Pancreatic cancer early recurrence prediction method and system based on interpretable machine model

Abstract

The present application relates to the technical field of medical imageomics, and particularly relates to a pancreatic cancer early recurrence prediction method and system based on an interpretable machine model, comprising the following steps: extracting intratumoral and peritumoral imageomics features in CT images; performing single factor analysis and multivariate logistic regression analysis on body composition parameters and clinical pathological data to obtain clinical features; constructing six groups of classifier models based on the intratumoral and peritumoral imageomics features through six machine learning algorithms, and obtaining an imageomics model according to model performance comparison; constructing a clinical-imageomics combined model by combining the clinical features and the imageomics model, and performing an interpretable SHAP analysis on the clinical-imageomics combined model. The present application combines intratumoral and peritumoral CT imageomics features with body composition parameters, constructs a machine learning model for predicting the early recurrence risk of PDAC after resection, and incorporates the interpretable SHAP analysis to enhance the transparency of the machine learning model decision-making process.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of medical imageomics, and in particular to a pancreatic cancer early recurrence prediction method and system based on an interpretable machine model. BACKGROUND

[0002] Pancreatic ductal adenocarcinoma (PDAC) is one of the most devastating malignancies, and computed tomography (CT) is the preferred imaging modality for diagnosis and assessment of PDAC. In recent years, as a technique for analyzing image features in a region of interest, imageomics has shown excellent results in predicting tumor biological complexity and intra-tumor heterogeneity. Through high-throughput analysis of CT images within the tumor, imageomics has also been used to predict the biological behavior of PDAC tumors. Although most studies have focused on primary tumors, recent studies have highlighted the important auxiliary role of the peritumoral region in assessing cancer heterogeneity and its importance in prognostic prediction. In addition to tumor characteristics, individual differences in patients are also key factors affecting the recurrence of pancreatic cancer. Previous studies have shown that body composition, as an indicator of patient nutritional status and metabolism, is an important risk factor for PDAC recurrence and poor prognosis. However, this easily accessible parameter is often overlooked in the development of prognostic models.

[0003] In the prior art, pancreatic cancer recurrence prediction is usually constructed by machine learning algorithms to build a prognosis model, and only imageomics features are used, ignoring individual state and metabolic indicator features, which limits the performance of prognosis recurrence prediction. At the same time, due to the complexity and lack of interpretability of machine learning models, pancreatic cancer recurrence prediction lacks credibility. SUMMARY

[0004] The present application aims to provide a pancreatic cancer early recurrence prediction method and system based on an interpretable machine model, to solve the technical problem that the prior art ignores individual state and metabolic indicator features, which limits the performance of prognosis recurrence prediction, and at the same time, due to the complexity and lack of interpretability of machine learning models, pancreatic cancer recurrence prediction lacks credibility.

[0005] To solve the above technical problems, the present application specifically provides the following technical solutions:

[0006] A pancreatic cancer early recurrence prediction method and system based on an interpretable machine model, comprising the following steps:

[0007] Obtaining CT image data, clinical pathology data, and postoperative follow-up data of a case;

[0008] Segmenting the intratumoral and peritumoral regions in the CT image to determine the regions of interest VOIs, and extracting intratumoral and extratumoral imageomics features in the regions of interest VOIs;

[0009] quantitative analysis of the body composition in the CT image to obtain body composition parameters, and performing single factor analysis and multivariate logistic regression analysis on the body composition parameters and the clinical pathological data to obtain clinical characteristics;

[0010] Six sets of classifier models were constructed based on the intratumoral and extratumoral imaging features and postoperative follow-up data by using six machine learning algorithms, and the imaging feature model was obtained by comparing the performances of the six sets of classifier models;

[0011] A clinical-imaging feature combined model for predicting early recurrence of pancreatic cancer was constructed by combining the clinical characteristics and the imaging feature model, and SHAP analysis was performed on the clinical-imaging feature combined model.

[0012] As a preferred scheme of the present application, the method for determining the intratumoral and extratumoral imaging features comprises:

[0013] A pre-established neural network model is used to segment the tumor region of interest and the tumor peripheral region of interest on the CT image;

[0014] The CT image is resampled to a standardized pixel size of 1.0 × 1.0 × 1.0 mm 3 , the gray level is discretized into a fixed interval of 5, and Pyradiomics is used to extract the imaging features of the tumor region of interest and the tumor peripheral region of interest as the intratumoral and extratumoral imaging features, wherein the intratumoral and extratumoral imaging features include 216 first-order features, 264 gray level co-occurrence matrix GLCM features, 168 gray level dependence matrix GLDM features, 192 gray level run-length matrix GLRLM features, 192 gray level size zone matrix GLSZM features, 60 adjacent gray level difference matrix NGTDM features and 14 shape features.

[0015] As a preferred scheme of the present application, the neural network model includes three sets of Unet network structures, the first set of Unet network structures is used to segment the tumor region of interest and the tumor peripheral region of interest in the CT image, the second set of Unet network structures is used to correct the tumor region of interest, and the third set of Unet network structures is used to correct the tumor peripheral region of interest;

[0016] The segmentation formula of the first set of Unet network structures is: (maskA, maskB) = Unet1(G CT );

[0017] The segmentation formula of the second set of Unet network structures is:

[0018] maskA_fixed = Add(Unet2(Concat(maskA, GCT )), maskA);

[0019] The segmentation of the third group of Unet network structure is:

[0020] maskB_fixed=Add(Unet3(Concat(maskB, G CT )), maskB);

[0021] In the formula, maskA is the tumor region of interest segmented by the first group of Unet network structure, maskB is the tumor peripheral region of interest segmented by the first group of Unet network structure, G CT is a CT image, Unet1 is the first group of Unet network structure, maskA_fixed is the corrected maskA of the second group of Unet network structure, Unet2 is the second group of Unet network structure, maskB_fixed is the corrected maskB of the third group of Unet network structure, Unet3 is the third group of Unet network structure, Concat is a channel dimension merging operator, and Add is a feature dimension merging operator.

[0022] The loss function of the neural network model is: ;

[0023] ;

[0024] ;

[0025] ;

[0026] In the formula, is the total loss, is the segmentation loss of the tumor region of interest, is the segmentation loss of the tumor peripheral region of interest, is the reconstruction loss of the tumor region of interest and the tumor peripheral region of interest, is a hyperparameter, is the gold standard for the tumor region of interest segmented alone in the CT image, is the gold standard for the tumor peripheral region of interest segmented alone in the CT image, is the gold standard for the tumor region of interest and the tumor peripheral region of interest segmented simultaneously in the CT image, and MSE is a mean square error operator.

[0027] As a preferred scheme of the present application, the quantitative analysis method of the body composition parameter comprises:

[0028] The skeletal muscle index SMI, the subcutaneous fat index SFI, the intermuscular fat index IMFI, the visceral fat index VFI, and the skeletal muscle density SMD are obtained by SliceOMatic V5.0 to measure the cross-sectional images of the third lumbar vertebra level of the CT image;

[0029] The subcutaneous fat-visceral fat ratio VSR is obtained by dividing the subcutaneous fat index SFI by the visceral fat index VFI;

[0030] The visceral fat-skeletal muscle tissue ratio VMR is obtained by dividing the skeletal muscle index SMI by the visceral fat index VFI.

[0031] As a preferred scheme of the present application, the method for determining the clinical characteristics comprises:

[0032] The clinical factors statistically significant p<0.05 in the body composition parameters and the clinicopathological data are screened out by single factor analysis;

[0033] The clinical factors are input into multivariate logistic regression analysis to screen out independent predictors for identifying the risk of early recurrence of pancreatic cancer, as the clinical characteristics, which include serum CA12-5, visceral fat index VFI, adjuvant therapy, lymphatic vessel invasion LVI, tumor grade and TNM stage.

[0034] As a preferred scheme of the present application, the method for obtaining the postoperative follow-up data comprises:

[0035] According to the follow-up scheme, follow-up is performed once every 3 months in the first two years, once every 6 months from 3 to 4 years, and once a year thereafter, wherein the follow-up scheme includes physical examination, serum tumor marker detection and abdominal and pelvic CT scan;

[0036] According to the reference time of recurrence of pancreatic ductal adenocarcinoma after surgery, cases that recur within 1 year after surgery are marked as the ER group, and cases that do not recur or recur more than 1 year after surgery are marked as the non-ER group.

