

Open AccessArticle

Anomaly Detection and Biomarkers Localization in Retinal Images

by Liran Tiosano, Ron Abutbul, Rivkah Lender, Yahel Shwartz, Itay Chowers, Yedid Hoshen and Jaime Levy

J. Clin. Med. 2024, 13(11), 3093; https://doi.org/10.3390/jcm13113093 - 24 May 2024

Viewed by 799

Abstract

Background: To design a novel anomaly detection and localization approach using artificial intelligence methods using optical coherence tomography (OCT) scans for retinal diseases. Methods: High-resolution OCT scans from the publicly available Kaggle dataset and a local dataset were used by four [...] Read more.

Background: To design a novel anomaly detection and localization approach using artificial intelligence methods using optical coherence tomography (OCT) scans for retinal diseases. Methods: High-resolution OCT scans from the publicly available Kaggle dataset and a local dataset were used by four state-of-the-art self-supervised frameworks. The backbone model of all the frameworks was a pre-trained convolutional neural network (CNN), which enabled the extraction of meaningful features from OCT images. Anomalous images included choroidal neovascularization (CNV), diabetic macular edema (DME), and the presence of drusen. Anomaly detectors were evaluated by commonly accepted performance metrics, including area under the receiver operating characteristic curve, F1 score, and accuracy. Results: A total of 25,315 high-resolution retinal OCT slabs were used for training. Test and validation sets consisted of 968 and 4000 slabs, respectively. The best performing across all anomaly detectors had an area under the receiver operating characteristic of 0.99. All frameworks were shown to achieve high performance and generalize well for the different retinal diseases. Heat maps were generated to visualize the quality of the frameworks’ ability to localize anomalous areas of the image. Conclusions: This study shows that with the use of pre-trained feature extractors, the frameworks tested can generalize to the domain of retinal OCT scans and achieve high image-level ROC-AUC scores. The localization results of these frameworks are promising and successfully capture areas that indicate the presence of retinal pathology. Moreover, such frameworks have the potential to uncover new biomarkers that are difficult for the human eye to detect. Frameworks for anomaly detection and localization can potentially be integrated into clinical decision support and automatic screening systems that will aid ophthalmologists in patient diagnosis, follow-up, and treatment design. This work establishes a solid basis for further development of automated anomaly detection frameworks for clinical use. Full article

(This article belongs to the Topic AI in Medical Imaging and Image Processing)

►▼ Show Figures

Figure 1

Figure 1
<p>Demonstration of local region anomaly detection workflow. The non-medical pre-trained neural network learns local region representations. Regions with similar properties are mapped to numerically similar representations, while dissimilar local retinal regions are mapped to dissimilar representations. The last row presents a test scan with choroidal neovascularization which the fluid representation differs from the present of the pigment epithelial detachment and drusen.</p> Full article ">Figure 2
<p>Heat maps of anomalous retinal OCT scans containing (<b>A</b>) choroidal neovascularization, (<b>B</b>) diabetic macular edema, and (<b>C</b>) drusen. Column 1 presents the original OCT scan. Heat-maps were extracted by local region using aligned density estimation (column 2) and local region with global density estimation (column 3).</p> Full article ">

16 pages, 1888 KiB  

Open AccessArticle

Prediction of Lymph Node Metastasis in T1 Colorectal Cancer Using Artificial Intelligence with Hematoxylin and Eosin-Stained Whole-Slide-Images of Endoscopic and Surgical Resection Specimens

by Joo Hye Song, Eun Ran Kim, Yiyu Hong, Insuk Sohn, Soomin Ahn, Seok-Hyung Kim and Kee-Taek Jang

Cancers 2024, 16(10), 1900; https://doi.org/10.3390/cancers16101900 - 16 May 2024

Viewed by 1059

Abstract

According to the current guidelines, additional surgery is performed for endoscopically resected specimens of early colorectal cancer (CRC) with a high risk of lymph node metastasis (LNM). However, the rate of LNM is 2.1–25.0% in cases treated endoscopically followed by surgery, indicating a [...] Read more.

According to the current guidelines, additional surgery is performed for endoscopically resected specimens of early colorectal cancer (CRC) with a high risk of lymph node metastasis (LNM). However, the rate of LNM is 2.1–25.0% in cases treated endoscopically followed by surgery, indicating a high rate of unnecessary surgeries. Therefore, this study aimed to develop an artificial intelligence (AI) model using H&E-stained whole slide images (WSIs) without handcrafted features employing surgically and endoscopically resected specimens to predict LNM in T1 CRC. To validate with an independent cohort, we developed a model with four versions comprising various combinations of training and test sets using H&E-stained WSIs from endoscopically (400 patients) and surgically resected specimens (881 patients): Version 1, Train and Test: surgical specimens; Version 2, Train and Test: endoscopic and surgically resected specimens; Version 3, Train: endoscopic and surgical specimens and Test: surgical specimens; Version 4, Train: endoscopic and surgical specimens and Test: endoscopic specimens. The area under the curve (AUC) of the receiver operating characteristic curve was used to determine the accuracy of the AI model for predicting LNM with a 5-fold cross-validation in the training set. Our AI model with H&E-stained WSIs and without annotations showed good performance power with the validation of an independent cohort in a single center. The AUC of our model was 0.758–0.830 in the training set and 0.781–0.824 in the test set, higher than that of previous AI studies with only WSI. Moreover, the AI model with Version 4, which showed the highest sensitivity (92.9%), reduced unnecessary additional surgery by 14.2% more than using the current guidelines (68.3% vs. 82.5%). This revealed the feasibility of using an AI model with only H&E-stained WSIs to predict LNM in T1 CRC. Full article

(This article belongs to the Topic AI in Medical Imaging and Image Processing)

►▼ Show Figures

Figure 1

Figure 1
<p>Pipeline of the approach for classifying lymph node metastasis. WSI, whole slide image.</p> Full article ">Figure 2
<p>Flow chart of study population. CCRT, concurrent chemoradiotherapy.</p> Full article ">Figure 3
<p>Area under the ROC curve for attention-based WSI-level classification deep learning model for predicting lymph node metastasis in T1 colorectal cancer. (<b>A</b>) Previous model, (<b>B</b>) Version 1, (<b>C</b>) Version 2, (<b>D</b>) Version 3, (<b>E</b>) Version 4 ROC, receiver operating characteristic; WSI, whole-slide image; Version 1, train and test: surgical specimen; Version 2, train and test: endoscopic and surgical specimen; Version 3, train: endoscopic and surgical specimens and test: surgical specimen; Version 4, train: endoscopic and surgical specimens and test: endoscopic specimen; AUC: area under the curve; Avg: average.</p> Full article ">Figure 4
<p>Attention score visualization in WSIs of positive lymph node metastasis. (<b>A</b>) Spatial distribution of attention scores on the WSIs, (<b>B</b>) Patch-images stratified by high attention scores WSIs, whole slide images.</p> Full article ">

