Lukas Fischer

In the last decade, industry’s demand for deep learning (DL) has increased due to its high performance in complex scenarios. Due to the DL method’s complexity, experts and non-experts rely on blackbox software packages such as Tensorflow and Pytorch. The frameworks are constantly improving, and new versions are released frequently. As a natural process in software development, the released versions contain improvements/changes in the methods and their implementation. Moreover, versions may be bug-polluted, leading to the model performance decreasing or stopping the model from working. The aforementioned changes in implementation can lead to variance in obtained results. This work investigates the effect of implementation changes in different major releases of these frameworks on the model performance. We perform our study using a variety of standard datasets. Our study shows that users should consider that changing the framework version can affect the model performance. Moreover, th...

In order to develop machine learning and deep learning models that take into account the guidelines and principles of trustworthy AI, a novel information theoretic approach is introduced in this article. A unified approach to privacy-preserving interpretable and transferable learning is considered for studying and optimizing the trade-offs between the privacy, interpretability, and transferability aspects of trustworthy AI. A variational membership-mapping Bayesian model is used for the analytical approximation of the defined information theoretic measures for privacy leakage, interpretability, and transferability. The approach consists of approximating the information theoretic measures by maximizing a lower-bound using variational optimization. The approach is demonstrated through numerous experiments on benchmark datasets and a real-world biomedical application concerned with the detection of mental stress in individuals using heart rate variability analysis.

Zsfassung in dt. SpracheIn den letzten Jahren lieferten Segmentierungsansätze basierend auf sequenzielle Monte Carlo Methoden vielversprechende Ergebnisse bei der Lokalisierung und Beschreibung anatomischer Strukturen in medizinisch relevanten Bildern. Auch bekannt unter der Bezeichnung Shape Particle Filter wurden diese Methoden für die Segmentierung von Wirbelkörpern, Lungenflügeln und Herzen eingesetzt. Ihr großer Vorteil liegt darin, dass sie auch bei Bildern mit starkem Rauschen wie zum Beispiel MR Aufnahmen, sowie bei sich überlagernden Strukturen, bei denen eine eindeutige Unterscheidung der Objekte schwierig ist, noch sehr gute Segmentierungsergebnisse liefern. Shape Particle Filter benötigen eine Maske, welche auf dem Mean Shape eines Shape Models basiert und in existierenden Implementierungen immer manuell definiert wird. Während der Suche nach einem Objekt wird die Wahrscheinlichkeit für jeden Pixel zu einer gewissen Region der Maske zu gehören durch das Klassifizieren vo...

The recent success of Generative Adversarial Networks (GAN) is a result of their ability to generate high quality images from a latent vector space. An important application is the generation of images from a text description, where the text description is encoded and further used in the conditioning of the generated image. Thus the generative network has to additionally learn a mapping from the text latent vector space to a highly complex and multi-modal image data distribution, which makes the training of such models challenging. To handle the complexities of fashion image and meta data, we propose Ontology Generative Adversarial Networks (O-GANs) for fashion image synthesis that is conditioned on an hierarchical fashion ontology in order to improve the image generation fidelity. We show that the incorporation of the ontology leads to better image quality as measured by Fréchet Inception Distance and Inception Score. Additionally, we show that the O-GAN achieves better conditionin...

Poster: ESSR 2013 / P-0014 / Evaluation of cortical bone density and microarchitecture of lung transplant recipients by HR-pQCT" by: "L. Fischer1, A. Valentinitsch2, M. D. DiFranco1, C. Schueller-Weidekamm1, B. Zweytick1, F. Kainberger1, G. Langs1, J. M. Patsch1; 1Vienna/AT, 2Wien/AT;

Poster: "ECR 2012 / B-0145 / Trabecular direction and deformation distribution in lung transplant patients with severe osteoporosis risk" by: "L. Fischer1, J. M. Patsch2, A. Valentinitsch2, C. Schueller-Weidekamm1, B. Zweytick1, F. Kainberger1, G. Langs1; 1Vienna/AT, 2San Francisco, CA/US"

Animated illustration of the cross-sectional cut planes. Green plane: Proximal plane (50% of the distance between the tip of the styloid process and the most dorsal point of the tuberculum listerii); Blue plane: Middle plane (Half way between the distal and proximal sectional plane); Red plane: Distal plane (The most dorsal point of the tuberculum listerii). (PDF 140 kb)