[0037] As a preferred scheme of the present application, the method for determining the imaging omics model comprises:

[0038] The imaging omics features with ICC≥0.75 are retained in the intra- and extra-tumor imaging omics features by correlation coefficient ICC evaluation;

[0039] In the imaging omics features with ICC≥0.75, 7 intra-tumor imaging omics features, 9 peritumor imaging omics features and 8 intra- and peritumor imaging omics features are obtained by maximum correlation minimum redundancy mRMR method and least absolute shrinkage and selection operator LASSO method analysis;

[0040] Six classification models for predicting early recurrence of pancreatic cancer based on intratumoral imaging features were constructed by LR, SVM, RF, ExtraTrees, LightGBM and MLP algorithms based on 7 intratumoral imaging features and case labels in postoperative follow-up data.

[0041] Six classification models for predicting early recurrence of pancreatic cancer based on peritumoral imaging features were constructed by LR, SVM, RF, ExtraTrees, LightGBM and MLP algorithms based on 9 peritumoral imaging features and case labels in postoperative follow-up data.

[0042] Six classification models for predicting early recurrence of pancreatic cancer based on intratumoral-peritumoral imaging features were constructed by LR, SVM, RF, ExtraTrees, LightGBM and MLP algorithms based on 8 intratumoral-peritumoral imaging features and case labels in postoperative follow-up data.

[0043] Based on AUC index, ACC index, SEN index, SPE index, PPV index and NPV index, the performance differences between the three types of classifiers, i.e., the classification model for predicting early recurrence of pancreatic cancer based on intratumoral imaging features, the classification model for predicting early recurrence of pancreatic cancer based on peritumoral imaging features and the classification model for predicting early recurrence of pancreatic cancer based on intratumoral-peritumoral imaging features, were evaluated, and it was found that the classification model for predicting early recurrence of pancreatic cancer based on intratumoral-peritumoral imaging features was superior to the classification model for predicting early recurrence of pancreatic cancer based on intratumoral imaging features and the classification model for predicting early recurrence of pancreatic cancer based on peritumoral imaging features.

[0044] Based on AUC index, ACC index, SEN index, SPE index, PPV index and NPV index, the performance differences between the six classification models for predicting early recurrence of pancreatic cancer based on intratumoral-peritumoral imaging features were evaluated, and it was found that the classification model for predicting early recurrence of pancreatic cancer based on LR and intratumoral-peritumoral imaging features was superior to the classification model for predicting early recurrence of pancreatic cancer based on LR, SVM, ExtraTrees, LightGBM, MLP and intratumoral-peritumoral imaging features.

[0045] The classification model for predicting early recurrence of pancreatic cancer based on LR and intratumoral-peritumoral imaging features was used as an imaging feature model.

[0046] As a preferred scheme of the present application, the determination method of the clinical-imaging feature combined model comprises:

[0047] The serum CA12-5, visceral fat index VFI, adjuvant therapy, lymphatic vessel invasion LVI, tumor grade and TNM stage are combined into the imageomics model to obtain the clinical-imageomics combined model;

[0048] The performance of the clinical-imageomics combined model is evaluated by the ROC index, the AUC index and the DCA index.

[0049] As a preferred scheme of the present application, the method for performing the explainability SHAP analysis on the clinical-imageomics combined model comprises:

[0050] The global explanation of the clinical-imageomics combined model is performed at the feature level by the SHAP visualization analysis to obtain the contribution degrees of the intratumoral-peritumoral imageomics features and the clinical features to the output of the clinical-imageomics combined model, wherein the intratumoral-peritumoral imageomics features have the largest contribution, followed by the clinical features, and the high imageomics feature values are closely related to the increased risk of early recurrence, the high visceral fat index VFI, CA12-5 ≥ 35U / ml, lymphatic vessel invasion LVI positive, TNM stage ≥ IIA and no adjuvant therapy are all related to the increased risk of early recurrence, and the intratumoral-peritumoral imageomics features and the VFI have the interaction of increasing the risk of early recurrence;

[0051] The local explanation of the clinical-imageomics combined model is performed at the individual level by the SHAP visualization analysis to obtain the cases without early recurrence of pancreatic ductal adenocarcinoma and the cases with early recurrence of pancreatic ductal adenocarcinoma during follow-up, and the matching decision logic between the contribution degrees of the intratumoral-peritumoral imageomics features and the clinical features to the output of the clinical-imageomics combined model.

[0052] As a preferred scheme of the present application, the present application provides a pancreatic cancer early recurrence prediction system based on an explainable machine model, which is applied to a pancreatic cancer early recurrence prediction method based on an explainable machine model, and the system comprises:

[0053] A data acquisition unit is configured to acquire CT image data, clinical pathology data and postoperative follow-up data of a case;

[0054] A feature processing unit is configured to segment intratumoral and peritumoral regions in the CT image, determine the VOIs, and extract intratumoral and peritumoral imageomics features in the VOIs; perform quantitative analysis on the body composition in the CT image to obtain body composition parameters, and perform single factor analysis and multivariate logistic regression analysis on the body composition parameters and the clinical pathology data to obtain clinical features;

[0055] The model construction unit constructs six classifier models based on intratumoral and peritumoral radiomics features and postoperative follow-up data through six machine learning algorithms, and obtains the radiomics model according to the performance comparison of the six classifier models; the clinical features and the radiomics model are used to construct a clinical-radiomics combined model for early recurrence prediction of pancreatic cancer, and the clinical-radiomics combined model is subjected to SHAP analysis for interpretability.

[0056] Compared with the prior art, the present application has the following beneficial effects:

[0057] The present application combines intratumoral and peritumoral CT radiomics features with body composition parameters to construct a machine learning model for predicting the early recurrence risk of PDAC after resection, and incorporates SHAP analysis for interpretability, enhances the transparency of the decision-making process of the machine learning model, improves the credibility and interpretability of the machine learning model, increases the trust of clinicians in the prediction results, and helps to develop personalized treatment strategies. BRIEF DESCRIPTION OF DRAWINGS

[0058] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the following will briefly introduce the drawings needed to be used in the embodiments or prior art description. Obviously, the drawings in the following description are only exemplary, and for those skilled in the art, other drawings can be obtained without creative labor on the basis of the provided drawings.

[0059] Figure 1 The flow chart of the pancreatic cancer early recurrence prediction method based on the interpretable machine model provided by the embodiment of the present application;

[0060] Figure 2 The block diagram of the pancreatic cancer early recurrence prediction system based on the interpretable machine model provided by the embodiment of the present application;

[0061] Figure 3 The performance comparison chart of intratumoral radiomics features, peritumoral radiomics features and intratumoral-peritumoral radiomics features in six machine learning algorithms in the training set provided by the embodiment of the present application;

[0062] Figure 4 The performance comparison chart of intratumoral radiomics features, peritumoral radiomics features and intratumoral-peritumoral radiomics features in six machine learning algorithms in the test set provided by the embodiment of the present application;

[0063] Figure 5 The performance comparison chart of the clinical feature classification model, the radiomics model and the clinical-radiomics combined model provided by the embodiment of the present application;

[0064] Figure 6A bar chart of global explanation in the SHAP method provided by the embodiment of the present application;

[0065] Figure 7 A bar chart of local explanation in the SHAP method provided by the embodiment of the present application;

[0066] Figure 8 A dependency graph of the SHAP method provided by the embodiment of the present application;

[0067] Figure 9 A schematic diagram of a pancreatic tumor segmentation candidate region provided by the embodiment of the present application;

[0068] Figure 10 The neural network model structure flowchart provided by the embodiment of the present application. DETAILED DESCRIPTION

[0069] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the protection scope of the present application.

[0070] As shown in Figure 1 The present application provides a pancreatic cancer early recurrence prediction method based on an interpretable machine model, which comprises the following steps:

[0071] Obtaining CT image data, clinical pathology data and postoperative follow-up data of a case;

[0072] Segmenting an intratumoral and peritumoral region in the CT image, determining a volume of interest (VOI) and extracting intratumoral and extratumoral imageomic features in the VOI;

[0073] Quantitatively analyzing body composition in the CT image to obtain body composition parameters, and performing single factor analysis and multivariate logistic regression analysis on the body composition parameters and the clinical pathology data to obtain clinical features;

[0074] Constructing six groups of classifier models based on the intratumoral and extratumoral imageomic features and the postoperative follow-up data through six machine learning algorithms, and obtaining an imageomic model according to performance comparison of the six groups of classifier models;

[0075] A clinical-imaging feature combined model for predicting early recurrence of pancreatic cancer is constructed by combining clinical features and imaging feature models, and a SHAP (Shapley Additive exPlanations) analysis is performed on the clinical-imaging feature combined model.

[0076] The imaging feature is used to construct a machine learning model for predicting early recurrence of pancreatic cancer, i.e., an imaging feature model, to realize early recurrence prediction of pancreatic cancer based on the imaging feature.