18 pages, 2277 KiB  

Open AccessArticle

Revolutionizing Radiological Analysis: The Future of French Language Automatic Speech Recognition in Healthcare

by Mariem Jelassi, Oumaima Jemai and Jacques Demongeot

Diagnostics 2024, 14(9), 895; https://doi.org/10.3390/diagnostics14090895 - 25 Apr 2024

Viewed by 1164

Abstract

This study introduces a specialized Automatic Speech Recognition (ASR) system, leveraging the Whisper Large-v2 model, specifically adapted for radiological applications in the French language. The methodology focused on adapting the model to accurately transcribe medical terminology and diverse accents within the French language [...] Read more.

This study introduces a specialized Automatic Speech Recognition (ASR) system, leveraging the Whisper Large-v2 model, specifically adapted for radiological applications in the French language. The methodology focused on adapting the model to accurately transcribe medical terminology and diverse accents within the French language context, achieving a notable Word Error Rate (WER) of 17.121%. This research involved extensive data collection and preprocessing, utilizing a wide range of French medical audio content. The results demonstrate the system’s effectiveness in transcribing complex radiological data, underscoring its potential to enhance medical documentation efficiency in French-speaking clinical settings. The discussion extends to the broader implications of this technology in healthcare, including its potential integration with electronic health records (EHRs) and its utility in medical education. This study also explores future research directions, such as tailoring ASR systems to specific medical specialties and languages. Overall, this research contributes significantly to the field of medical ASR systems, presenting a robust tool for radiological transcription in the French language and paving the way for advanced technology-enhanced healthcare solutions. Full article

(This article belongs to the Topic AI in Medical Imaging and Image Processing)

►▼ Show Figures

Figure 1

Figure 1
<p>Comparative performance of the Whisper model in long-form transcription against leading ASR systems [<a href="#B26-diagnostics-14-00895" class="html-bibr">26</a>].</p> Full article ">Figure 2
<p>Audio file sample transcribed using Whisper Large-v2.</p> Full article ">Figure 3
<p>Exemplar list of technical terms.</p> Full article ">Figure 4
<p>System deployment and operational flowchart.</p> Full article ">Figure 5
<p>Loss curves.</p> Full article ">

14 pages, 2340 KiB  

Open AccessArticle

Classification of Osteophytes Occurring in the Lumbar Intervertebral Foramen

by Abdullah Emre Taçyıldız and Feyza İnceoğlu

Tomography 2024, 10(4), 618-631; https://doi.org/10.3390/tomography10040047 - 19 Apr 2024

Viewed by 1498

Abstract

Background: Surgeons have limited knowledge of the lumbar intervertebral foramina. This study aimed to classify osteophytes in the lumbar intervertebral foramen and to determine their pathoanatomical characteristics, discuss their potential biomechanical effects, and contribute to developing surgical methods. Methods: We conducted a retrospective, [...] Read more.

Background: Surgeons have limited knowledge of the lumbar intervertebral foramina. This study aimed to classify osteophytes in the lumbar intervertebral foramen and to determine their pathoanatomical characteristics, discuss their potential biomechanical effects, and contribute to developing surgical methods. Methods: We conducted a retrospective, non-randomized, single-center study involving 1224 patients. The gender, age, and anatomical location of the osteophytes in the lumbar intervertebral foramina of the patients were recorded. Results: Two hundred and forty-nine (20.34%) patients had one or more osteophytes in their lumbar 4 and 5 foramina. Of the 4896 foramina, 337 (6.88%) contained different types of osteophytes. Moreover, four anatomical types of osteophytes were found: mixed osteophytes in 181 (3.69%) foramina, osteophytes from the lower endplate of the superior vertebrae in 91 (1.85%) foramina, osteophytes from the junction of the pedicle and lamina of the upper vertebrae in 39 foramina (0.79%), and osteophytes from the upper endplate of the lower vertebrae in 26 (0.53%) foramina. The L4 foramen contained a significantly higher number of osteophytes than the L5 foramen. Osteophyte development increased significantly with age, with no difference between males and females. Conclusions: The findings show that osteophytic extrusions, which alter the natural anatomical structure of the lumbar intervertebral foramina, are common and can narrow the foramen. Full article

(This article belongs to the Topic AI in Medical Imaging and Image Processing)

