CN117132573A - Method for predicting pulmonary nodule multiplication time based on single chest CT examination - Google Patents

Abstract

The application relates to the field of lung nodule monitoring, and discloses a lung nodule multiplication time prediction method based on single chest CT examination, which is suitable for being executed in computing equipment, and comprises the following steps: generating feature data for each sample in a pre-acquired sample set, the sample set comprising a plurality of lung nodules, the feature data comprising a location, a size, a proximity structure, and a calculated doubling time of the lung nodules; performing radiometric feature extraction on the CT image according to the feature data to obtain a first feature set; performing feature dimension reduction according to the first feature set to determine a training feature set; training the initial model according to the training feature set and the calculated multiplication time to obtain a prediction model; and inputting CT images of a single chest examination of the patient to be predicted into the model to obtain the predicted multiplication time of the lung nodule of the patient to be predicted. The application can predict the multiplication time of the lung nodule according to the CT image of the single examination of the lung nodule, thereby guiding the scientific examination treatment and reducing the radiation dose of multiple follow-up examinations.

Description

A prediction method for pulmonary nodule doubling time based on a single chest CT examination

Technical field

The present invention relates to the field of pulmonary nodule prediction, and in particular to a prediction method of pulmonary nodule doubling time based on a single chest CT examination.

Background technique

Pulmonary nodules can be divided into solid nodules and ground-glass nodules based on imaging findings. Ground glass nodules (GGN) can be divided into two categories: pure GGN (pGGN) without solid components; mixed GGN (mixed GGN, pGGN) with solid components and covering part of the lung texture. mGGN). Compared with solid nodules, GGN is more closely related to lung adenocarcinoma, which includes adenocarcinoma in situ (AIS), minimally invasive adenocarcinoma (MIA), and invasive adenocarcinoma (IAC). The growth prediction of GGN is of great value in judging its properties. Doubling time is an important indicator to evaluate whether a nodule will grow. Currently, multiple CT examinations are needed to calculate the doubling time and determine whether the nodule is growing. During the follow-up review period, patients are often anxious, and multiple CT examinations also increase the radiation dose.

In the existing technology, the doubling time of the nodule cannot be obtained based on the CT image of a single examination. To this end, a new prediction method for pulmonary nodule doubling time based on a single chest CT examination is needed.

Contents of the invention

To this end, the present invention provides a prediction method for pulmonary nodule doubling time based on a single chest CT examination, in an attempt to solve the above existing problems.

According to one aspect of the present invention, a method for predicting pulmonary nodule doubling time based on a single chest CT examination is provided, which is suitable for execution in a computing device. The method includes the steps of: generating features for each sample in a pre-acquired sample set Data, the sample set includes multiple chest CT examinations of multiple pulmonary nodules, and the characteristic data includes the location, size, density, shape, edge, tumor-lung interface, internal characteristics, adjacent structures and calculated doubling time of the pulmonary nodules; according to the required Perform radiomics feature extraction on CT images with the above feature data to obtain a first feature set; perform feature dimensionality reduction based on the first feature set to determine a training feature set; train the initial model based on the training feature set and calculate the doubling time to obtain Prediction model: input the CT image of a single chest examination of the patient to be predicted into the prediction model to obtain the predicted doubling time of the patient's pulmonary nodule to be predicted.

Optionally, in the method according to the present invention, the samples in the sample set are determined according to the selection criteria and the exclusion criteria.

Optionally, in the method according to the present invention, the selection criteria are the criteria that need to be met as sample pulmonary nodules, including: the diameter of the pulmonary nodule in the baseline CT scan is ≤3cm; more than two thin-section CT scans have been performed; baseline The time interval between the CT scan and the last CT scan before surgery was greater than or equal to 180 days; surgery was performed within one month of the last CT scan.

Optionally, in the method according to the present invention, it also includes: Exclusion criteria for pulmonary nodules that cannot be used as samples, including: nodules that have undergone any treatment before surgery, and the treatment includes radiotherapy or chemotherapy; lack of clinical information or Nodules that have lost nodule image data; nodule image artifacts affect nodule display.

Optionally, in the method according to the present invention, generating characteristic data for each sample in the pre-acquired sample set includes: the time interval from the baseline CT scan of the GGN to the last CT scan before surgery, the time interval of the GGN baseline CT scan The volume of the GGN and the volume of the last CT scan of the GGN before surgery are determined according to the following formula to calculate the doubling time:

VDT=(log2*T)/[log(V _f /V _i )]

Among them, T is the time interval from the baseline CT scan of the GGN to the last CT scan before surgery, V _i is the volume of the GGN at the baseline CT scan, and V _f is the volume of the last CT scan of the GGN before surgery.