[0077] In addition to the imaging feature, body composition reflects the nutritional metabolic state of the patient, which has been proven to be an important factor affecting tumor recurrence and survival outcome. At the same time, individual differences are also a key factor affecting the recurrence of pancreatic cancer. Therefore, the imaging feature is combined with body composition and individual pathological features, such as serum CA12-5 (carbohydrate antigen 12-5), adjuvant therapy, lymphatic vessel invasion LVI, tumor grade and TNM stage (tumor-lymph node-metastasis stage), to form a clinical-imaging feature combined model with better performance based on the imaging feature model, thereby improving the performance of early recurrence prediction of pancreatic cancer.

[0078] To ensure the reproducibility and reliability of the imaging feature, the segmentation is repeated in 30 randomly selected cases, and the intra-group and inter-group consistency is calculated after 2 weeks. The intra-group correlation coefficient (ICC) is calculated to evaluate the performance. Subsequently, the imaging features with good intra-group and inter-group consistency (ICC≥0.75) are retained for subsequent analysis. Initially, only the features selected by the Mann-Whitney U test (p<0.05) are retained. Then, the correlation analysis is performed to exclude features with a correlation coefficient greater than 0.9. The feature dimension is reduced by the maximum correlation minimum redundancy (mRMR) method and the least absolute shrinkage and selection operator (LASSO) method. The regularization parameter λ is selected by ten-fold cross-validation. The optimal imaging feature is selected by support vector machine (SVM) analysis and an imaging feature signature is constructed.

[0079] The best classifier is selected by using six machine learning algorithms (LR, SVM, RF, ExtraTrees, LightGBM and MLP). The performance of the machine learning model constructed on the optimal imaging feature is compared by using various machine learning algorithms including LR, SVM, RF, ExtraTrees, LightGBM and MLP, and the best classifier is obtained: a classification model for predicting early recurrence of pancreatic cancer based on LR and intratumoral-peritumoral imaging features.

[0080] The clinical pathology data of the present application consists of basic case characteristics and medical history information, including age, gender, BMI, smoking status and history of diabetes. Tumor markers, including CA 19-9 (carbohydrate antigen 19-9), CA 12-5, carcinoembryonic antigen (CEA) levels, white blood cell count and platelet count, were measured preoperatively, and at least 2 weeks before surgery. Inflammation-related prognostic indicators, including platelet-lymphocyte ratio (PLR), neutrophil-lymphocyte ratio (NLR), lymphocyte-monocyte ratio (LMR) and prognostic nutritional index (PNI), were calculated. The location and size of the tumor were measured by preoperative CT scanning, and the presence of pancreatitis was confirmed by CT, clinical and laboratory examinations at the time of diagnosis. Pathological information was extracted from postoperative pathological specimens, including margin status (R0 / R1), tumor staging according to the eighth edition of the American Joint Committee on Cancer (AJCC) staging system, tumor differentiation, perineural invasion and lymphatic vessel invasion.

[0081] Segmenting the tumor region of interest and the tumor surrounding region of interest on the CT image using a pre-established neural network model;

[0082] As shown in Figure 10 , the neural network model includes three groups of Unet network structures, the first group of Unet network structures is used to segment the tumor region of interest and the tumor surrounding region of interest in the CT image, the second group of Unet network structures is used to correct the tumor region of interest, and the third group of Unet network structures is used to correct the tumor surrounding region of interest;

[0083] The segmentation formula of the first group of Unet network structures is: CT (maskA, maskB) = Unet1(G CT );

[0084] The segmentation formula of the second group of Unet network structures is:

[0085] maskA_fixed = Add(Unet2(Concat(maskA, G CT )), maskA);

[0086] The segmentation formula of the third group of Unet network structures is:

[0087] maskB_fixed = Add(Unet3(Concat(maskB, G CT )), maskB);

[0088] In the formula, maskA is the tumor region of interest segmented by the first group of Unet network structures, maskB is the tumor surrounding region of interest segmented by the first group of Unet network structures, G CTCT image, Unet1 is a first group of Unet network structure, maskA_fixed is a maskA modified by a second group of Unet network structure, Unet2 is a second group of Unet network structure, maskB_fixed is a maskB modified by a third group of Unet network structure, Unet3 is a third group of Unet network structure, Concat is a channel dimension merging operator, and Add is a feature dimension merging operator;

[0089] The loss function of the neural network model is: ;

[0090] ;

[0091] ;

[0092] ;

[0093] In the formula, is a total loss, is a segmentation loss of a tumor region of interest, is a segmentation loss of a tumor peripheral region of interest, is a reconstruction loss of the tumor region of interest and the tumor peripheral region of interest, is a hyperparameter, is a gold standard of a tumor region of interest in a CT image, is a gold standard of a tumor peripheral region of interest in a CT image, is a gold standard of a tumor region of interest and a tumor peripheral region of interest in a CT image, and MSE is a mean square error operator.

[0094] In the positioning of the region of interest, two parts are separately segmented, the first part is to segment a tumor region of interest (i.e., to segment tumor tissue in a CT image), and the second part is to segment a tumor peripheral region of interest (i.e., to segment tumor peripheral tissue in a CT image).

[0095] The application constructs a neural network model, first, a first group of Unet network structure is set to segment the tumor tissue and the tumor surrounding tissue in the CT image, to obtain the segmentation results of the two parts, the Unet network structure simultaneously segments the two tissue parts, the segmentation efficiency is improved, but at the same time, the segmentation accuracy of the two tissue parts is balanced in the segmentation process, so that the segmentation accuracy of the two tissue parts is easily compromised to each other, the low-precision segmentation occurs at the intersection edge, and the segmentation accuracy of the two tissue parts cannot reach the best, therefore, the application adds accuracy correction in the first part of segmenting the tumor region of interest (i.e. segmenting the tumor tissue in the CT image) and the second part of segmenting the tumor surrounding region of interest (i.e. segmenting the tumor surrounding tissue in the CT image).

[0096] The accuracy correction of segmenting the tumor region of interest is that the tumor tissue segmentation result obtained by the first group of Unet network structure is fused with the original CT image in the channel dimension, so that the tumor tissue segmentation result lost in the segmentation process of the first group of Unet network structure (i.e. in the feature extraction process, some smaller tissue components almost disappear after multiple pooling processing) is supplemented from the original CT image, which is helpful to accurately segment the tissue edge, then the tumor tissue segmentation result with the lost information supplemented is further segmented by the second group of Unet network structure by fusing the channel dimension of the original CT image, the tissue segmentation result lost in the tumor tissue segmentation process is obtained, which is superimposed to the tumor tissue segmentation result obtained by the first group of Unet network structure to obtain the accurate tumor tissue segmentation result, so as to realize the accuracy correction of the tumor tissue segmentation.

[0097] Similarly, the accuracy correction of segmenting the tumor surrounding region of interest is that the tumor surrounding tissue segmentation result obtained by the first group of Unet network structure is fused with the original CT image in the channel dimension, so that the tumor surrounding tissue segmentation result lost in the segmentation process of the first group of Unet network structure (i.e. in the feature extraction process, some smaller tissue components almost disappear after multiple pooling processing) is supplemented from the original CT image, which is helpful to accurately segment the tissue edge, then the tumor surrounding tissue segmentation result with the lost information supplemented is further segmented by the third group of Unet network structure by fusing the channel dimension of the original CT image, the tissue segmentation result lost in the tumor surrounding tissue segmentation process is obtained, which is superimposed to the tumor surrounding tissue segmentation result obtained by the first group of Unet network structure to obtain the accurate tumor surrounding tissue segmentation result, so as to realize the accuracy correction of the tumor surrounding tissue segmentation.

[0098] In the training of the neural network model, the loss function is constructed with the segmentation gold standard as the measurement standard, the loss function includes two parts, the first part is the mean square error between the corrected tumor tissue segmentation result and the tumor tissue gold standard, the mean square error between the corrected tumor peripheral tissue segmentation result and the tumor peripheral tissue gold standard, the second part is the mean square error between the superimposed combination of the corrected tumor tissue segmentation result and the tumor peripheral tissue segmentation result and the combined segmentation gold standard of the tumor tissue and the peripheral tissue.

[0099] The first part can ensure that the corrected tumor tissue segmentation result and the corrected tumor peripheral tissue segmentation result are close to the single tissue segmentation gold standard, and ensure the high precision performance of the corrected tumor tissue segmentation result and the corrected tumor peripheral tissue segmentation result.