►▼ Show Figures

Figure 1

Figure 1
<p>Mixed-type osteophytosis is observed. (<b>A</b>) An osteophyte (red arrow) arising from the pedicle–lamina junction and the inferior endplate of the superior vertebrae is observed in the L5–S1 foramen. Vertebral osteophytes (blue arrow) are also observed at a position anterior to the vertebrae. (<b>B</b>) Osteophytes (red arrow) arising from three different points united in the foramen. A mixed-type osteophyte arising from the pedicle–lamina junction, lower endplate of the superior vertebrae, and upper endplate of the inferior vertebrae formed an interesting pathoanatomical structure in the foramen. (<b>C</b>) A mixed-type osteophyte (blue arrow) arising from the pedicle–lamina junction in the L4–5 foramen and an osteophyte (red arrow) arising from the lower endplate of the superior vertebrae in the L5–S1 foramen are observed. (<b>D</b>) The L5–S1 disk space appears to have collapsed. A mixed-type osteophyte (red arrow) is seen to be arising from both the lower endplate of the superior vertebrae and the upper endplate of the inferior vertebrae. The fact that the spines are not in the same plane in sagittal images gives the impression of degenerative scoliosis.</p> Full article ">Figure 2
<p>Osteophytes arising from the lower endplate of the superior vertebrae are observed. (<b>A</b>) An osteophyte arising from the lower endplate of the superior vertebrae (red arrow) in the L5–S1 foramen. It is observed that the intervertebral foramen is narrowed. (<b>B</b>) An osteophyte arising from the lower endplate of the superior vertebrae (red arrow) in the L5–S1 foramen. (<b>C</b>) In the L5–S1 foramen, the osteophyte arising from the inferior endplate of the superior vertebrae (red arrow) almost merges with the superior articular facet. The disk space is narrowed, and the vacuum phenomenon (blue arrow) is observed. (<b>D</b>) An osteophyte (red arrow) extending cranially from the inferior endplate of the superior vertebrae potentially narrows the foraminal space. Vacuum phenomenon (blue arrow) is observed.</p> Full article ">Figure 3
<p>In the L5 (L5–S1) foramina, osteophytes arising from the pedicle–lamina junction and osteophytes arising from the superior endplate of the lower vertebrae are observed. (<b>A</b>) An osteophyte (red arrow) is observed in the L5–S1 foramen originating from the pedicle–lamina junction and extending to the anterior of the superior articular process. Hyperostosis (blue arrow) is observed in the pars region. (<b>B</b>) An osteophyte (red arrow) arising from the pedicle–lamina junction is observed in the L5–S1 foramen. (<b>C</b>) An osteophyte (red arrow) arising from the lower vertebral endplate of the intervertebral foramen is observed. Moreover, an osteophyte (blue arrow) arising from the lower endplate of the superior vertebrae is observed. (<b>D</b>) An osteophyte (red arrow) arising from the superior endplate of the lower vertebrae in the L5–S1 foramen exhibits oblique extension and occupies the foramen.</p> Full article ">Figure 4
<p>Intervertebral foramina and osteophytes are illustrated. (<b>A</b>) Osteophytes (red area) arising from three different points. (<b>B</b>) An osteophyte (red area) arising from the inferior endplate of the superior vertebrae. (<b>C</b>) An osteophyte (red area) arising from the pedicle–lamina junction. (<b>D</b>) A big osteophyte (red area) arising from the lower vertebral endplate of the intervertebral foramen is observed. Simultaneously, osteophytization is observed at two distinct points.</p> Full article ">

attachment

Supplementary material:
Supplementary File 1 (ZIP, 3519 KiB)

17 pages, 9206 KiB  

Open AccessArticle

Contrast Agent Dynamics Determine Radiomics Profiles in Oncologic Imaging

by Martin L. Watzenboeck, Lucian Beer, Daria Kifjak, Sebastian Röhrich, Benedikt H. Heidinger, Florian Prayer, Ruxandra-Iulia Milos, Paul Apfaltrer, Georg Langs, Pascal A. T. Baltzer and Helmut Prosch

Cancers 2024, 16(8), 1519; https://doi.org/10.3390/cancers16081519 - 16 Apr 2024

Viewed by 1025

Abstract

Background: The reproducibility of radiomics features extracted from CT and MRI examinations depends on several physiological and technical factors. The aim was to evaluate the impact of contrast agent timing on the stability of radiomics features using dynamic contrast-enhanced perfusion CT (dceCT) or [...] Read more.

Background: The reproducibility of radiomics features extracted from CT and MRI examinations depends on several physiological and technical factors. The aim was to evaluate the impact of contrast agent timing on the stability of radiomics features using dynamic contrast-enhanced perfusion CT (dceCT) or MRI (dceMRI) in prostate and lung cancers. Methods: Radiomics features were extracted from dceCT or dceMRI images in patients with biopsy-proven peripheral prostate cancer (pzPC) or biopsy-proven non-small cell lung cancer (NSCLC), respectively. Features that showed significant differences between contrast phases were identified using linear mixed models. An L2-penalized logistic regression classifier was used to predict class labels for pzPC and unaffected prostate regions-of-interest (ROIs). Results: Nine pzPC and 28 NSCLC patients, who were imaged with dceCT and/or dceMRI, were included in this study. After normalizing for individual enhancement patterns by defining seven individual phases based on a reference vessel, 19, 467 and 128 out of 1204 CT features showed significant temporal dynamics in healthy prostate parenchyma, prostate tumors and lung tumors, respectively. CT radiomics-based classification accuracy of healthy and tumor ROIs was highly dependent on contrast agent phase. For dceMRI, 899 and 1027 out of 1118 features were significantly dependent on time after contrast agent injection for prostate and lung tumors. Conclusions: CT and MRI radiomics features in both prostate and lung tumors are significantly affected by interindividual differences in contrast agent dynamics. Full article

(This article belongs to the Topic AI in Medical Imaging and Image Processing)

►▼ Show Figures

Figure 1

Figure 1
<p>Synchronization of CT image series across prostate cancer patients. (<b>A</b>) Median voxel intensity as a function of time for each of the seven pzPC patients who were imaged with dceCT after intravenous injection of contrast medium. Lines are colored according to patients. (<b>B</b>) Definition of seven contrast agent phases on a single patient, to enable synchronization of image series across patients. (<b>C</b>) Representative slices for the seven defined contrast agent phases of a single prostate cancer patient.</p> Full article ">Figure 2
<p>Significantly affected CT radiomics features in healthy prostate tissue and pzPC. (<b>A</b>) Heatmap of arterial lumen median voxel intensity. (<b>B</b>,<b>C</b>) Heatmaps of radiomics features with significant temporal dynamics after intravenous contrast agent injection for healthy prostate parenchyma (<b>B</b>) and pzPC (<b>C</b>) sections. For each feature, mean values for all seven patients were z-score transformed over the seven timepoints. Features are grouped according to hierarchical clustering, which is represented by dendrograms. (<b>D</b>,<b>E</b>) Number of radiomics features with and without significant temporal dynamics, grouped by feature class and colored according to clusters as defined in (<b>B</b>,<b>C</b>). glcm: gray level co-occurrence matrix, gldm: gray level dependence matrix, glrlm: gray level run length matrix, glszm: gray level size zone matrix.</p> Full article ">Figure 3
<p>Machine-learning-based classification of healthy and pzPC ROIs depends on contrast agent phase. (<b>A</b>) ROC curves for logistic regression classifiers with an L2 penalty differentiating between healthy and tumor ROIs. Numbering above the curves denotes the timepoint after contrast agent injection. (<b>B</b>) AUC and 95% confidence intervals for each timepoint, which is represented on the <span class="html-italic">x</span>-axis. ROI, region-of-interest; pzPC, peripheral zone prostate cancer; TPR, true-positive rate; FPR, false-positive rate.</p> Full article ">Figure 4
<p>CT radiomics feature dynamics in lung cancer. (<b>A</b>) Representative slices for the seven defined contrast agent phases of a single lung cancer patient. (<b>B</b>) Heatmap of aortal lumen median voxel intensity. (<b>C</b>) Heatmap of radiomics features with significant temporal dynamics after intravenous contrast agent injection in solid lung tumors. For each feature, mean values for all seven patients are z-score transformed over the seven timepoints. Features are grouped by hierarchical clustering. (<b>D</b>) Distribution of radiomics features across different feature classes. Features are colored according to clusters defined in (<b>C</b>). glcm: gray level co-occurrence matrix, gldm: gray level dependence matrix, glrlm: gray level run length matrix, glszm: gray level size zone matrix.</p> Full article ">Figure 5
<p>MRI radiomics feature dynamics in prostate cancer. (<b>A</b>) Representative slice illustrating segmentation of femoral artery and pzPC lesions in a single pzPC patient. (<b>B</b>) Heatmap of the femoral artery lumen median voxel intensity. (<b>C</b>) Heatmap of radiomics features with significant temporal dynamics after intravenous contrast agent injection in pzPC lesions. Features are grouped by hierarchical clustering. (<b>D</b>) Distribution of radiomics features across different feature classes. Features are colored according to clusters defined in (<b>C</b>). glcm: gray level co-occurrence matrix, gldm: gray level dependence matrix, glrlm: gray level run length matrix, glszm: gray level size zone matrix.</p> Full article ">Figure 6
<p>MRI radiomics feature dynamics in lung cancer. (<b>A</b>) Representative slice illustrating segmentation of thoracic aorta and NSCLC lesions in a single NSCLC patient. (<b>B</b>) Heatmap of aortal lumen median voxel intensity. (<b>C</b>) Heatmap of radiomics features with significant temporal dynamics after intravenous contrast agent injection in solid lung tumors. Features are grouped by hierarchical clustering. (<b>D</b>) Distribution of radiomics features across different feature classes. Features are colored according to clusters defined in (<b>C</b>). glcm: gray level co-occurrence matrix, gldm: gray level dependence matrix, glrlm: gray level run length matrix, glszm: gray level size zone matrix.</p> Full article ">