Optionally, performing radiomic feature extraction on the CT image according to the feature data to obtain the first feature set includes: performing Gaussian Laplacian filter transformation and wavelet transform on the CT image, and extracting high-level features of the image to obtain the radiomic feature Set; extract first-order shape and second-order texture features from CT images to obtain radiomic features; use inter-class correlation coefficients to perform intra-observer and inter-observer consistency tests on radiomic features, and select from the radiomic feature set The radiomic features that meet the preset conditions for consistency testing are used as the first feature set.

Optionally, in the method according to the present invention, performing feature dimensionality reduction to determine the training feature set based on the first feature set includes: obtaining multiple predetermined features for the features in the first feature set through the minimum redundancy and maximum correlation methods. Assume a training feature set, each preset training feature set includes a different number of features; calculate the AUC for each preset training feature set based on the initial model, and use the preset training set with the largest AUC as the training feature set, where AUC is The average area under the receiver operating characteristic curve.

Optionally, in the method according to the present invention, the initial model includes: random forest, linear kernel support vector machine, radial basis function kernel support vector machine, linear extreme gradient boosting, K nearest neighbor, artificial neural network and naive Baye Sri Lanka.

Optionally, in the method according to the present invention, training the initial model according to the training feature set and calculating the doubling time to obtain the prediction model includes: using the determined training feature set, nesting according to the training samples of the training set in the sample set Cross-validation trains each initial model; the predictive model is determined from the initial model based on the AUC, accuracy, and stability of each model in the nested cross-validation training results.

According to another aspect of the present invention, a computing device is provided, including: one or more processors; a memory; and one or more programs, wherein the one or more programs are stored in the memory and configured to be configured by one or more A plurality of processors are executed, and one or more programs include instructions for executing any of the methods for predicting pulmonary nodule doubling time based on a single chest CT examination according to the present invention.

According to yet another aspect of the present invention, there is provided a computer-readable storage medium storing one or more programs, the one or more programs including instructions that, when executed by a computing device, cause the computing device to perform a method according to the present invention. Any one of the methods for predicting pulmonary nodule doubling time based on a single chest CT examination.

The present invention is a method for predicting pulmonary nodule doubling time based on a single chest CT examination, which is suitable for execution in a computing device. The method includes the steps of: generating characteristic data for each sample in a pre-acquired sample set, and the sample set includes multiple A pulmonary nodule, the characteristic data includes the location, size, density, shape, edge, tumor-lung interface, internal characteristics, adjacent structures and calculated doubling time of the pulmonary nodule; radiomics feature extraction is performed on the CT image based on the characteristic data Obtain the first feature set; perform feature dimensionality reduction according to the first feature set to determine the training feature set; train the initial model according to the training feature set and calculate the doubling time to obtain the prediction model; input the CT image of the patient to be predicted according to the prediction model, and obtain The predicted doubling time of pulmonary nodules in patients to be predicted. The present invention can train a prediction model based on CT images of pulmonary nodules, thereby predicting the doubling time of pulmonary nodules based on a single chest CT examination, thereby guiding scientific examination processing and reducing the radiation dose of multiple follow-up examinations.

Description of the drawings

To carry out the above and related purposes, certain illustrative aspects are described herein in conjunction with the following description and accompanying drawings, which are indicative of various ways in which the principles disclosed herein may be practiced, and all aspects and their equivalents are intended to within the scope of the claimed subject matter. The above and other objects, features and advantages of the present disclosure will become more apparent by reading the following detailed description in conjunction with the accompanying drawings. Throughout this disclosure, the same reference numbers generally refer to the same parts or elements.

Figure 1 shows a schematic flow chart of a method 100 for predicting pulmonary nodule doubling time based on a single chest CT examination according to an exemplary embodiment of the present invention;

Figure 2 shows a schematic diagram of the process of training a prediction model according to an exemplary embodiment of the present invention;

Figure 3 shows a schematic diagram of changes in the mean and standard deviation of features before and after data calibration according to an exemplary embodiment of the present invention;

Figure 4 shows a schematic diagram of the average AUC calculated from four feature sets according to an exemplary embodiment of the present invention;

Figure 5 shows a schematic diagram of the performance of seven models after training according to an exemplary embodiment of the present invention;

Figure 6 shows a schematic diagram of the test results of the stability of seven models according to an exemplary embodiment of the present invention;

Figure 7a shows a schematic diagram of the ROC curve of the NNet model according to an exemplary embodiment of the present invention;

Figure 7b shows a schematic diagram of the confusion matrix predicted by the NNet model according to an exemplary embodiment of the present invention;

Figures 8a and 8b respectively show a schematic diagram of the first patient's baseline CT scan and a schematic diagram of the last CT scan before surgery according to an exemplary embodiment of the present invention;

8c and 8d respectively show a schematic diagram of a baseline CT scan of a second patient and a schematic diagram of the last CT scan before surgery according to an exemplary embodiment of the present invention.