[0100] The second part can ensure that the combined result of the corrected tumor tissue segmentation result and the corrected tumor peripheral tissue segmentation result can still restore the multi-tissue segmentation gold standard, avoid the problem of overfitting caused by the high precision performance of the corrected tumor tissue segmentation result and the corrected tumor peripheral tissue segmentation result, and improve the generalization of the model.

[0101] In summary, the pre-established neural network model can obtain high efficiency and high precision of tissue segmentation performance, so as to obtain accurate tumor tissue segmentation and tumor peripheral tissue segmentation, and provide high precision guarantee for subsequent intratumoral and peritumoral image feature.

[0102] Further, the present application provides another embodiment, in particular:

[0103] In order to improve the accuracy of the segmentation of the tumor region of interest and the tumor peripheral region of interest on the CT image, the present application firstly locates the pancreatic tumor region on the CT image, marks the pancreatic tumor region on the CT image by using the region proposal network RPN, and then performs intratumoral-peritumoral region segmentation by using the marked pancreatic tumor region on the CT image, so that the accurate position information of the target can be provided by positioning, which is helpful for more accurately identifying and separating the target during image segmentation, can ensure that the segmentation algorithm is concentrated in the correct area, reduces noise interference, and thus improves the accuracy and efficiency of segmentation.

[0104] Furthermore, in order to improve the accuracy of locating pancreatic tumor regions on CT images using the Region Proposal Network (RPN), this invention sets up two feature extraction networks as teacher models to guide the training of a lightweight MobileNet network. This results in a feature extraction network structure that combines high-precision localization performance with a lightweight structure, reducing the computational load of the feature extraction process while achieving high-precision pancreatic tumor localization performance.

[0105] The methods for constructing feature extraction networks include:

[0106] A global feature extraction network structure is used to extract global features from CT images to obtain global image features.

[0107] A local feature extraction network structure is used to extract local features from CT images to obtain local image features.

[0108] Both the global feature extraction network structure and the local feature extraction network structure are used as teacher models to guide the training of the lightweight MobileNet network, and the trained MobileNet network is used as the feature extraction network.

[0109] The network structure of the feature extraction network is as follows:

[0110] ;

[0111] In the formula, Image features For MobileNet networks, For CT images;

[0112] The structure of the regional recommendation network is as follows:

[0113] ;

[0114] In the formula, Candidate regions for segmentation of pancreatic tumors, such as Figure 9 As shown, Image features A network is recommended for the region.

[0115] Global feature extraction methods for global feature extraction network structures include:

[0116] The CT images are sequentially passed through a 3×3 convolutional layer, a global average pooling layer, a 1×1 convolutional layer, a normalization layer, a ReLU activation function, a 1×1 convolutional layer, and a normalization layer to obtain a global feature extraction method.

[0117] The structural formula of the global feature extraction network is:

[0118] ;

[0119] In the formula, The global feature extraction method is, The CT image is, The 3*3 convolution layer is, The 1*1 convolution layer is, GAP is a global average pooling, and BN is a normalization layer, The ReLU activation function is.

[0120] In the present application, the global feature extraction network structure is subjected to a global average pooling operation (GAP) to extract global context information, and then is subjected to a 1*1 convolution layer to reduce the calculation amount, and then is subjected to a batch normalization layer (BN) and a ReLU activation function to improve the nonlinear expression ability of the network, and finally is subjected to a 1*1 convolution to restore the original channel number, and is subjected to a BN layer to prevent overfitting, to obtain global features.

[0121] The local feature extraction method of the local feature extraction network structure includes:

[0122] The CT image is sequentially subjected to a 3*3 convolution layer, a 1*1 convolution layer, a normalization layer, a ReLU activation function, a 1*1 convolution layer and a normalization layer, to obtain a local feature extraction method;

[0123] The structure formula of the local feature extraction network structure is:

[0124] ;

[0125] In the formula, The local feature extraction method is, The CT image is, The 3*3 convolution layer is, The 1*1 convolution is, and BN is a normalization layer, The ReLU activation function is.

[0126] In the present application, the local feature extraction network structure is subjected to a 1*1 convolution layer to reduce the calculation amount, and then is subjected to a batch normalization layer (BN) and a ReLU activation function which can improve the nonlinear expression ability of the network, and finally is subjected to a 1*1 convolution to restore the original channel number, and is subjected to a BN layer to prevent overfitting, to obtain local features.

[0127] The method for guiding the training of a light MobileNet network by taking the global feature extraction network structure and the local feature extraction network structure as teacher models includes:

[0128] The global feature extraction network structure is taken as a first teacher model, a light MobileNet network is guided to train, and the trained MobileNet network is taken as a first training model;

[0129] The local feature extraction network structure is taken as a second teacher model, the first training model is guided to train, and the trained first training model is taken as a feature extraction network;

[0130] The loss function for guiding the MobileNet network to train is:

[0131]

[0132] The loss function for guiding the first training model to train is:

[0133]

[0134] In the formula, is the loss for guiding the MobileNet network to train, is the loss for guiding the first training model to train, is a pancreatic tumor segmentation candidate region obtained by a region proposal network according to global features output by the global feature extraction network structure, is a pancreatic tumor segmentation candidate region obtained by the region proposal network according to image features output by the MobileNet network, is a pancreatic tumor segmentation candidate region obtained by the region proposal network according to local features output by the local feature extraction network structure, is a pancreatic tumor segmentation candidate region obtained by the region proposal network according to image features output by the first training model, is a true value of the pancreatic tumor segmentation candidate region, and is obtained by an experienced doctor from 3Dslicer software, , are all hyperparameters, and are usually set to 1, is a 2-norm formula.

[0135] The global feature extraction network structure guides the light MobileNet network to train, and the trained MobileNet network is taken as the first training model, as a loss function

[0136] The image features extracted by the trained MobileNet network can obtain a similar pancreatic tumor positioning result through the global features, that is, satisfies is minimized, and in the teacher model guided training process (knowledge distillation), the image features extracted by the trained MobileNet network also need to obtain a similar pancreatic tumor positioning result to the real tumor positioning, that is, satisfies is minimized.​​

[0137] The global feature is used for pancreatic tumor region positioning, and generally has an advantage in large target detection, but the precision is limited in small target detection. The image feature extracted by the trained MobileNet network obtains a similar pancreatic tumor positioning result as the global feature, which enables the MobileNet network to learn the feature extraction process of detecting large targets.

[0138] The application guides the training of the lightweight MobileNet network through the local feature extraction network structure, and sets the loss function as

[0139] The image feature extracted by the trained MobileNet network obtains a similar pancreatic tumor positioning result as the local feature, that is, satisfies minimization, and in the teacher model guided training process (knowledge distillation), the image feature extracted by the trained MobileNet network also needs to obtain a pancreatic tumor positioning result similar to the real tumor positioning, that is, satisfies minimization.

[0140] The local feature is used for pancreatic tumor region positioning, and generally has an advantage in small target detection. The image feature extracted by the trained MobileNet network obtains a similar pancreatic tumor positioning result as the local feature, which enables the MobileNet network to learn the feature extraction process of detecting small targets.

[0141] The loss function for training the global feature extraction network structure and the local feature extraction network structure includes:

[0142]

[0143] In the formula, is the loss value of training the global feature extraction network structure and the local feature extraction network structure, is the pancreatic tumor segmentation candidate region obtained by the region proposal network according to the global feature output by the global feature extraction network structure, is the pancreatic tumor segmentation candidate region obtained by the region proposal network according to the local feature output by the local feature extraction network structure, is the KL divergence operation formula.

[0144] The application sets the loss function of the global feature extraction network structure and the local feature extraction network structure as , wherein ​​The global feature extraction network structure learns from the local feature extraction network structure, i.e., the tumor region positioning result obtained by the global feature extracted by the global feature extraction network structure is fitted to the tumor region positioning result obtained by the local feature extracted by the local feature extraction network structure. The local feature extraction network structure learns from the global feature extraction network structure, i.e., the tumor region positioning result obtained by the local feature extracted by the local feature extraction network structure is fitted to the tumor region positioning result obtained by the global feature extracted by the global feature extraction network structure. Therefore, the tumor region positioning result obtained by the global feature extraction network structure is constrained by the tumor region positioning result extracted by the local feature extraction network structure, and the tumor region positioning result obtained by the local feature extraction network structure is constrained by the tumor region positioning result extracted by the global feature extraction network structure.