16 pages, 2832 KiB  

Open AccessArticle

Impact of Deep Learning Denoising Algorithm on Diffusion Tensor Imaging of the Growth Plate on Different Spatial Resolutions

by Laura Santos, Hao-Yun Hsu, Ronald R. Nelson, Jr., Brendan Sullivan, Jaemin Shin, Maggie Fung, Marc R. Lebel, Sachin Jambawalikar and Diego Jaramillo

Tomography 2024, 10(4), 504-519; https://doi.org/10.3390/tomography10040039 - 2 Apr 2024

Viewed by 1109

Abstract

To assess the impact of a deep learning (DL) denoising reconstruction algorithm applied to identical patient scans acquired with two different voxel dimensions, representing distinct spatial resolutions, this IRB-approved prospective study was conducted at a tertiary pediatric center in compliance with the Health [...] Read more.

To assess the impact of a deep learning (DL) denoising reconstruction algorithm applied to identical patient scans acquired with two different voxel dimensions, representing distinct spatial resolutions, this IRB-approved prospective study was conducted at a tertiary pediatric center in compliance with the Health Insurance Portability and Accountability Act. A General Electric Signa Premier unit (GE Medical Systems, Milwaukee, WI) was employed to acquire two DTI (diffusion tensor imaging) sequences of the left knee on each child at 3T: an in-plane 2.0 × 2.0 ${\mathrm{m}\mathrm{m}}^2$ with section thickness of 3.0 mm and a 2 mm³ isovolumetric voxel; neither had an intersection gap. For image acquisition, a multi-band DTI with a fat-suppressed single-shot spin-echo echo-planar sequence (20 non-collinear directions; b-values of 0 and 600 s/mm²) was utilized. The MR vendor-provided a commercially available DL model which was applied with 75% noise reduction settings to the same subject DTI sequences at different spatial resolutions. We compared DTI tract metrics from both DL-reconstructed scans and non-denoised scans for the femur and tibia at each spatial resolution. Differences were evaluated using Wilcoxon-signed ranked test and Bland–Altman plots. When comparing DL versus non-denoised diffusion metrics in femur and tibia using the 2 mm × 2 mm × 3 mm voxel dimension, there were no significant differences between tract count (p = 0.1, p = 0.14) tract volume (p = 0.1, p = 0.29) or tibial tract length (p = 0.16); femur tract length exhibited a significant difference (p < 0.01). All diffusion metrics (tract count, volume, length, and fractional anisotropy (FA)) derived from the DL-reconstructed scans, were significantly different from the non-denoised scan DTI metrics in both the femur and tibial physes using the 2 mm³ voxel size (p < 0.001). DL reconstruction resulted in a significant decrease in femorotibial FA for both voxel dimensions (p < 0.01). Leveraging denoising algorithms could address the drawbacks of lower signal-to-noise ratios (SNRs) associated with smaller voxel volumes and capitalize on their better spatial resolutions, allowing for more accurate quantification of diffusion metrics. Full article

(This article belongs to the Topic AI in Medical Imaging and Image Processing)