Detailed ways

Exemplary embodiments of the present disclosure will be described in more detail below with reference to the accompanying drawings. Although exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided to provide a thorough understanding of the disclosure, and to fully convey the scope of the disclosure to those skilled in the art. The same reference numbers generally refer to the same parts or elements.

Figure 1 shows a schematic flowchart of a method 100 for predicting pulmonary nodule doubling time based on a single chest CT examination according to an exemplary embodiment of the present invention. The method 100 for predicting pulmonary nodule doubling time based on a single chest CT examination of the present invention is suitable for execution in a computing device. The prediction method of pulmonary nodule doubling time based on a single chest CT examination starts from step 110, generating feature data for each sample in a pre-obtained sample set, the sample set includes multiple pulmonary nodules, and the feature data includes all Describe the location, size, density, morphology, edge, tumor-lung interface, internal characteristics, adjacent structures and calculation of doubling time of pulmonary nodules.

Figure 2 shows a schematic diagram of the process of training a prediction model according to an exemplary embodiment of the present invention. As shown in Figure 2, in order to establish a pulmonary nodule multiplication model, the present invention acquires CT images of pulmonary ground glass nodules (GGN) in advance. According to one embodiment of the present invention, nodule data of pulmonary ground-glass nodules confirmed by surgical pathology were collected from 4 hospitals as sample data. When collecting sample data, the samples in the sample set are determined based on the selection criteria and exclusion criteria.

Pulmonary ground-glass nodules that can be used as samples are selected according to the following selection criteria:

Pulmonary nodule diameter ≤3cm on baseline CT scan;

Have had more than two thin-section CT scans;

The time interval between the baseline CT scan and the last CT scan before surgery was greater than or equal to 180 days;

Surgery was performed within one month of the last CT scan.

And pulmonary ground-glass nodules that cannot be used as samples are excluded according to the following exclusion criteria:

Nodules that have undergone any treatment before surgery, including radiotherapy or chemotherapy;

Missing clinical information or missing nodule image data;

Nodule image artifacts or other factors affecting GGN display.

According to the above selection criteria and exclusion criteria, 176 GGNs of 172 patients from 4 hospitals were selected as samples, resulting in a sample set including 176 GGNs. Among the 172 patients, 112 were female and 60 were male, with a median age of 55 years old and an age range of 13 to 83 years old; among them: 31 cases were from the first hospital, 26 cases were from the second hospital, 33 cases were from the third hospital, and the remaining 82 patients were from the Fourth Hospital. According to an embodiment of the present invention, the present invention uses 90 GGNs of 90 patients from the first hospital, the second hospital and the third hospital as training samples to obtain a training set including 90 GGNs. The CT images of the GGNs in the training set are CT images of multiple chest examinations of multiple patients; 86 GGNs from 82 patients in the Fourth Hospital were used as test samples to obtain a verification set including 86 GGNs. The sample set includes training set and validation set.

According to an embodiment of the present invention, the GNNs used as samples all collect CT images. Acquired CT images are stored in the Image Archiving and Communications System (PACS) in Digital Imaging and Communications in Medicine (DICOM) format for retrieval. CT images of GNN can be collected through the following image acquisition equipment: GE MEDICAL SYSTEMS Discovery HD750CT, GE MEDICAL SYSTEMS Optima CT670, Philips Brilliance iCT, Philips IngenuityCT, SIEMENS SOMATOM Definition as and United Imaging CTu510.

During the CT image acquisition process of GNN, image segmentation was completed by two doctors, the first radiologist with 5 years of experience in chest CT image diagnosis and the second radiologist with 3 years of experience in chest CT image diagnosis. The first radiologist delineated the three-dimensional regions of interest (3D-ROI) of all samples in the sample set, i.e., 176 GGNs, and randomly selected 25 samples from the sample set, followed by the second radiologist 1 month later. Draw the 3D-ROI again on the GGN. When the second radiologist draws the 3D-ROI of 25 randomly selected samples, the lung window settings include a window width of 1400HU and a window level of -600HU. The GNN is drawn layer by layer in the axial plane to obtain the 3D-ROI.