[0145] Correspondingly, when the global feature extraction network guides the MobileNet network training, when the MobileNet network is taught the feature extraction skill for large target detection, it is constrained by the small target detection feature skill of the local feature extraction network structure, so as to avoid the trained MobileNet network extracting features only for large target detection, i.e., the features extracted by the trained MobileNet network are over-fitted to the features for large target detection, resulting in ineffective feature extraction. When the local feature extraction network guides the MobileNet network training, when the MobileNet network is taught the feature extraction skill for small target detection, it is constrained by the large target detection feature skill of the global feature extraction network structure, so as to avoid the trained MobileNet network extracting features only for small target detection, i.e., the features extracted by the trained MobileNet network are over-fitted to the features for small target detection, resulting in ineffective feature extraction. Thus, in the training process guided by the two teacher models, the mutual constraint of the feature extraction of the two teacher models is balanced, so that the MobileNet network can extract image features that are superior to both large target and small target detection, and there is no over-fitting defect.

[0146] Correspondingly, the neural network model is modified in structure based on the feature extraction network and the region proposal network RPN:

[0147] The segmentation formula of the first group of Unet network structures is: (maskA, maskB) = Unet1(G CT-Box );

[0148] The segmentation formula of the second group of Unet network structures is:

[0149] maskA_fixed = Add(Unet2(Concat(maskA, G CT-Box )), maskA);

[0150] The segmentation of the third group of Unet network structure is as follows:

[0151] maskB_fixed = Add(Unet3(Concat(maskB, G CT-Box )), maskB);

[0152] In the formula, maskA is the tumor region of interest segmented by the first group of Unet network structure, maskB is the tumor peripheral region of interest segmented by the first group of Unet network structure, G CT-Box is the CT image in the pancreatic tumor segmentation candidate region, Unet1 is the first group of Unet network structure, maskA_fixed is the maskA corrected by the second group of Unet network structure, Unet2 is the second group of Unet network structure, maskB_fixed is the maskB corrected by the third group of Unet network structure, Unet3 is the third group of Unet network structure, Concat is a channel dimension merging operator, and Add is a feature dimension merging operator.

[0153] The loss function is as follows: ;

[0154] ;

[0155] ;

[0156] ;

[0157] In the formula, is the total loss, is the segmentation loss of the tumor region of interest, is the segmentation loss of the tumor peripheral region of interest, is the reconstruction loss of the tumor region of interest and the tumor peripheral region of interest, is a hyperparameter, usually set to 1, is the gold standard for the tumor region of interest in the CT image in the pancreatic tumor segmentation candidate region, is the gold standard for the tumor peripheral region of interest in the CT image in the pancreatic tumor segmentation candidate region, is the pancreatic tumor segmentation candidate region, and MSE is a mean square error operator.

[0158] The method for determining the intratumoral and extratumoral imageomic features comprises:

[0159] segmenting the tumor region of interest and the tumor peripheral region of interest on the CT image using a pre-established neural network model;

[0160] The neural network model comprises three groups of Unet network structures, the first group of Unet network structures is used for segmenting a tumor region of interest and a tumor peripheral region of interest in the CT image, the second group of Unet network structures is used for correcting the tumor region of interest, and the third group of Unet network structures is used for correcting the tumor peripheral region of interest.

[0161] The segmentation formula of the first group of Unet network structures is (maskA, maskB) = Unet1(G CT-Box );

[0162] The segmentation formula of the second group of Unet network structures is:

[0163] maskA_fixed = Add(Unet2(Concat(maskA, G CT-Box )), maskA);

[0164] The segmentation formula of the third group of Unet network structures is:

[0165] maskB_fixed = Add(Unet3(Concat(maskB, G CT-Box )), maskB);

[0166] In the formula, maskA is the tumor region of interest segmented by the first group of Unet network structures, maskB is the tumor peripheral region of interest segmented by the first group of Unet network structures, G CT-Box is the CT image in the pancreatic tumor segmentation candidate region, Unet1 is the first group of Unet network structures, maskA_fixed is the maskA corrected by the second group of Unet network structures, Unet2 is the second group of Unet network structures, maskB_fixed is the maskB corrected by the third group of Unet network structures, Unet3 is the third group of Unet network structures, Concat is a channel dimension merging operator, and Add is a feature dimension merging operator.

[0167] The loss function of the neural network model is: ;

[0168] ;

[0169] ;

[0170] ;

[0171] In the formula, is the total loss, is the segmentation loss of the tumor region of interest, Segmentation loss of the region of interest around the tumor, Reconstruction loss of the region of interest around the tumor and the region of interest around the tumor, Hyperparameter, usually set to 1, Gold standard for separate segmentation of the region of interest around the tumor in the CT image in the pancreatic tumor segmentation candidate region, Gold standard for separate segmentation of the region of interest around the tumor in the CT image in the pancreatic tumor segmentation candidate region, Pancreatic tumor segmentation candidate region, MSE is the mean square error operator.

[0172] Resample the CT image to a standardized pixel size of 1.0 × 1.0 × 1.0 mm 3 Discretize the gray level into a fixed 5 interval, and use Pyradiomics (Python image feature extraction tool) to extract the image features of the region of interest around the tumor and the region of interest around the tumor as the intratumoral and extratumoral image features, wherein the 1106 intratumoral and extratumoral image features include 216 first-order features, 264 gray level co-occurrence matrix GLCM features, 168 gray level dependence matrix GLDM features, 192 gray level run length matrix GLRLM features, 192 gray level size zone matrix GLSZM features, 60 adjacent gray level difference matrix NGTDM features and 14 shape features.

[0173] The quantitative analysis method of body composition parameters comprises:

[0174] Measure the cross-sectional image of the third lumbar vertebra level of the CT image by SliceOMatic V5.0 to obtain the skeletal muscle index SMI, subcutaneous fat index SFI, intermuscular fat index IMFI, visceral fat index VFI, and skeletal muscle density SMD;

[0175] The subcutaneous fat- visceral fat ratio VSR is calculated by dividing the subcutaneous fat index SFI by the visceral fat index VFI;

[0176] The visceral fat-skeletal muscle tissue ratio VMR is calculated by dividing the skeletal muscle index SMI by the visceral fat index VFI.

[0177] The determination method of clinical characteristics comprises:

[0178] Statistically significant clinical factors p<0.05 are screened out from the body composition parameters and clinical pathological data by single factor analysis;

[0179] Clinical factors were input into a multivariable logistic regression analysis to select independent predictors for identifying the risk of early recurrence of pancreatic cancer, as clinical features, the clinical features included serum CA12-5, visceral fat index VFI, adjuvant therapy, lymphatic vessel invasion LVI, tumor grade and TNM stage.

[0180] The training set and the test set included 311 (mean age 60.4 years ± standard deviation [SD] 8.6; 200 males) and 104 (mean age 60.7 years ± SD 8.3; 59 males) patients, respectively. During follow-up, 49.2% (153 / 311) and 38.46% (40 / 104) of patients in the training set and the test set, respectively, had early recurrence (P = 0.072).

[0181] Univariate analysis showed that preoperative clinical information (including CA12-5, VFI) and preoperative data (including no adjuvant therapy, lymphatic vessel invasion LVI, tumor grade and TNM stage) were significantly associated with early recurrence of PDAC (pancreatic ductal adenocarcinoma) (all p<0.05). Multivariate analysis found that CA12-5, VFI, adjuvant therapy, lymphatic vessel invasion LVI and TNM stage were independent risk factors for early recurrence (all p<0.05), which were used to establish a classification model of clinical features.

[0182] The method for obtaining postoperative follow-up data includes:

[0183] According to the follow-up plan, follow-up is carried out once every 3 months in the first two years, once every 6 months from 3 to 4 years, and once a year thereafter, wherein the follow-up plan includes physical examination, serum tumor marker detection and abdominal and pelvic CT scan;

[0184] According to the reference time of recurrence of pancreatic ductal adenocarcinoma after surgery, cases that recur within 1 year after surgery are marked as the ER group, and cases that do not recur or recur more than 1 year after surgery are marked as the non-ER group, so the case labels are "ER" and "non-ER".