►▼ Show Figures

Figure 1

Figure 1
<p>Same subject DTI acquisition using voxel sizes of (<b>A</b>) 2 mm³ and (<b>D</b>) 2 mm × 2 mm × 3 mm. The image quality is visibly better when DL reconstruction is applied in the 2 mm³ voxel size (<b>A</b>), with sharper bone contours and an increase in fiber tracts on both the femur and tibia. Conversely, in the 2 mm × 2 mm × 3 mm voxel size, both the non-denoised and DL denoise images (<b>D</b>) appear equally pixelated, and there are minimal changes in tractography. The tensor ellipsoid representation of intravoxel eigen vectors is substantially larger and more numerous in the 2 mm³ voxel size (<b>B</b>,<b>C</b>) compared to the 2 mm × 2 mm × 3 mm voxel size (<b>E</b>,<b>F</b>).</p> Full article ">Figure 2
<p>Bland–Altman plots regarding the difference in DTI metrics for same physeal ROIs between DL reconstructed-and non-denoised (Non-DL) image DTIs. The horizontal axis represents the mean of the two methods and the vertical axis, the difference between them. The solid line (red) shows the mean difference (close to zero) and the dashed lines show the 95% limits of agreement.</p> Full article ">Figure 3
<p>Bland–Altman plots regarding the difference in DTI metrics for same physeal ROIs between DL-reconstructed and non-denoised (non-DL) image DTIs. The horizontal axis represents the mean of the two methods and the vertical axis, the difference between them. The solid line (red) shows the mean difference (close to zero) and the dashed lines show the 95% limits of agreement.</p> Full article ">Figure 4
<p>Bland–Altman plots regarding the difference in DTI metrics for same physeal ROIs between DL reconstructed and non-denoised (non-DL) DTIs. The zero value is indicated by the red line. The horizontal axis represents the mean of the two methods and the vertical axis, the difference between them. The solid line (red) shows the mean difference (close to zero) and the dashed lines show the 95% limits of agreement. The mean is very close to zero for most cases, indicating little difference between the methods, and the range of LoA is relatively small indicating a good numerical agreement in the methods among the majority of patients for the 2 mm × 2 mm × 3 mm voxel size.</p> Full article ">Figure 5
<p>Bland–Altman plots regarding the difference in DTI metrics for same physeal ROIs between DL reconstructed and non-denoised (non-DL) DTIs. The zero value is indicated by the red line. The horizontal axis represents the mean of the two methods and the vertical axis, the difference between them. The solid line (red) shows the mean difference (close to zero) and the dashed lines show the 95% limits of agreement. The mean is very close to zero for most cases, indicating little difference between the methods, and the range of LoA is relatively small indicating a good numerical agreement in the methods among the majority of patients for the 2 mm × 2 mm × 3 mm voxel size.</p> Full article ">Figure 6
<p>(<b>a</b>). 2 mm × 2 mm × 3 mm (<b>A</b>) non-denoised (<b>left</b>) and 2 mm × 2 mm × 3 mm denoised (<b>right</b>) versus 2 mm³ (<b>B</b>) non-denoised (<b>left</b>) and denoised (<b>right</b>) in a 9-year-old girl. There is an evident increase in fiber tract on the denoised images acquired with a 2 mm³, while denoised and non-denoised images are very similar using 2 mm × 2 mm × 3 mm voxel size. (<b>b</b>). 2 mm × 2 mm × 3 mm (<b>C</b>) non-denoised (<b>left</b>) and denoised (<b>right</b>), versus 2 mm³ (<b>D</b>) non-denoised (<b>left</b>) and denoised (<b>right</b>) in a10 year-old boy. There is an evident increase in fiber tract on the denoised images acquired with a 2 mm³, while denoised and non-denoised images are very similar using 2 mm × 2 mm × 3 mm voxel size.</p> Full article ">

attachment

Supplementary material:
Supplementary File 1 (ZIP, 5777 KiB)

17 pages, 2263 KiB  

Open AccessArticle

Characterization of CD34⁺ Cells from Patients with Acute Myeloid Leukemia (AML) and Myelodysplastic Syndromes (MDS) Using a t-Distributed Stochastic Neighbor Embedding (t-SNE) Protocol

by Cathrin Nollmann, Wiebke Moskorz, Christian Wimmenauer, Paul S. Jäger, Ron P. Cadeddu, Jörg Timm, Thomas Heinzel and Rainer Haas

Cancers 2024, 16(7), 1320; https://doi.org/10.3390/cancers16071320 - 28 Mar 2024

Viewed by 1272

Abstract

Using multi-color flow cytometry analysis, we studied the immunophenotypical differences between leukemic cells from patients with AML/MDS and hematopoietic stem and progenitor cells (HSPCs) from patients in complete remission (CR) following their successful treatment. The panel of markers included CD34, CD38, CD45RA, CD123 [...] Read more.

Using multi-color flow cytometry analysis, we studied the immunophenotypical differences between leukemic cells from patients with AML/MDS and hematopoietic stem and progenitor cells (HSPCs) from patients in complete remission (CR) following their successful treatment. The panel of markers included CD34, CD38, CD45RA, CD123 as representatives for a hierarchical hematopoietic stem and progenitor cell (HSPC) classification as well as programmed death ligand 1 (PD-L1). Rather than restricting the evaluation on a 2- or 3-dimensional analysis, we applied a t-distributed stochastic neighbor embedding (t-SNE) approach to obtain deeper insight and segregation between leukemic cells and normal HPSCs. For that purpose, we created a t-SNE map, which resulted in the visualization of 27 cell clusters based on their similarity concerning the composition and intensity of antigen expression. Two of these clusters were “leukemia-related” containing a great proportion of CD34⁺/CD38⁻ hematopoietic stem cells (HSCs) or CD34⁺ cells with a strong co-expression of CD45RA/CD123, respectively. CD34⁺ cells within the latter cluster were also highly positive for PD-L1 reflecting their immunosuppressive capacity. Beyond this proof of principle study, the inclusion of additional markers will be helpful to refine the differentiation between normal HSPCs and leukemic cells, particularly in the context of minimal disease detection and antigen-targeted therapeutic interventions. Furthermore, we suggest a protocol for the assignment of new cell ensembles in quantitative terms, via a numerical value, the Pearson coefficient, based on a similarity comparison of the t-SNE pattern with a reference. Full article

(This article belongs to the Topic AI in Medical Imaging and Image Processing)