3D-ROIs of 25 randomly selected samples were examined and corrected by a third radiologist to exclude cavities, calcifications, and adjacent structures such as blood vessels, bronchi, and pleura. The third radiologist is a senior doctor with 19 years of diagnostic experience in chest images.

The CT images and 3D-ROI of 176 GGNs and the CT images and 3D-ROI of 25 randomly selected GGNs were used as sample data sets. One sample data in each sample data set is the CT image and 3G-ROI of GGN. CT images and 3G-ROI of 176 GGNs were used for feature extraction and model construction, and CT images of 25 GGNs were used for consistency evaluation.

Subsequently, characteristic data of each sample data was generated based on each sample data (including CT images and 3G-ROI of 176 GGNs). The characteristic data is the CT image evaluation result of GGN. Feature data include location, size, density, morphology, margins, tumor-lung interface, internal features and adjacent structures.

Among them, the location includes the position where the GGN is equally divided into 1/3 of the hilus and the lung lobe where it is located; the location of pulmonary nodules is divided into three categories: inner 1/3, middle 1/3, and outer 1/3. The lobes of the lungs include the upper lobe of the right lung, the middle lobe of the right lung, the lower lobe of the right lung, the upper lobe of the left lung, and the lower lobe of the left lung. Different location characteristics are divided according to the location of nodules in different partitions.

The size includes the maximum cross-sectional diameter of the GGN and the minimum cross-sectional diameter perpendicular to it;

Density includes pure ground-glass nodules (pGGN) in which the ground-glass component is visible in the lung window but not in the mediastinal window or part-solid nodule (pGGN) which contains both solid and ground-glass components. PSN);

Shapes include irregular or round/round-like;

Marginal features include lobulations, spicules, and spinous processes, which are structures that extend from the lesion but, unlike the borders of the lung parenchyma, have at least a raised edge;

The tumor-lung interface includes three types: fuzzy boundary, clear and smooth boundary, and clear but rough boundary;

Internal features include vacuoles, cavities/cysts, cavities, calcifications, bronchial truncation, and bronchial tortuosity/dilatation;

Adjacent structures include pleural depression sign and vascular bundle sign;

Characteristic data also include whether the patient's bronchial walls are thickened and whether emphysema is present.

The differences in clinical and morphological features between GGNs in the training set and the validation set include a greater proportion of spiculation signs (P=0.009) and clear and rough tumor-lung interfaces (P<0.001) in the training set compared with the test set.

The present invention also calculates the calculated doubling time VDT of GGN. Specifically: based on the time interval from the baseline CT scan of GGN to the last CT scan before surgery, the volume of GGN during baseline CT scan, and the volume of GGN's last CT scan before surgery, according to Calculated using the following formula:

VDT=(log2*T)/[log(V _f /V _i )]

Among them, T is the time interval from the baseline CT scan of the GGN to the last CT scan before surgery, V _i is the volume of the GGN at the baseline CT scan, and v _f is the volume of the last CT scan of the GGN before surgery.

According to an embodiment of the present invention, the present invention uses VDT of 400 days as a threshold and divides GGN into two groups: VDT>400 days and VDT≤400 days. For ground-glass nodules, CT follow-up (with an interval of 3 to 12 months) is recommended for VDT >400 days, while further intervention is recommended for VDT ≤ 400 days.

In the sample set, compared with the VDT>400 days group: patients in the VDT≤400 days group were older (P=0.001), and more patients had emphysema (emphysema in the lobe where the nodule was located, P<0.001). emphysema in the remaining lung (P=0.011); the maximum diameter of the cross-section of the GGN is larger (P=0.013), the shape is more irregular (P=0.009), and the lobulation sign (P=0.028) and spiculation sign (P=0.017 ) and bronchial truncation sign (P=0.014) were more frequent.

Subsequently, step 120 is performed to perform radiomics feature extraction on the CT image according to the feature data to obtain the first feature set.

The present invention performs radiomics feature extraction and radiomics feature extraction on each training sample.

Before performing radiomics feature extraction, the CT images of GNN were preprocessed, including: resampling the volume element size to a resolution of 1×1×1mm3, and using a bandwidth of 25 to discretize the image grayscale.

The preprocessed CT images can be imaged using Pyradiomics software (version number 3.0.1 https://pyradiomics.readthedocs.io/en/latest/) that complies with the standards of the International Symposium on Biomedical Imaging (ISBI) Omics feature extraction, extracting first-order shape and second-order texture features.