[0185] The method for determining the radiomics model includes:

[0186] By the correlation coefficient ICC, 978 radiomics features with ICC≥0.75 are retained from 1106 intratumoral and peritumoral radiomics features;

[0187] Among the radiomics features with ICC≥0.75, 7 intratumoral radiomics features, 9 peritumoral radiomics features and 8 intratumoral-peritumoral radiomics features are obtained by maximum relevance minimum redundancy mRMR method and least absolute shrinkage and selection operator LASSO method analysis;

[0188] Based on the 7 intratumoral imaging features and the case labels in the postoperative follow-up data, six classification models for predicting early recurrence of pancreatic cancer based on intratumoral imaging features were constructed by LR, SVM, RF, ExtraTrees, LightGBM and MLP algorithms: Label(ER, NER)=K(featureA), K∈[LR, SVM, RF, ExtraTrees, LightGBM, MLP];

[0189] Based on the 9 peritumoral imaging features and the case labels in the postoperative follow-up data, six classification models for predicting early recurrence of pancreatic cancer based on peritumoral imaging features were constructed by LR, SVM, RF, ExtraTrees, LightGBM and MLP algorithms: Label(ER, NER)=K(featureB), K∈[LR, SVM, RF, ExtraTrees, LightGBM, MLP];

[0190] Based on the 8 intratumoral-peritumoral imaging features and the case labels in the postoperative follow-up data, six classification models for predicting early recurrence of pancreatic cancer based on intratumoral-peritumoral imaging features were constructed by LR, SVM, RF, ExtraTrees, LightGBM and MLP algorithms: Label(ER, NER)=K(featureAB), K∈[LR, SVM, RF, ExtraTrees, LightGBM, MLP];

[0191] In the formula, Label(ER, NER) is the case label, featureA is the intratumoral imaging feature, featureB is the peritumoral imaging feature, featureAB is the intratumoral-peritumoral imaging feature, and K is the identifier of the model body, which can be selected from LR, SVM, RF, ExtraTrees, LightGBM and MLP algorithms.

[0192] Based on the AUC index, ACC index, SEN index, SPE index, PPV index and NPV index, the performance differences between the three types of classifiers, i.e., the classification model for predicting early recurrence of pancreatic cancer based on intratumoral imaging features, the classification model for predicting early recurrence of pancreatic cancer based on peritumoral imaging features and the classification model for predicting early recurrence of pancreatic cancer based on intratumoral-peritumoral imaging features, were evaluated, and it was found that the classification model for predicting early recurrence of pancreatic cancer based on intratumoral-peritumoral imaging features was superior to the classification model for predicting early recurrence of pancreatic cancer based on intratumoral imaging features and the classification model for predicting early recurrence of pancreatic cancer based on peritumoral imaging features;

[0193] Based on the AUC index, ACC index, SEN index, SPE index, PPV index, NPV index, the performance difference between the six intratumoral-peritumoral image-based classification models for predicting early recurrence of pancreatic cancer was evaluated, and the classification model based on LR and intratumoral-peritumoral image-based features for predicting early recurrence of pancreatic cancer was superior to the classification model based on LR, SVM, ExtraTrees, LightGBM, MLP and intratumoral-peritumoral image-based features for predicting early recurrence of pancreatic cancer.

[0194] The classification model based on LR and intratumoral-peritumoral image-based features for predicting early recurrence of pancreatic cancer is used as an image-based model.

[0195] The present application evaluates the performance of intratumoral, peritumoral and intratumoral-peritumoral image-based classification models under six machine learning algorithms in the training set and test set. As shown in Figure 3 and Figure 4 The classification model based on intratumoral-peritumoral image-based features is superior to the classification model based on intratumoral and peritumoral image-based features in both sets, wherein Intra is the classification model based on intratumoral image-based features, Peri is the classification model based on peritumoral image-based features, and IntraPeri is the classification model based on intratumoral-peritumoral image-based features. Among the six algorithms, the RF model performs best in intratumoral-peritumoral image-based features, with AUC values of 0.830 (95% CI: 0.785-0.874) and 0.770 (95% CI: 0.675-0.865) in the training set and test set, respectively. In addition, its accuracy (ACC) in the training set and test set is 0.759 and 0.721, respectively, the sensitivity is 0.889 and 0.750, respectively, the specificity is 0.663 and 0.703, respectively, the positive predictive value (PPV) is 0.701 and 0.612, respectively, and the negative predictive value (NPV) is 0.855 and 0.818, respectively (as shown in Table 1), wherein AUC is the area under the receiver operating characteristic curve; ACC is the accuracy; SEN is the sensitivity; SPE is the specificity; PPV is the positive predictive value; NPV is the negative predictive value; LR is the logistic regression; SVM is the support vector machine; RF is the random forest; ExtraTrees is the extreme random tree; LightGBM is the light gradient boosting machine; MLP is the multilayer perceptron. Given its excellent performance, the classification model based on RF algorithm and intratumoral-peritumoral image-based features is selected for the construction of the subsequent combined model.

[0196]

[0197] Therefore, the application proposes and verifies a machine learning prediction model (classification model based on RF algorithm and intra-peritumoral radiomics features) integrating intra-tumoral and peritumoral radiomics features for evaluating the early recurrence risk after radical resection of PDAC. The results show that the classification model combined with intra-peritumoral radiomics features outperforms the classification model of single intra-tumoral or peritumoral features. This is due to the fact that tumor heterogeneity not only exists within the tumor body, but is also closely related to the surrounding microenvironment. Although traditional imaging techniques are difficult to directly capture the tumor microenvironment, radiomics can identify subtle changes by quantitative analysis, providing important supplementary information for tumor heterogeneity. Therefore, the classification model based on RF algorithm and intra-peritumoral radiomics features proposed by the application conforms to the actual prediction scenario, provides important supplementary information through tumor heterogeneity, and exhibits the best prediction ability, with AUC indicators of 0.830 and 0.770 for the training set and the test set, respectively.

[0198] In order to further improve the prediction ability, the application incorporates independent clinical risk factors (including serum CA12-5 (carcinoembryonic antigen 125), VFI (visceral fat index), adjuvant therapy, LVI (lymphatic and vascular invasion), tumor grade, and TNM stage (tumor stage)) into the RF algorithm to construct a clinical feature classification model. The AUC values of this model in the training set and the test set are 0.763 (95% CI: 0.711-0.816) and 0.717 (95% CI: 0.615-0.819), respectively, as shown in Table 2, where Clinical is the clinical feature classification model, IntraPeri is the radiomics model, and Combined is the clinical-radiomics combined model. The combined model (clinical-radiomics combined model) based on the RF algorithm combining intra-peritumoral radiomics features with independent clinical factors performs better, with AUC values of 0.890 (95% CI: 0.854-0.925) and 0.822 (95% CI: 0.742-0.901) for the training set and the test set, respectively, as shown in Figure 5 Figure 5 wherein A-C are the ROC curve, calibration curve, and decision curve of the three models in the training set, and D-F are the ROC curve, calibration curve, and decision curve of the three models in the test set). This combined model outperforms the classification model based on RF and intra-peritumoral radiomics features, as well as the classification model based on RF and clinical features in terms of accuracy and decision utility, and the ROC curve, calibration curve, and decision curve analysis results are shown in Figure 5

[0199]

[0200] ​​The area under the receiver operating characteristic curve (AUC) of the training set and the test set was calculated by the receiver operating characteristic curve (ROC) when the present application evaluated the performance of all models. The goodness of fit of the image-omics-clinical model was evaluated by the calibration curve. The clinical utility was evaluated by the decision curve analysis (DCA), covering two groups of data sets.

[0201] The determination method of the clinical-image-omics combined model comprises:

[0202] The serum CA12-5, visceral fat index VFI, adjuvant therapy, lymphatic vessel invasion LVI, tumor grade and TNM stage are combined into the image-omics model to obtain the clinical-image-omics combined model.

[0203] The performance of the clinical-image-omics combined model is evaluated by the ROC index, the AUC index and the DCA index.

[0204] The present application further improves the prediction performance of the model by integrating preoperative tumor image-omics features, body composition and postoperative pathological data, and exhibits better overall performance than the above research in the training set and the test set. The clinical-image-omics combined model constructed by the present application not only focuses on the image features of the tumor itself, but also considers the influence of individual differences, especially the body composition and postoperative pathology on the early recurrence of PDAC. Various machine learning algorithms are used to deeply mine the complex relationship between features and outcomes, and to improve the accuracy and flexibility of the prediction model.

[0205] In addition to radiomics, body composition reflects the nutritional metabolic state of patients, which has been proven to be an important factor affecting tumor recurrence and survival outcomes. Previous studies have shown that increased visceral fat is associated with early recurrence and poor prognosis of various tumors, including PDAC. This invention further supports this finding by showing that patients with high visceral fat are more likely to relapse within 1 year after surgery. This is due to the accumulation of visceral fat, which is associated with pancreatic fat infiltration, and the accumulated fat promotes tumor dissemination by altering the tumor microenvironment. In addition, lymphovascular invasion (LVI), TNM staging, and adjuvant therapy have also been proven to be important factors affecting PDAC recurrence, providing theoretical support for their inclusion in the prediction model. It is worth mentioning that, although many studies have used CA199 as a marker for early recurrence of PDAC, this invention found that CA125 is closely related to early recurrence of PDAC and included it in the machine learning model. Unlike CA199, which reflects the overall tumor burden of PDAC, CA125 is more focused on "metastasis-related burden." Preoperative elevation of CA125 may indicate the presence of "subclinical metastatic lesions" that have not been detected by conventional imaging. These lesions may significantly increase the risk of postoperative recurrence, which explains why some patients relapse within 1 year despite receiving radical resection. Although inflammatory prognostic scores such as LMR, NLR, and PLR were also evaluated, these indicators did not show significant correlation in early recurrence of PDAC, and therefore were not included in the model. Finally, the combined model integrating radiomics features, VFI, TNM staging, lymphovascular invasion (LVI), and CA125 achieved AUC of 0.890 and 0.822 in the training set and test set, respectively, further confirming the importance of integrating multiple data types for prognosis prediction.