►▼ Show Figures

Figure 1

Figure 1
<p>In (<b>A</b>) and (<b>B</b>), steps (5) and (6) of the gating strategy are shown. The orange arrow means that in B only the cells from the gate of interest (GOI, black frame) were analyzed. (<b>A</b>) The exclusion of the population of granulocytes with (5) a GOI; (<b>B</b>) The final step (6), selection of the CD34⁺ cells; In (<b>C</b>), the numbers of CD34⁺ cells after the gating are shown for each patient as well as the numbers thereof included in the t-SNE analysis.</p> Full article ">Figure 2
<p>(<b>A</b>) t-SNE representation of the combined data set (cells of patients with active disease (AD) and in complete remission (CR)), and the contributions from patients in CR (<b>B</b>) as well as from the patients with AD (<b>C</b>). The color scale for (<b>A</b>–<b>C</b>) corresponds to the local density of cells in arbitrary units. In each t-SNE plot, the color scale starts at zero and is normalized to the maximum density in the respective plot. In (<b>D</b>–<b>H</b>), the prevalence of the various markers entering the t-SNE algorithm are reproduced, for CD34, CD38, CD45RA, CD123 and PD-L1. The color scale represents the expression level. The classification in terms of <span class="html-italic">positive</span> (+) and <span class="html-italic">negative</span> (−) expression is indicated by the black horizontal lines in the color bar.</p> Full article ">Figure 3
<p>(<b>A</b>) Density-based gate definition on the t-SNE plot of the CR data set, performed by visual inspection; (<b>B</b>) Application of the gates on the AD data set; In (<b>C</b>), the ratio of the percentage distribution of the cells for the CR to the AD samples is given for each gate; The gates are inked in red if this ratio is smaller than 1, i.e., most of the cells in this gate come from AD samples, and blue for ratios larger than 1. The cell type identification for the gates is represented in (<b>D</b>,<b>E</b>) for the CR samples and for the AD samples, respectively; (<b>F</b>) Evolution scheme for the relevant cell types; The ratios of the percentage distribution of cells for the different types in CR samples vs. AD samples are listed in (<b>G</b>).</p> Full article ">Figure 4
<p>Box plots of the antigen expression levels (red: CD34, green: CD38, blue: CD45RA, orange: CD123, purple: PD-L1) within 5 selected gates from the 27 shown in <a href="#cancers-16-01320-f003" class="html-fig">Figure 3</a>. The fluorescence data were scaled biexponentially in a preliminary step. The light blue and red background indicate the expression levels classified as <span class="html-italic">negative</span> (blue) and <span class="html-italic">positive</span> (red), respectively. The PD-L1 antigen was not used for the cell classification.</p> Full article ">Figure 5
<p>Graphic representation of the calculation of the r value as an example for two patients, 6 (CR group, top center) and 16 (AD group, bottom center). For patient 6, the density distribution is compared with those containing all patients with AD (top right) and for all patients of the CR group except patient 6 (top left). Likewise, we compare for patient 16 the density with the t-SNE plots containing all patients with AD except patient 16 (bottom right) and all patients of the CR group (bottom left). The orange arrows indicate which two plots were compared with each other. The r values are given in between the compared plots. The density is normalized to the respective maximum value of the plot. The gates are shown as an overlay for all plots.</p> Full article ">Figure 6
<p>(<b>A</b>) Density plot of all patients with AD without patient 4 and patient 11; (<b>B</b>) density plot of patient 4, and (<b>C</b>) of patient 11. The gates are shown as an overlay in all plots.</p> Full article ">

12 pages, 2975 KiB  

Open AccessArticle

Validation of Left Atrial Volume Correction for Single Plane Method on Four-Chamber Cine Cardiac MRI

by Hosamadin Assadi, Nicholas Sawh, Ciara Bailey, Gareth Matthews, Rui Li, Ciaran Grafton-Clarke, Zia Mehmood, Bahman Kasmai, Peter P. Swoboda, Andrew J. Swift, Rob J. van der Geest and Pankaj Garg

Tomography 2024, 10(4), 459-470; https://doi.org/10.3390/tomography10040035 - 25 Mar 2024

Viewed by 1312

Abstract

Background: Left atrial (LA) assessment is an important marker of adverse cardiovascular outcomes. Cardiovascular magnetic resonance (CMR) accurately quantifies LA volume and function based on biplane long-axis imaging. We aimed to validate single-plane-derived LA indices against the biplane method to simplify the post-processing [...] Read more.

Background: Left atrial (LA) assessment is an important marker of adverse cardiovascular outcomes. Cardiovascular magnetic resonance (CMR) accurately quantifies LA volume and function based on biplane long-axis imaging. We aimed to validate single-plane-derived LA indices against the biplane method to simplify the post-processing of cine CMR. Methods: In this study, 100 patients from Leeds Teaching Hospitals were used as the derivation cohort. Bias correction for the single plane method was applied and subsequently validated in 79 subjects. Results: There were significant differences between the biplane and single plane mean LA maximum and minimum volumes and LA ejection fraction (EF) (all p < 0.01). After correcting for biases in the validation cohort, significant correlations in all LA indices were observed (0.89 to 0.98). The area under the curve (AUC) for the single plane to predict biplane cutoffs of LA maximum volume ≥ 112 mL was 0.97, LA minimum volume ≥ 44 mL was 0.99, LA stroke volume (SV) ≤ 21 mL was 1, and LA EF ≤ 46% was 1, (all p < 0.001). Conclusions: LA volumetric and functional assessment by the single plane method has a systematic bias compared to the biplane method. After bias correction, single plane LA volume and function are comparable to the biplane method. Full article

(This article belongs to the Topic AI in Medical Imaging and Image Processing)

►▼ Show Figures

Figure 1

Figure 1
<p>Cumulative frequency distribution of biplane left atrial (LA) parameters in the whole study cohort (<span class="html-italic">n</span> = 179). Each graph’s two distinct background colors divide them at the halfway point of the <span class="html-italic">y</span>-axis, indicating different zones or reference ranges for easier visual interpretation. The figure effectively visualizes the distributions of the biplane LA parameters measured in the study.</p> Full article ">Figure 2
<p>Scatter plots with heat maps demonstrating correlations between single plane four-chamber cine (4CH) left atrial (LA) parameters and biplane LA parameters. The color-coding of data points represents data density, with red indicating higher density and blue signifying lower density. Each plot includes a linear regression line with confidence intervals, demonstrating the strength and direction of the correlations. The figure effectively visualizes the strong positive correlations between the single plane and biplane LA parameters measured in the study.</p> Full article ">Figure 3
<p>Bland–Altman plots for all left atrial (LA) variables demonstrate significant bias between biplane and single plane methods for LA maximum and minimum volumes. Both are systematically underestimated by the single plane method. Stroke volume (SV) is similar between both groups. Due to the relative difference in maximum and minimum volumes, the single plane method systematically overestimates the LA ejection fraction (EF). Other than for stroke volume, all <span class="html-italic">p</span>-values are significant for paired comparison.</p> Full article ">Figure 4
<p>Post-bias correction agreement investigation in the validation cohort (<span class="html-italic">n</span> = 79): Bland–Altman plot agreement and correlation between biplane and single plane methods of left atrial (LA) maximum and minimum volumes assessment methods. Both Bland–Altman and regression analyses demonstrate the agreement of biplane versus single plane measurements for these two variables.</p> Full article ">Figure 5
<p>Post-bias correction agreement investigation in the validation cohort (<span class="html-italic">n</span> = 79): Bland–Altman plot agreement and correlation between biplane and single plane methods of left atrial (LA) ejection fraction (EF) and stroke volume (SV) assessment. Both Bland–Altman and regression analyses demonstrate the agreement of biplane versus single plane measurements for these two variables.</p> Full article ">Figure 6
<p>Receiver operating characteristic (ROC) Curves for left atrial (LA) parameters in predicting biplane cutoffs of abnormal values using the upper limit of normal values. [(<b>A</b>) maximum volume ≥ 112 mL, (<b>B</b>) minimum volume ≥ 44 mL, (<b>C</b>) stroke volume ≤ 21 mL, and (<b>D</b>) ejection fraction ≤ 46%]. Each panel displays the sensitivity (<span class="html-italic">y</span>-axis) against 100-specificity (<span class="html-italic">x</span>-axis). All <span class="html-italic">p</span>-values are less than 0.001, indicating significant predictive power for all metrics.</p> Full article ">