When performing radiomics feature extraction, the CT image is subjected to Laplacian of Gaussian (LOG; sigma: 2-5) filter transformation and wavelet transformation to extract high-level features of the image. Finally, 1158 radiomic features of each GGN were obtained.

In order to select robust and reproducible features from the extracted features, the interclass correlation coefficient (ICC) was used to evaluate the radiomics features extracted from 25 GGNs delineated by the first radiologist and the second radiologist. Interobserver and interobserver agreement, and volumetric agreement of intraobserver and interobserver segmentation of radiomic features were calculated to calculate DICE coefficients.

In the test of intra- and inter-observer agreement assessing radiomic features, the median intra-observer ICC for 1158 radiomic features was 0.93 (25th, 75th percentiles: 0.86, 0.97 ), the median inter-observer ICC was 0.89 (25th and 75th percentiles: 0.77, 0.95). The 900 features that met ICC>0.75 in the inter-observer agreement analysis were retained for subsequent analysis. The intra- and inter-observer DICE coefficients were 0.88 [0.85, 0.90] and 0.82 [0.78, 0.83] respectively, indicating that the consistency of volume segmentation was acceptable.

Subsequently, the present invention uses combot to perform multi-center calibration on the samples from each hospital to reduce the impact of different data batches between centers. According to the different hospitals where the samples come from, the samples are divided into cz from the first hospital. center, the dy center from the Second Hospital, the wf center from the Third Hospital, and the zj center from the Fourth Hospital.

FIG. 3 shows a schematic diagram of changes in the mean and standard deviation of features before and after data calibration according to an exemplary embodiment of the present invention.

Part a of Figure 3 plots the mean ± standard deviation of each center before calibration, among which, cz center: -0.024±3.640; dy center: -0.052±2.997; wf center: 0.025±2.680; zj center: -0.378±4.818.

Part b in Figure 3 plots the mean ± standard deviation of each center after calibration, among which, cz center: -0.102±3.322; dy center: -0.059±3.290; wf center: -0.016±3.290; zj center: -0.120±3.457 .

Subsequently, for the calibrated samples, the present invention retains the features with an ICC greater than 0.75 obtained by calculating the DICE coefficient to evaluate the volume consistency of intra-observer and inter-observer segmentation of radiomics features, and eliminates the others. Features with an ICC greater than 0.75 are considered to be repeatable and highly robust.

Subsequently, step 130 is performed to perform feature dimensionality reduction based on the first feature set to determine a training feature set, and use the Minimum Redundancy-Maximum Relevance (MRMR) method to perform feature dimensionality reduction based on the eliminated features. Since the number of features affects the performance of the model, MRMR generates 4 feature sets, containing 3, 6, 9 and 12 features respectively. The average area under the receiver operating characteristic (ROC) curve of the seven initial models was used to determine the optimal number of features and the final feature set. The seven initial models are: random forest (RF), linear kernel support vector machine (support vector machines with Linear kernel, svmLinear), radial basis function kernel support vector machine (support Vector machines with Radial basis function kernel, svmRadial) ), linear extreme gradient boosting (extremegradient boosting with linear booster, xgbLinear), K-nearest neighbors (KNN), artificial neural network (NNet) and naive bayes (NB).

According to an embodiment of the present invention, Figure 4 shows a schematic diagram of the average AUC calculated from 4 feature sets according to an exemplary embodiment of the present invention. As shown in Figure 4, four types of AUC generated by MRMR under 7 initial models Among the average performances of feature sets with different numbers of features, the feature set containing 3 features has the highest average AUC (AUC=0.7984). The average AUCs of feature sets containing 3, 6, 9 and 12 features are 0.7984, 0.7971, 0.7776 and 0.7732 respectively. Therefore, the present invention selects a feature set containing three features for subsequent model training.

Subsequently, step 140 is performed to train the initial model according to the training feature set and the calculated doubling time to obtain a prediction model.

Using the determined feature set, nested cross-validation is performed to train each initial model based on the training samples of the training set. The CT images of GGN in the training set are CT images of multiple chest examinations of multiple patients. For each initial model, 5-fold cross-validation is performed internally to adjust model parameters and prevent model overfitting, and external cross-validation is 100 times of leave group out cross validation (LGOCV) to test the stability of the model. . In 100 times of LGOCV, each time the training set is divided into a training subset and a test subset. The training subset is used to train the initial model, and the test subset is used to test the model after training to obtain performance indicators, including AUC, accuracy degree, specificity, positive predictive value (PPV), negative predictive value (NPV), and area under the ROC curve (AUC). 100 times of LGOCV can produce 100 model performance indicators, and the relative standard deviation (RSD) is used to evaluate the stability of each model. The calculation formula is as follows:

RSD%=(SD _metric /Mean _metric )*100%

In the formula, SD _metric represents the standard deviation of 100 metric values, Mean _metric represents the average of 100 metric values, and metric (metric value) is accuracy or AUC respectively. The lower the RSD value, the higher the stability of the model.