To determine whether stress fractures are associated with bone microstructural deterioration we quantified distal radial and the unfractured distal tibia using high resolution peripheral quantitative computed tomography in 26 cases with lower limb stress fractures (15 males, 11 females; mean age 37.1 ± 3.1 years) and 62 age-matched healthy controls (24 males, 38 females; mean age 35.0 ± 1.6 years). Relative to controls, in men, at the distal radius, cases had smaller cortical cross sectional area (CSA) (p = 0.012), higher porosity of the outer transitional zone (OTZ) (p = 0.006), inner transitional zone (ITZ) (p = 0.043) and the compact-appearing cortex (CC) (p = 0.023) while trabecular vBMD was lower (p = 0.002). At the distal tibia, cases also had a smaller cortical CSA (p = 0.008). Cortical porosity was not higher, but trabecular vBMD was lower (p = 0.001). Relative to controls, in women, cases had higher distal radial porosity of the OTZ (p = 0.028), ITZ (p = 0.030) not CC (p = ...

Separating and labeling each instance of a nucleus (instance-aware segmentation) is the key challenge in segmenting single cell nuclei on fluorescence microscopy images. Deep Neural Networks can learn the implicit transformation of a nuclear image into a probability map indicating the class membership of each pixel (nucleus or background), but the use of post-processing steps to turn the probability map into a labeled object mask is error-prone. This especially accounts for nuclear images of tissue sections and nuclear images across varying tissue preparations. In this work, we aim to evaluate the performance of state-of-the-art deep learning architectures to segment nuclei in fluorescence images of various tissue origins and sample preparation types without post-processing. We compare architectures that operate on pixel to pixel translation and an architecture that operates on object detection and subsequent locally applied segmentation. In addition, we propose a novel strategy to ...

Guidelines and principles of trustworthy AI should be adhered to in practice during the development of AI systems. This work suggests a novel information theoretic trustworthy AI framework based on the hypothesis that information theory enables taking into account the ethical AI principles during the development of machine learning and deep learning models via providing a way to study and optimize the inherent tradeoffs between trustworthy AI principles. Under the proposed framework, a unified approach to “privacy-preserving interpretable and transferable learning” is considered to introduce the information theoretic measures for privacy-leakage, interpretability, and transferability. A technique based on variational optimization, employing conditionally deep autoencoders, is developed for practically calculating the defined information theoretic measures for privacy-leakage, interpretability, and transferability.

BACKGROUND Neoadjuvant chemotherapy in patients with primary osteosarcoma improves survival rates, but it also causes side effects in various organs including bone. Low bone mineral density (BMD) can occur owing partly to chemotherapy or limited mobility. This can cause a higher risk of fractures compared with those who do not receive such treatment. Changes in BMD alone cannot explain the propensity of fractures. Studying microarchitectural changes of bone might help to understand the effect. QUESTIONS/PURPOSES (1) Do patients who were treated for osteosarcoma (more than 20 years previously) have low BMD? (2) Do these patients experience more fractures than controls who do not have osteosarcoma? (3) What differences in bone microarchitecture are present between patients treated for high-grade osteosarcoma and individuals who have never had osteosarcoma? METHODS We contacted 48 patients who were treated for osteosarcoma and who participated in an earlier study. These patients underwent multimodal treatment including chemotherapy more than 20 years ago. Of the original patient group, 60% (29 of 48) were missing, leaving 40% (19 of 48) available for inclusion in this study; all 19 agreed to participate. There were nine men and 10 women with a mean age of 46 ± 4 years and a mean time from surgery to examination of 28 ± 3 years. BMD was measured by dual-energy x-ray absorptiometry, and any fracture history was assessed using a questionnaire. Additionally, high-resolution peripheral quantitative CT was performed to compare the groups in terms of microarchitectural changes, such as cortical and trabecular area, cortical and trabecular thickness, cortical porosity, and endocortical perimeter. Participants in the control group were selected from a cohort consisting of a population-based random sample of 499 healthy adult women and men. Osteoporosis or low BMD was not an exclusion criterion for entering this study; however, the patients in the control group were selected based on a normal BMD (that is, T score > -1.0 at both the spine and hip). Also, the participants were matched based on age and sex. Differences between patients and controls were assessed using the Wilcoxon rank sum test for continuous variables and a chi-square test for categorical variables. A multiple regression analysis was performed. Model assumptions were checked using histograms and quantile-quantile plots of residuals. RESULTS Twelve of 19 patients who were treated for osteosarcoma had either osteopenia (eight patients) or osteoporosis (four patients). More patients with osteosarcoma reported sustaining fractures (11 of 19 patients) than did control patients (2 of 19 controls; p < 0.001). Among all microarchitectural parameters, only the endocortical perimeter was increased in patients compared with the control group (75 ± 15 mm versus 62 ± 18 mm; p = 0.04); we found no differences between the groups in terms of cortical and trabecular area, cortical and trabecular thickness, or cortical porosity. CONCLUSION Although patients who were treated for osteosarcoma had osteopenic or osteoporotic BMD and a higher proportion of patients experienced fractures than did patients in the control group, we could not confirm differences in microarchitectural parameters using high-resolution peripheral quantitative CT. Therefore, it seems that bone geometry and microstructural parameters are not likely the cause of the increased proportion of fractures observed in our patients who were treated for osteosarcoma. Until we learn more about the bone changes associated with chemotherapy in patients with osteosarcoma, we recommend that patients undergo regular BMD testing, and we recommend that physicians consider osteoporosis treatment in patients with low BMD. These data might provide the impetus for future multicenter prospective studies examining the association between chemotherapy and bone microarchitecture. LEVEL OF EVIDENCE Level III, therapeutic study.