14 pages, 1931 KiB  

Open AccessArticle

Automatic Active Lesion Tracking in Multiple Sclerosis Using Unsupervised Machine Learning

by Jason Uwaeze, Ponnada A. Narayana, Arash Kamali, Vladimir Braverman, Michael A. Jacobs and Alireza Akhbardeh

Diagnostics 2024, 14(6), 632; https://doi.org/10.3390/diagnostics14060632 - 16 Mar 2024

Viewed by 1381

Abstract

Background: Identifying active lesions in magnetic resonance imaging (MRI) is crucial for the diagnosis and treatment planning of multiple sclerosis (MS). Active lesions on MRI are identified following the administration of Gadolinium-based contrast agents (GBCAs). However, recent studies have reported that repeated administration [...] Read more.

Background: Identifying active lesions in magnetic resonance imaging (MRI) is crucial for the diagnosis and treatment planning of multiple sclerosis (MS). Active lesions on MRI are identified following the administration of Gadolinium-based contrast agents (GBCAs). However, recent studies have reported that repeated administration of GBCA results in the accumulation of Gd in tissues. In addition, GBCA administration increases health care costs. Thus, reducing or eliminating GBCA administration for active lesion detection is important for improved patient safety and reduced healthcare costs. Current state-of-the-art methods for identifying active lesions in brain MRI without GBCA administration utilize data-intensive deep learning methods. Objective: To implement nonlinear dimensionality reduction (NLDR) methods, locally linear embedding (LLE) and isometric feature mapping (Isomap), which are less data-intensive, for automatically identifying active lesions on brain MRI in MS patients, without the administration of contrast agents. Materials and Methods: Fluid-attenuated inversion recovery (FLAIR), T2-weighted, proton density-weighted, and pre- and post-contrast T1-weighted images were included in the multiparametric MRI dataset used in this study. Subtracted pre- and post-contrast T1-weighted images were labeled by experts as active lesions (ground truth). Unsupervised methods, LLE and Isomap, were used to reconstruct multiparametric brain MR images into a single embedded image. Active lesions were identified on the embedded images and compared with ground truth lesions. The performance of NLDR methods was evaluated by calculating the Dice similarity (DS) index between the observed and identified active lesions in embedded images. Results: LLE and Isomap, were applied to 40 MS patients, achieving median DS scores of 0.74 ± 0.1 and 0.78 ± 0.09, respectively, outperforming current state-of-the-art methods. Conclusions: NLDR methods, Isomap and LLE, are viable options for the identification of active MS lesions on non-contrast images, and potentially could be used as a clinical decision tool. Full article

(This article belongs to the Topic AI in Medical Imaging and Image Processing)

►▼ Show Figures

Figure 1

Figure 1
<p>The Isomap algorithm can be summarized in four steps: (1) select a nearest neighbor algorithm, (2) calculate geodesic distances for all data points (3) apply multidimensional scaling to geodesic distance matrix, (4) output single embedded image.</p> Full article ">Figure 2
<p>Embedding of the 3-dimensional Swiss roll to 2 dimensions using MDS, LLE, and Isomap. Neighborhood size, k, was set to 12 for all methods. Both NLDR methods, LLE and Isomap, unfolded the Swiss roll in lower dimensions while the linear DR method, MDS, failed to unfold and preserve the manifold. Results that best retained the shape of the Swiss roll were derived by Isomap.</p> Full article ">Figure 3
<p>Multiparametric input brain MRI integration framework for embedding reconstruction. (<b>Top row</b>) Four brain MRIs are inputted and then integrated into a single embedded image using NLDR methods. The NLDR maps are thresholded for lesions. (<b>Bottom row</b>) T1-pre and post are subtracted. Next, active lesions are masked to obtain the ground truths binary masks. At the end, the two binary images are compared using dice similarity (DS) score, for the evaluation of proposed methods.</p> Full article ">Figure 4
<p>Dice Similarity is used to describe the overlap between embedded and ground truth active lesions.</p> Full article ">Figure 5
<p>Examples of LLE algorithm sensitivity to different image matrix sizes and resolutions. We compared the active lesion (yellow arrow) in the embedded image to ground truth lesions using DS scores. (<b>A</b>–<b>C</b>) As shown in 64² images, (bottom row) DS was significantly lower. However, no significant differences were observed between 256² and 128².</p> Full article ">Figure 6
<p>Evaluation of the sensitivity of NLDR methods to control parameters (K; neighborhood size): (<b>A</b>) Isomap (<b>B</b>) LLE.</p> Full article ">Figure 7
<p>(<b>A</b>) Multiparametric brain MRI data as input to nonlinear dimensionality reduction (NLDR) methods. (<b>B-upper</b>) Output embedded images from locally linear embedding (LLE) and isometric feature mapping (Isomap). (<b>B-lower</b>) The resulting binary masks are shown. (<b>C-upper</b>) The ground truth (contrast subtraction image). (<b>C-lower</b>) Corresponding binary image for ground truth, used to evaluate methods using dice similarity scores.</p> Full article ">Figure 8
<p>The performance of nonlinear dimensionality reduction methods, locally linear embedding (LLE) and isometric feature mapping (Isomap), are shown in the box plot. Isomap marginally outperformed LLE with a median and standard deviation for dice similarity (DS) score, respectively, as 0.78 ± 0.09 and 0.74 ± 0.09.</p> Full article ">

attachment

Supplementary material:
Supplementary File 1 (ZIP, 102 KiB)

13 pages, 3128 KiB  

Open AccessArticle

Impact of AI-Based Post-Processing on Image Quality of Non-Contrast Computed Tomography of the Chest and Abdomen

by Marcel A. Drews, Aydin Demircioğlu, Julia Neuhoff, Johannes Haubold, Sebastian Zensen, Marcel K. Opitz, Michael Forsting, Kai Nassenstein and Denise Bos

Diagnostics 2024, 14(6), 612; https://doi.org/10.3390/diagnostics14060612 - 13 Mar 2024

Cited by 1 | Viewed by 1514

Abstract

Non-contrast computed tomography (CT) is commonly used for the evaluation of various pathologies including pulmonary infections or urolithiasis but, especially in low-dose protocols, image quality is reduced. To improve this, deep learning-based post-processing approaches are being developed. Therefore, we aimed to compare the [...] Read more.