Finally, the model with good prediction performance and high stability is selected as the final model, and the model is evaluated through an independent validation set. Calculate the accuracy, sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), and area under the ROC curve (AUC) of the model.

According to an embodiment of the present invention, the present invention can use R software (version 4.2.0, https://cran.r-project.org/) and SPSS23.0 for Windows (SPSS, Chicago, Illinois) in statistical analysis ))conduct statistical analysis. Categorical variables were analyzed using the χ2 test or FISHER exact test, and continuous variables were analyzed using the Mann-Whitney test. Use ROC analysis to examine the performance of the model. The larger the AUC, the better the performance of the model, and a two-tailed P value <0.05 is considered statistically significant.

Figure 5 shows a performance diagram after training of seven initial models according to an exemplary embodiment of the present invention. As shown in Figure 5, in the nested cross-validation results, the xgbLiner algorithm model has the best prediction performance, with accuracy and AUC of 0.890 and 0.896 respectively; the performance of the NNet algorithm model is second, with accuracy and AUC of 0.865 and 0.886 respectively; and The NB algorithm model has the lowest prediction performance among the seven algorithms, with accuracy and AUC being only 0.748 and 0.790 respectively. The specific performance parameters are shown in Table 1:

Table 1

The data in the table are means ± standard deviation

In order to select the best model among the 7 models, the stability of the 7 models was also evaluated.

Figure 6 shows a schematic diagram of the test results of the stability of seven models according to an exemplary embodiment of the present invention. As shown in Figure 6: The left side is the RSD of the average accuracy, and the right side is the RSD of the average AUC. The smaller the abscissa value, the better the stability of the table algorithm prediction; the larger the ordinate value, the higher the average accuracy or average AUC of the model.

The results show that the NNet algorithm model has the highest robustness against data interference, with an accuracy and AUC RSD of 11.9% and 10.9% respectively; while the xgbLiner algorithm model with the best prediction performance has an accuracy and AUC RSD of 14.4% respectively. (third place) and 14.9% (second place). The RSD of the accuracy of the NB algorithm model (24.5%) and the RSD of the AUC of the svmRadial algorithm model (22.2%) are the highest and the robustness is the worst.

The stability parameter test results of different models are shown in Table 2:

Table 2

As shown in Table 2, the lower the value of RSD, the better the stability of the model.

Furthermore, compare the average AuC difference between the NNet model and other models. Although the average AUC value of the NNet model is slightly lower than the xgbLiner model, the difference is not statistically significant (P=0.851). Comparison of the average AUC values of the NNet model and other models as shown in Table 3:

table 3

In summary, considering the prediction performance and robustness of the model, the present invention selected the most robust NNet model as the prediction model, and further evaluated the NNet model on an independent verification set.

Figure 7a shows a schematic diagram of the NNet model ROC curve according to an exemplary embodiment of the present invention; Figure 7b shows a schematic diagram of the NNet model prediction confusion matrix according to an exemplary embodiment of the present invention.

As shown in Figure 7a and Figure 7b, the confusion matrix of ROC analysis and prediction shows that the AUC of the model is 0.709 (95% CI: 0.515~0.879); the accuracy, sensitivity, specificity, PPV and NPV are 0.756 and 0.667 respectively. , 0.766, 0.250 and 0.952.

Finally, step 150 is executed, and the CT image of a single chest examination of the patient to be predicted is input into the prediction model to obtain the predicted doubling time of the pulmonary nodules of the patient to be predicted. The prediction model in the present invention is trained based on CT images of multiple chest examinations. In actual prediction, prediction doubling time prediction can be achieved using only CT images of a single chest examination, thereby guiding scientific examination processing and reducing multiple follow-up examinations. radiation dose, shorten the patient’s examination cycle, reduce the number of hospital visits, and improve the patient’s medical experience.

The NNet model is trained according to the present invention to predict the CT images of pulmonary nodules of 2 patients in the verification set. 8a and 8b respectively show a schematic diagram of a first patient's baseline CT scan and a schematic diagram of the last CT scan before surgery according to an exemplary embodiment of the present invention. 8c and 8d respectively show a schematic diagram of a baseline CT scan of a second patient and a schematic diagram of the last CT scan before surgery according to an exemplary embodiment of the present invention.