[0206] The method for performing explainability SHAP analysis on the clinical-radiomics combined model comprises:

[0207] Global explanation of the clinical-radiomics combined model at the feature level is obtained through SHAP visual analysis, and the contribution of intratumoral-peritumoral radiomics features and clinical features to the output of the clinical-radiomics combined model is obtained, wherein the intratumoral-peritumoral radiomics features have the largest contribution, followed by clinical features, and high radiomics feature values are closely related to increased risk of early recurrence. High visceral fat index (VFI), CA12-5 ≥ 35U / ml, positive lymphovascular invasion (LVI), TNM stage ≥IIA, and no adjuvant therapy are all associated with increased risk of early recurrence, and intratumoral-peritumoral radiomics features and VFI have an interaction that increases the risk of early recurrence.

[0208] The SHAP visual analysis is used to perform individual-level local interpretation on the clinical-radiomics combined model, and the cases of no early recurrence of pancreatic ductal adenocarcinoma (PDAC) and the cases of early recurrence of PDAC during follow-up are obtained, and the contribution of intratumoral-peritumoral radiomics features and clinical features to the output of the clinical-radiomics combined model is matched with the decision logic.

[0209] The SHAP method is used to enhance the model interpretability, and more intuitive insights are provided for the contribution of each feature to the prediction and the individualized decision-making process.

[0210] The SHAP method is used to explain the output of the model by calculating the contribution of each variable to the prediction. This interpretable method provides two aspects: global explanation at the feature level and local explanation at the individual level. In global explanation, for the clinical-radiomics combined model, the SHAP bar chart (as shown in FIG. Figure 6 A) shows that the radiomics features have the largest contribution to the output of the combined model, followed by clinical factors such as VFI, lymphatic vessel invasion (LVI), TNM stage, adjuvant therapy, and CA12-5. The violin plot (as shown in FIG. Figure 6 B) and the dependence plot (as shown in FIG. Figure 8 A) show that as the radiomics feature value increases, the SHAP value shows a clear upward trend, indicating that a higher radiomics feature value is closely related to an increased risk of early recurrence. Similarly, high VFI, CA12-5 ≥ 35 U / ml, LVI positive, TNM stage ≥ IIA, and no adjuvant therapy are all associated with an increased risk of early recurrence (as shown in FIG. Figure 8 B-F). In addition, the interaction between radiomics and VFI (as shown by the color gradient in FIG. Figure 6 C-D) indicates that patients with higher radiomics feature values and increased VFI have a higher risk of recurrence.

[0211] In local explanation, Figure 7 FIGS. Figure 7 A and B show a case of a patient who did not experience early recurrence of PDAC during follow-up. According to the prediction model, the probability of this patient being classified as the "ER" class is 17.5%, which is significantly lower than the baseline probability of 52.5%. The specific contribution of each feature is also shown in the waterfall Figure 7 A and B. As shown in FIGS. Figure 7 A and B, the radiomics feature value, LVI negative, TNM ≤ IIA, adjuvant therapy, and CA12-5 lower than 35 U / ml push the decision towards the "non-ER" class, but high VFI does not have this effect. Figure 7 FIGS. C and D show another case of a patient who experienced early recurrence. According to the prediction model, the probability of this patient being classified as the "ER" class is 81.8%.

[0212] The explainability of machine learning model is the key to its clinical translation application. The SHAP value is used to analyze the contribution of each feature to the model prediction. The SHAP analysis shows that the intratumoral-peritumoral radiomics features contribute the most to the prediction of early recurrence, followed by VFI, lymphatic vessel invasion LVI, TNM stage, adjuvant therapy and CA125 level. In addition, the local explainability of the clinical-radiomics combined model highlights its potential for individualized treatment strategy. By understanding the specific contribution of each factor, clinicians can more accurately assess the individual recurrence risk and make postoperative monitoring and adjuvant therapy decisions.

[0213] Figures 6-8 Radiomics refers to radiomics features.

[0214] As shown in Figure 2 The application provides an early recurrence prediction system for pancreatic cancer based on an explainable machine model, which is applied to an early recurrence prediction method for pancreatic cancer based on an explainable machine model. The system comprises:

[0215] A data acquisition unit is configured to acquire CT image data, clinical pathology data and postoperative follow-up data of a case;

[0216] A feature processing unit is configured to segment intratumoral and peritumoral regions in the CT image, determine VOIs (Volumetric of Interest), and extract intratumoral and peritumoral radiomics features in the VOIs; perform quantitative analysis on body composition in the CT image to obtain body composition parameters, and perform single factor analysis and multivariate logistic regression analysis on the body composition parameters and the clinical pathology data to obtain clinical features;

[0217] A model construction unit is configured to construct six groups of classifier models based on intratumoral and peritumoral radiomics features and postoperative follow-up data by using six machine learning algorithms, and obtain a radiomics model according to performance comparison of the six groups of classifier models; construct a clinical-radiomics combined model for early recurrence prediction of pancreatic cancer by combining the clinical features and the radiomics model, and perform SHAP analysis on the explainability of the clinical-radiomics combined model.

[0218] The application combines intratumoral and peritumoral CT radiomics features with body composition parameters to construct a machine learning model for predicting the early recurrence risk of PDAC after resection, and incorporates SHAP analysis to enhance the transparency of the decision-making process of the machine learning model, improve the credibility and explainability of the machine learning model, increase the trust of clinicians in the prediction results, and help to develop individualized treatment strategies.

[0219] The above examples are only exemplary embodiments of the present application, and are not intended to limit the present application, and the protection scope of the present application is defined by the claims. Those skilled in the art can make various modifications or equivalent replacements to the present application within the spirit and protection scope of the present application, and such modifications or equivalent replacements are also considered to fall within the protection scope of the present application.