Background Bone marrow edema (BME) is a localised painful bone lesion which is diagnosed by MRI and a frequent cause for severe pain in the joints of the lower limbs and. Ischemia, local osteoporosis and bone bruise/stress fractures are possible pathophysiological pathways. Besides these pathways BME also appear reactively within the inflammatory form of rheumatologic diseases as it is associated with future development of bone erosion. Only few data are available on bone micro structure of cortical and trabecular sites measured by high-resolution peripheral quantitative computed tomography (HR-pQCT) and serum bone turnover markers (BTM). The aim of this cross-sectional pilot study was to investigate bone mineral density (BMD) measured by DXA, bone micro structure and serum BTM values in patients presenting with arthralgia and BME at the lower limbs compared to age matched healthy controls (HC). Methods We compared 14 patients (mean age 43.7 ± 19.2 yr) with atraumatic BME of lower limb (2 femoral head, 7 proximal tibia, 5 ankle) to 35 age matched healthy controls (HC). All subjects had DXA (GE Lunar iDXA) of spine and hip, a HR-pQCT (Scanco Medical) examination of distal radius and tibia including microarchitectural parameters. The volume of interest was separated into a trabecular and a cortical region. The periosteal and endosteal boundaries were defined using an automated contouring. The region between the two contours was considered the cortical compartment volume for measuring cortical porosity. Further serum examinations of BSR, CRP and of intact amino terminal propeptide of type I procollagen (PINP), type 1 collagen cross-linked C-telopeptide (CTX), 25-OH vitamin D3, intact parathyroid hormone (iPTH) were measured. Results Areal BMD/BTM: BMD values measured by DXA as well as bone turnover marker (BTM), BSR and CRP were not statistically different between the groups. Mean BMD was in osteopenic range. HR-pQCT measurements-tibia: BME patients compared to HC had increased total bone area (TotalArea) (773.88 ± 238 vs 659.19 ± 113 mm², p < 0.05) and increased trabecular area (TrabArea) (689.89 ± 238.25 vs 555.74 ± 109.05 mm², p < 0.01), but a lower density of the compacta (Dcomp) (809.19 ± 65.78 vs 870.64 ± 74.49 mgHA/ccm, p < 0.01) and diminished average bone density (D100) (245.25±46.50 vs 286.98 ± 64.38 mgHA/ccm, p < 0.05) at the tibia compared to HC. Intracortical porosity (Ct.Po) at the tibia of patients with BME was significantly higher (8 ± 1.4 vs 5 ± 0.3 %, p < 0.05) and cortical thickness (Ct.th) (0.88 ± 0.24 vs 1.09 ± 0.31 mm, p < 0.05) was reduced. Trabecular thickness (Tb.th) (0.07 ± 0.01 vs 0.08 ± 0.01 mm, p < 0.05) was significantly decreased, whereas the number of trabeculae (Tb.N) did not differ from HC (1.83 ± 0.29 vs 1.74 ± 0.29 I/mm, p = 0.19). Conclusions Our data strongly suggest that altered structural properties at both, cortical and trabecular compartments contribute to the susceptibility of BME. An increased bone area is in contrast to reduced bone density, and an enhanced cortical porosity seems to be combined with reduced cortical and trabecular thickness. This structural impairment might be responsible for the development of atraumatic BME and our findings contribute to the understanding and to the treatment of this localised bone lesion. Disclosure of Interest None Declared