The first patient is a 65-year-old male. As shown in Figure 8a, the baseline CT showed mGGN in the upper lobe of the left lung. As shown in Figure 8b, the patient’s last preoperative CT follow-up showed that the nodule increased significantly, the solid component increased, and both The scan interval was 343 days, the doubling time predicted by the model was ≤400 days, the actual calculated VDT was 267 days, and the pathology confirmed invasive adenocarcinoma.

The second patient was a 55-year-old female. As shown in Figure 8c, the baseline CT showed pGGN in the right middle lobe of the lung. As shown in Figure 8d, the patient's last preoperative CT follow-up showed no significant changes in the nodules. The interval between two scans was 273 days, the doubling time predicted by the model was >400 days, the actual calculated VDT was 1000 days, and the pathology confirmed it was adenocarcinoma in situ.

It should be noted that the storage medium (computer-readable medium) mentioned above in this application may be a computer-readable signal medium or a non-transitory computer-readable storage medium, or any combination of the above two. The non-transitory computer-readable storage medium may be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, device or device, or any combination thereof. More specific examples of non-transitory computer readable storage media may include, but are not limited to: an electrical connection having one or more wires, a portable computer disk, a hard drive, random access memory (RAM), read only memory (ROM), Erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination of the above.

As used herein, a non-transitory computer-readable storage medium may be any tangible medium that contains or stores a program for use by or in connection with an instruction execution system, apparatus, or device. In this application, the computer-readable signal medium may include a data signal propagated in baseband or as part of a carrier wave, in which computer-readable program code is carried. Such propagated data signals may take many forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination of the above. A computer-readable signal medium may also be any computer-readable medium other than non-transitory computer-readable storage media that can be sent, propagated, or transmitted for use by or in connection with an instruction execution system, apparatus, or device program of. Program code embodied on a computer-readable medium may be transmitted using any suitable medium, including but not limited to: wire, optical cable, RF (radio frequency), etc., or any suitable combination of the above.

The above-mentioned computer-readable medium may be included in the above-mentioned electronic device; it may also exist independently without being assembled into the electronic device.

Computer program code for performing the operations of the present application may be written in one or more programming languages, including but not limited to object-oriented programming languages such as Java, Smalltalk, C++, or a combination thereof. Includes conventional procedural programming languages, such as the "C" language or similar programming languages. The program code may execute entirely on the user's computer, partly on the user's computer, as a stand-alone software package, partly on the user's computer and partly on a remote computer or entirely on the remote computer or server.

The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operations of possible implementations of systems, methods, and computer program products according to various embodiments of the present application. In this regard, each block in the flowchart or block diagram may represent a module, segment, or portion of code that contains one or more logic functions that implement the specified executable instructions. It should also be noted that, in some alternative implementations, the functions noted in the block may occur out of the order noted in the figures. For example, two blocks shown one after the other may actually execute substantially in parallel, or they may sometimes execute in the reverse order, depending on the functionality involved. It will also be noted that each block of the block diagram and/or flowchart illustration, and combinations of blocks in the block diagram and/or flowchart illustration, can be implemented by special purpose hardware-based systems that perform the specified functions or operations. , or can be implemented using a combination of specialized hardware and computer instructions.

The units involved in the embodiments of this application can be implemented in software or hardware. Among them, the name of a unit does not constitute a limitation on the unit itself under certain circumstances.

The functions described above herein may be performed, at least in part, by one or more hardware logic components. For example, and without limitation, exemplary types of hardware logic components that may be used include: Field Programmable Gate Arrays (FPGAs), Application Specific Integrated Circuits (ASICs), Application Specific Standard Products (ASSPs), Systems on a Chip (Systems on a Chip), Complex Programmable Programmable logic device (CPLD), etc.

The above description is only some embodiments of the present application and an explanation of the technical principles used. Persons skilled in the art should understand that the disclosure scope involved in this application is not limited to technical solutions composed of specific combinations of the above technical features, but should also cover solutions consisting of the above technical features or without departing from the above disclosed concept. Other technical solutions formed by any combination of equivalent features. For example, a technical solution is formed by replacing the above features with technical features with similar functions disclosed in this application (but not limited to).