Non-contrast computed tomography (CT) is commonly used for the evaluation of various pathologies including pulmonary infections or urolithiasis but, especially in low-dose protocols, image quality is reduced. To improve this, deep learning-based post-processing approaches are being developed. Therefore, we aimed to compare the objective and subjective image quality of different reconstruction techniques and a deep learning-based software on non-contrast chest and low-dose abdominal CTs. In this retrospective study, non-contrast chest CTs of patients suspected of COVID-19 pneumonia and low-dose abdominal CTs suspected of urolithiasis were analysed. All images were reconstructed using filtered back-projection (FBP) and were post-processed using an artificial intelligence (AI)-based commercial software (PixelShine (PS)). Additional iterative reconstruction (IR) was performed for abdominal CTs. Objective and subjective image quality were evaluated. AI-based post-processing led to an overall significant noise reduction independent of the protocol (chest or abdomen) while maintaining image information (max. difference in SNR 2.59 ± 2.9 and CNR 15.92 ± 8.9, p < 0.001). Post-processing of FBP-reconstructed abdominal images was even superior to IR alone (max. difference in SNR 0.76 ± 0.5, p ≤ 0.001). Subjective assessments verified these results, partly suggesting benefits, especially in soft-tissue imaging (p < 0.001). All in all, the deep learning-based denoising—which was non-inferior to IR—offers an opportunity for image quality improvement especially in institutions using older scanners without IR availability. Further studies are necessary to evaluate potential effects on dose reduction benefits. Full article

(This article belongs to the Topic AI in Medical Imaging and Image Processing)

►▼ Show Figures

Figure 1

Figure 1
<p>Patient flowchart and study design for assessment of thoracal (<b>A</b>) and abdominal LD-CT (<b>B</b>).</p> Full article ">Figure 2
<p>Axial CT slices reconstructed with FBP showing the positions of seven evaluated ROIs (green circle) for thoracal CT (<b>A</b>) and six ROIs for abdominal CT (<b>B</b>). AAC = ascending aorta, ADE = descending thoracal aorta, AOD = descending abdominal aorta, AIR = air, AUM = autochthonous muscle, FAT = subcutaneous fat, LLL = left liver lobule, LUN = lung tissue, TRP = pulmonary trunk, SPL = spleen.</p> Full article ">Figure 3
<p>Axial CT of a 65-year-old patient with COVID-19 pneumonia comparing FBP reconstruction (<b>A</b>,<b>C</b>) and post-processing with PS (<b>B</b>,<b>D</b>) using soft-tissue (<b>A</b>,<b>B</b>) and lung kernel (<b>C</b>,<b>D</b>). Window width (W) and centre (C) are given.</p> Full article ">Figure 4
<p>Axial CT of a 63-year-old patient suspected of urolithiasis comparing FBP reconstruction (<b>A</b>), FBP post-processed with PS (<b>B</b>), IR (<b>C</b>), and IR post-processed with PS (<b>D</b>). Window width (W) and centre (C) are given.</p> Full article ">

10 pages, 1275 KiB  

Open AccessArticle

Performance of ECG-Derived Digital Biomarker for Screening Coronary Occlusion in Resuscitated Out-of-Hospital Cardiac Arrest Patients: A Comparative Study between Artificial Intelligence and a Group of Experts

by Min Ji Park, Yoo Jin Choi, Moonki Shim, Youngjin Cho, Jiesuck Park, Jina Choi, Joonghee Kim, Eunkyoung Lee and Seo-Yoon Kim

J. Clin. Med. 2024, 13(5), 1354; https://doi.org/10.3390/jcm13051354 - 27 Feb 2024

Cited by 3 | Viewed by 1173

Abstract

Acute coronary syndrome is a significant part of cardiac etiology contributing to out-of-hospital cardiac arrest (OHCA), and immediate coronary angiography has been proposed to improve survival. This study evaluated the effectiveness of an AI algorithm in diagnosing near-total or total occlusion of coronary [...] Read more.

Acute coronary syndrome is a significant part of cardiac etiology contributing to out-of-hospital cardiac arrest (OHCA), and immediate coronary angiography has been proposed to improve survival. This study evaluated the effectiveness of an AI algorithm in diagnosing near-total or total occlusion of coronary arteries in OHCA patients who regained spontaneous circulation. Conducted from 1 July 2019 to 30 June 2022 at a tertiary university hospital emergency department, it involved 82 OHCA patients, with 58 qualifying after exclusions. The AI used was the Quantitative ECG (QCG™) system, which provides a STEMI diagnostic score ranging from 0 to 100. The QCG score’s diagnostic performance was compared to assessments by two emergency physicians and three cardiologists. Among the patients, coronary occlusion was identified in 24. The QCG score showed a significant difference between occlusion and non-occlusion groups, with the former scoring higher. The QCG biomarker had an area under the curve (AUC) of 0.770, outperforming the expert group’s AUC of 0.676. It demonstrated 70.8% sensitivity and 79.4% specificity. These findings suggest that the AI-based ECG biomarker could predict coronary occlusion in resuscitated OHCA patients, and it was non-inferior to the consensus of the expert group. Full article

(This article belongs to the Topic AI in Medical Imaging and Image Processing)

►▼ Show Figures

Figure 1

Figure 1
<p>Boxplot of (<b>A</b>) QCG scores of the coronary occlusion group and the non-occlusion group. (<b>B</b>) Experts’ consensus of the two groups. (<b>C</b>) Troponin I level in the two groups.</p> Full article ">Figure 2
<p>ROC curves for the prediction of coronary occlusion. Black solid line: ROC curve of QCG, Green solid line: ROC curve of experts’ consensus, Y (black filled circle): Youden Index Point of QCG, E (red hollow circles): Emergency Physicians, C (blue hollow triangles): Cardiologists, A (green filled circle): Youden Index Point of the Consensus score.</p> Full article ">

Show export options expand_more Show export options expand_less

Select all

Export citation of selected articles as:

Plain Text BibTeX BibTeX (without abstracts) Endnote Endnote (without abstracts) Tab-delimited RIS

 

Error

Oops... you haven't selected anything for export.

Ok

clear

Displaying articles 1-40