In recent years segmentation approaches based on sequential Monte Carlo Methods delivered promising results for the localization and delineation of anatomical structures in medical images. Also known as Shape Particle Filters, they were used for the segmentation of human vertebrae, lungs and hearts, being especially well suited to cope with the high levels of noise encountered in MR data and difficult overlaps in radiographs. This report surveys the robustness of these methods on different medical example images. A Differential Evolution approach on Shape Particle Filtering is applied for image segmentation. The goal of this work is to analyze the behavior, e.g. the robustness of the implemented Shape Particle Filter. Results on different data (synthetic rectangles, MRI slices and radiographs) are reported.

Fully-automated nuclear image segmentation is the prerequisite to ensure statistically significant, quantitative analyses of tissue preparations,applied in digital pathology or quantitative microscopy. The design of segmentation methods that work independently of the tissue type or preparation is complex, due to variations in nuclear morphology, staining intensity, cell density and nuclei aggregations. Machine learning-based segmentation methods can overcome these challenges, however high quality expert-annotated images are required for training. Currently, the limited number of annotated fluorescence image datasets publicly available do not cover a broad range of tissues and preparations. We present a comprehensive, annotated dataset including tightly aggregated nuclei of multiple tissues for the training of machine learning-based nuclear segmentation algorithms. The proposed dataset covers sample preparation methods frequently used in quantitative immunofluorescence microscopy. We...

The main challenges are discussed together with the lessons learned from past and ongoing research along the development cycle of machine learning systems. This will be done by taking into account intrinsic conditions of nowadays deep learning models, data and software quality issues and human-centered artificial intelligence (AI) postulates, including confidentiality and ethical aspects. The analysis outlines a fundamental theory-practice gap which superimposes the challenges of AI system engineering at the level of data quality assurance, model building, software engineering and deployment. The aim of this paper is to pinpoint research topics to explore approaches to address these challenges.

Diagnosis and risk stratification of cancer and many other diseases require the detection of genomic breakpoints as a prerequisite of calling copy number alterations (CNA). This, however, is still challenging and requires time-consuming manual curation. As deep-learning methods outperformed classical state-of-the-art algorithms in various domains and have also been successfully applied to life science problems including medicine and biology, we here propose Deep SNP, a novel Deep Neural Network to learn from genomic data. Specifically, we used a manually curated dataset from 12 genomic single nucleotide polymorphism array (SNPa) profiles as truth-set and aimed at predicting the presence or absence of genomic breakpoints, an indicator of structural chromosomal variations, in windows of 40,000 probes. We compare our results with well-known neural network models as well as Rawcopy though this tool is designed to predict breakpoints and in addition genomic segments with high sensitivity...

Various medical fields rely on detailed anatomical knowledge of the distal radius. Current studies are limited to two-dimensional analysis and biased by varying measurement locations. The aims were to 1) generate 3D shape models of the distal radius and investigate variations in the 3D shape, 2) generate and assess morphometrics in standardized cut planes, and 3) test the model's classification accuracy. The local radiographic database was screened for CT-scans of intact radii. 1) The data sets were segmented and 3D surface models generated. Statistical 3D shape models were computed (overall, gender and side separate) and the 3D shape variation assessed by evaluating the number of modes. 2) Anatomical landmarks were assigned and used to define three standardized cross-sectional cut planes perpendicular to the main axis. Cut planes were generated for the mean shape models and each individual radius. For each cut plane, the following morphometric parameters were calculated and com...

In recent years segmentation approaches based on sequential Monte Carlo Methods delivered promising results for the localization and delineation of anatomical structures in medical images. Also known as Shape Particle Filters, they were used for the segmentation of human vertebrae, lungs and hearts, being especially well suited to cope with the high levels of noise encountered in MR data and difficult overlaps in radiographs. This report surveys the robustness of these methods on different medical example images. A Differential Evolution approach on Shape Particle Filtering is applied for image segmentation. The goal of this work is to analyze the behavior, e.g. the robustness of the implemented Shape Particle Filter. Results on different data (synthetic rectangles, MRI slices and radiographs) are reported.