Furthermore, although operations are depicted in a specific order, this should not be understood as requiring that the operations be performed in the specific order shown, or in sequential order. Under certain circumstances, multitasking and parallel processing may be advantageous. Likewise, although several specific implementation details are included in the above discussion, these should not be construed as limiting the scope of the application. Certain features that are described in the context of separate embodiments can also be implemented in combination in a single embodiment. Conversely, various features that are described in the context of a single embodiment can also be implemented in multiple embodiments separately or in any suitable subcombination.

Although the subject matter has been described in language specific to structural features and/or methodological acts, it is to be understood that the subject matter defined in the appended claims is not necessarily limited to the specific features or acts described above. Rather, the specific features and acts described above are merely example forms of implementing the claims.

Claims (11)

1. A method of predicting lung nodule doubling time based on a single chest CT examination, adapted for execution in a computing device, the method comprising the steps of:

generating feature data for each sample in a pre-acquired sample set, the sample set comprising a plurality of lung nodules, the feature data comprising a location, a size, a density, a morphology, an edge, a tumor lung interface, internal features, an adjacent structure, and a calculated doubling time of the lung nodules;

performing radiometric feature extraction on the CT image according to the feature data to obtain a first feature set;

performing feature dimension reduction according to the first feature set to determine a training feature set;

training the initial model according to the training feature set and the calculated multiplication time to obtain a prediction model;

and inputting CT images of a single chest examination of the patient to be predicted into the model to obtain the predicted multiplication time of the lung nodule of the patient to be predicted.

2. The method of claim 1, wherein the samples in the sample set are determined based on selection criteria and exclusion criteria.

3. The method of claim 2, wherein the selection criteria are criteria that need to be met as a sample lung nodule, comprising:

the diameter of a lung nodule in baseline CT scanning is less than or equal to 3cm;

performing thin layer CT scanning for more than two times;

the time interval between the baseline CT scan and the last CT scan prior to surgery is greater than or equal to 180 days;

surgery was performed within one month after the last CT scan.

4. A method according to claim 2 or 3, wherein the exclusion criteria are criteria for lung nodules that cannot be taken as a sample, comprising:

any treated nodules prior to surgery, including radiation or chemotherapy;

nodules lacking clinical information or missing nodule image data;

the nodule image artifacts affect the nodule display.

5. The method of claim 1, wherein the generating feature data for each sample in the set of pre-acquired samples comprises:

according to the time interval from the baseline CT scanning of the GGN to the last CT scanning before the operation, the volume of the GGN at the baseline CT scanning and the volume of the GGN at the last CT scanning before the operation, the multiplication time is determined and calculated according to the following formula:

VDT＝(log2*T)/[log( _f / _i )]

wherein T is the time interval from baseline CT scan to last CT scan before surgery for GGN, V _i Is GGN baseline CT scanVolume at time V _f Is the volume of the last CT scan of the preoperative GGN.

6. The method of claim 1, wherein performing a radiological feature extraction on a CT image from the feature data to obtain a first feature set comprises:

performing Gaussian Laplace filtering transformation and wavelet transformation on the CT image, and extracting advanced features of the image to obtain a radiology feature set;

extracting first-order shape and second-order texture features from the CT image to obtain image histology features;

and carrying out consistency test on the image histology characteristics in and among observers by using the correlation coefficient among the classes, and selecting the radiology characteristics meeting the preset conditions by the consistency test from the radiology characteristic set as a first characteristic set.

7. The method of claim 1, wherein the feature-dimension-reduction determination of a training feature set from the first feature set comprises:

obtaining a plurality of preset training feature sets for the features in the first feature set through a minimum redundancy and maximum association method, wherein the number of the features included in each preset training feature set is different;

and calculating an AUC (automatic value) for each preset training feature set according to the initial model, and taking the preset training set with the largest AUC as the training feature set, wherein the AUC is the average area under the working feature curve of the subject.

8. The method of claim 1, wherein the initial model comprises: random forests, linear kernel support vector machines, radial basis function kernel support vector machines, linear extreme gradient boosting, K nearest neighbors, artificial neural networks, and naive bayes.

9. The method of any of claims 1-8, wherein the training the initial model from the training feature set and calculating doubling time to obtain a predictive model comprises:

using the determined training feature set, and performing nested cross-validation training on each initial model according to training samples of the training set in the sample set;

and determining a prediction model from the initial model according to the AUC, accuracy and stability of each model in the nested cross-validation training result.

10. A computing device, comprising:

one or more processors;

a memory; and

one or more devices comprising instructions for performing any of the methods of claims 1-9.

11. A computer readable storage medium storing one or more programs, the one or more programs comprising instructions, which when executed by a computing device, cause the computing device to perform any of the methods of claims 1-9.