Purpose To characterize bone microarchitecture and quantify bone strength in lung transplant (LT) recipients by using high-resolution (HR) peripheral quantitative computed tomographic (CT) imaging of the ultradistal radius. Materials and Methods After study approval by the local ethics committee, all participants provided written informed consent. Included were 118 participants (58 LT recipients [mean age, 46.8 years ± 1.9; 30 women, 28 men] and 60 control participants [mean age, 39.9 years ± 1.9; 41 women, 19 men]) between April 2010 and May 2012. HR peripheral quantitative CT of the ultradistal radius was performed and evaluated for bone mineral density and trabecular and cortical bone microarchitecture. Mechanical competence was quantified by microfinite element analysis. Differences between LT recipients and control participants were determined by using two-way factorial analysis of covariance with age adjustment. Results Total and trabecular bone mineral density were significantly lower (-13.4% and -16.4%, respectively; P = .001) in LT recipients than in healthy control participants. LT recipients had lower trabecular number (-9.7%; P = .004) and lower trabecular thickness (-8.1%; P = .025). Trabecular separation and trabecular network heterogeneity were higher (+24.3% and +63.9%, respectively; P = .007 and P = .012, respectively) in LT recipients. Moreover, there was pronounced cortical porosity (+31.3%; P = .035) and lower cortical thickness (-10.2%, P = .005) after LT. In addition, mechanical competence was impaired, which was reflected by low stiffness (-15.0%; P < .001), low failure force (-14.8%; P < .001), and low bone strength (-14.6%; P < .001). Conclusion Men and women with recent LT showed severe deficits in cortical and trabecular bone…

Alcohol-induced chronic pancreatitis is associated with bone loss, but bone histomorphometric data describing the mechanism of cortical (Ct) and trabecular (Tb) bone loss are scarce. In this case-control study, we investigated 13 black male patients aged 41.2 +/- 8.9 years with alcohol-induced chronic pancreatitis by routine iliac crest cortical and trabecular histomorphometry and by biochemistry relevant to bone, liver function, and iron overload. Patients showed lower values for Ct thickness (P = 0.018), endocortical (Ec) wall thickness (P = 0.0002), Tb bone volume (0.019), Tb thickness (0.001), Tb wall thickness (P < 0.0001), Ec osteoid thickness (P = 0.001), Ec mineral apposition rate (P = 0.011), and Ec bone formation rate (P = 0.035). Ec eroded surface (P = 0.004) was elevated compared to controls. Tb osteoid thickness (P = 0.14) and Tb mineral apposition rate (P = 0.195) tended to be lower than in controls. Levels of 25-hydroxyvitamin D (P < 0.005), serum magnesium (P = 0.02), and ascorbic acid (P = 0.049) were lower and urine calcium/creatinine ratios higher than in controls. Alkaline phosphatase and gamma-glutamyl transpeptidase (GGT) were negatively correlated but iron markers were positively correlated with bone structural and formation variables. The histomorphometric data were found to be consistent with alcohol bone disease. Osteomalacia was not a feature. Secondary pathogenetic factors were liver disease, hypovitaminosis D and C, diabetes mellitus, and possibly chronic pancreatitis.

Hemodialysis (HD) patients face increased fracture risk, which is further associated with elevated risk of hospitalization and mortality. High-resolution peripheral computed tomography (HR-pQCT) has advanced our understanding of bone disease in chronic kidney disease by characterizing distinct changes in both the cortical and trabecular compartments. Increased cortical porosity (Ct.Po) has been shown to be associated with fracture in patients with osteopenia or in postmenopausal diabetic women. We tested whether the degree of Ct.Po identifies hemodialysis patients with prevalent fragility fractures in comparison to bone mineral density (BMD) assessed by dual X-ray absorptiometry (DXA). We performed a post-hoc analysis of a cross-sectional study in 76 prevalent hemodialysis patients. Markers of mineral metabolism, coronary calcification score, DXA-, and HR-pQCT-data were analyzed, and Ct.Po determined at radius and tibia. Ct.Po was significantly higher in patients with fracture but a...