Claims

1. A method of predicting early recurrence of pancreatic cancer based on an interpretable machine model, the method comprising: The method comprises the following steps: Obtaining CT image data, clinical pathology data and postoperative follow-up data of a case; Segmenting intra-tumor and peritumor regions in the CT image, determining VOIs, and extracting intra- and extra-tumor image features in the VOIs; Quantitatively analyzing body composition in the CT image to obtain body composition parameters, and performing single factor analysis and multivariate logistic regression analysis on the body composition parameters and the clinical pathology data to obtain clinical features; Constructing six classifier models based on the intra- and extra-tumor image features and the postoperative follow-up data by using six machine learning algorithms, and obtaining an image feature model according to the performance comparison of the six classifier models; Constructing a clinical-image feature combined model for predicting early recurrence of pancreatic cancer by combining the clinical features and the image feature model, and performing SHAP analysis on the clinical-image feature combined model; The method for segmenting intra-tumor and peritumor regions in the CT image comprises: Marking pancreatic tumor segmentation candidate regions in the CT image by using a feature extraction network and a region proposal network; Segmenting tumor VOIs and peritumor VOIs in the pancreatic tumor segmentation candidate regions in the CT image by using a pre-established neural network model; The method for constructing the feature extraction network comprises: Extracting global features of the CT image by using a global feature extraction network structure to obtain image global features; Extracting local features of the CT image by using a local feature extraction network structure to obtain image local features; Using the global feature extraction network structure and the local feature extraction network structure as teacher models to guide the training of a lightweight MobileNet network, and using the trained MobileNet network as the feature extraction network; The network structure of the feature extraction network is as follows: ； In the formula, is an image feature, is a MobileNet network, is a CT image; The structure of the region proposal network is as follows: ； In the formula, segmenting a pancreatic tumor from the candidate region, image features, a region proposal network; The global feature extraction method of the global feature extraction network structure comprises: The CT image is sequentially passed through a 3×3 convolution layer, a global average pooling layer, a 1×1 convolution layer, a normalization layer, a ReLU activation function, a 1×1 convolution layer and a normalization layer to obtain the global feature extraction method; The structure formula of the global feature extraction network structure is as follows: ； In the formula, is a global feature extraction method, is a CT image, is a 3x3 convolution layer, is a 1x1 convolution layer, GAP is a global average pooling, and BN is a normalization layer, is a ReLU activation function. The local feature extraction method of the local feature extraction network structure comprises: The CT image is sequentially passed through a 3×3 convolution layer, a 1×1 convolution layer, a normalization layer, a ReLU activation function, a 1×1 convolution layer and a normalization layer to obtain the local feature extraction method; The structure formula of the local feature extraction network structure is as follows: ； In the formula, is a local feature extraction method, is a CT image, is a 3x3 convolution layer, is a 1x1 convolution, BN is a normalization layer, is a ReLU activation function; The method for guiding the training of the lightweight MobileNet network by using the global feature extraction network structure and the local feature extraction network structure as teacher models comprises: Using the global feature extraction network structure as a first teacher model to guide the training of the lightweight MobileNet network, and using the trained MobileNet network as a first training model; Using the local feature extraction network structure as a second teacher model to guide the training of the first training model, and using the trained first training model as the feature extraction network; The loss function for guiding the training of the MobileNet network is as follows: The loss function for guiding the training of the first trained model is: In the formula, is a loss function for guiding the MobileNet network training, is a loss function for guiding the first training model training, is a pancreatic tumor segmentation candidate region obtained by the region proposal network according to the global feature output by the global feature extraction network structure, is a pancreatic tumor segmentation candidate region obtained by the region proposal network according to the image feature output by the MobileNet network, is a pancreatic tumor segmentation candidate region obtained by the region proposal network according to the local feature output by the local feature extraction network structure, is a pancreatic tumor segmentation candidate region obtained by the region proposal network according to the image feature output by the first training model, is a true value of the pancreatic tumor segmentation candidate region, , are all hyperparameters, and are usually set to 1, is a 2-norm formula; The structure of the neural network model for segmenting the tumor region of interest and the tumor peripheral region of interest in the pancreatic tumor segmentation candidate region in the CT image includes: The segmentation of the first group of Unet network structures is: (maskA, maskB) = Unet1(G CT-Box ); The segmentation formula of the second group of Unet network structures is: maskA_fixed = Add(Unet2(Concat(maskA, G CT-Box )), maskA) The segmentation formula of the third group of Unet network structures is: maskB_fixed = Add(Unet3(Concat(maskB, G CT-Box )), maskB) In the formula, maskA is the tumor region of interest segmented by the first Unet network structure, maskB is the tumor peripheral region of interest segmented by the first Unet network structure, G CT-Box is the CT image in the pancreatic tumor segmentation candidate region, Unet1 is the first Unet network structure, maskA_fixed is the maskA corrected by the second Unet network structure, Unet2 is the second Unet network structure, maskB_fixed is the maskB corrected by the third Unet network structure, Unet3 is the third Unet network structure, Concat is a channel dimension merging operator, and Add is a feature dimension merging operator. The loss function is: ; ； ； ； wherein, is a total loss, is a segmentation loss of a tumor region of interest, is a segmentation loss of a tumor region of interest, is a reconstruction loss of a tumor region of interest and a tumor region of interest, is a hyper parameter, usually set to 1, is a gold standard of a tumor region of interest in a CT image within a pancreatic tumor segmentation candidate region, is a gold standard of a tumor region of interest in a CT image within a pancreatic tumor segmentation candidate region, is a pancreatic tumor segmentation candidate region, and MSE is a mean square error operator. 2.The method of claim 1, wherein the method is characterized by: The method for determining the intratumoral and extratumoral imageomic features includes: The CT images were resampled to a standardized pixel size of 1.0 x 1.0 x 1.0 mm 3 The gray level was discretized into a fixed 5 intervals, and the image features of the tumor region of interest and the tumor peripheral region of interest were extracted using Pyradiomics as the intratumoral and extratumoral image features, wherein the intratumoral and extratumoral image features included 216 first-order features, 264 gray level co-occurrence matrix (GLCM) features, 168 gray level dependence matrix (GLDM) features, 192 gray level run-length matrix (GLRLM) features, 192 gray level size zone matrix (GLSZM) features, 60 neighboring gray tone difference matrix (NGTDM) features, and 14 shape features. 3.The method of claim 2, wherein the method is characterized by: The method for quantitatively analyzing the body composition parameters includes: The skeletal muscle index SMI, the subcutaneous fat index SFI, the intermuscular fat index IMFI, the visceral fat index VFI, and the skeletal muscle density SMD are obtained by measuring the cross-sectional image of the third lumbar vertebra level of the CT image through SliceOMatic V5.0; The subcutaneous fat- visceral fat ratio VSR is calculated by dividing the subcutaneous fat index SFI by the visceral fat index VFI; The visceral fat-skeletal muscle tissue ratio VMR is calculated by dividing the skeletal muscle index SMI by the visceral fat index VFI.

4. The method of claim 3, wherein the method is based on an interpretable machine model. The method for determining the clinical characteristics includes: Clinical factors with statistical significance p < 0.05 are screened out from the body composition parameters and the clinical pathological data through single factor analysis; The clinical factors are input into multivariate logistic regression analysis to screen out independent predictors for identifying the early recurrence risk of pancreatic cancer, which are the clinical characteristics, including serum CA12-5, visceral fat index VFI, adjuvant therapy, lymphatic vessel invasion LVI, tumor grade, and TNM stage.

5. The method of claim 4, wherein the method is based on an interpretable machine model. The method for obtaining the postoperative follow-up data includes: According to the follow-up scheme, follow-up is performed once every 3 months in the first two years, once every 6 months from 3 to 4 years, and once a year thereafter, wherein the follow-up scheme includes physical examination, serum tumor marker detection, and abdominal and pelvic CT scanning; According to the reference time of pancreatic ductal adenocarcinoma recurrence after surgery, cases that recur within 1 year after surgery are marked as the ER group, and cases that do not recur or recur more than 1 year after surgery are marked as the non-ER group.

6. The method of claim 5, wherein the method is based on an interpretable machine model. The method for determining the imageomic model includes: Imageomic features with ICC ≥ 0.75 are retained in the intratumoral and extratumoral imageomic features through correlation coefficient ICC evaluation; Among the imageomic features with ICC ≥ 0.75, 7 intratumoral imageomic features, 9 peritumoral imageomic features, and 8 intratumoral-peritumoral imageomic features are obtained through maximum correlation minimum redundancy mRMR method and least absolute shrinkage and selection operator LASSO method analysis; Six classification models for predicting early recurrence of pancreatic cancer based on intratumoral imageomic features are constructed through LR, SVM, RF, ExtraTrees, LightGBM, and MLP algorithms based on 7 intratumoral imageomic features and case labels in postoperative follow-up data; Six classification models for predicting early recurrence of pancreatic cancer based on peritumoral imageomic features are constructed through LR, SVM, RF, ExtraTrees, LightGBM, and MLP algorithms based on 9 peritumoral imageomic features and case labels in postoperative follow-up data. Six classification models for predicting early recurrence of pancreatic cancer based on intratumoral-peritumoral radiomics features were constructed by LR, SVM, RF, ExtraTrees, LightGBM and MLP algorithms based on 8 intratumoral-peritumoral radiomics features and case labels in postoperative follow-up data; Based on AUC index, ACC index, SEN index, SPE index, PPV index and NPV index, the performance differences between the three types of classifiers, including the classification model for predicting early recurrence of pancreatic cancer based on intratumoral radiomics features, the classification model for predicting early recurrence of pancreatic cancer based on peritumoral radiomics features and the classification model for predicting early recurrence of pancreatic cancer based on intratumoral-peritumoral radiomics features, were evaluated, and it was found that the classification model for predicting early recurrence of pancreatic cancer based on intratumoral-peritumoral radiomics features was superior to the classification model for predicting early recurrence of pancreatic cancer based on intratumoral radiomics features and the classification model for predicting early recurrence of pancreatic cancer based on peritumoral radiomics features; Based on AUC index, ACC index, SEN index, SPE index, PPV index and NPV index, the performance differences between the six classification models for predicting early recurrence of pancreatic cancer based on intratumoral-peritumoral radiomics features were evaluated, and it was found that the classification model for predicting early recurrence of pancreatic cancer based on LR and intratumoral-peritumoral radiomics features was superior to the classification models for predicting early recurrence of pancreatic cancer based on LR, SVM, ExtraTrees, LightGBM, MLP and intratumoral-peritumoral radiomics features; The classification model for predicting early recurrence of pancreatic cancer based on LR and intratumoral-peritumoral radiomics features was used as the radiomics model.

7. The method of claim 6, wherein the method is based on an interpretable machine model. The method for determining the clinical-radiomics combined model comprises: The serum CA12-5, visceral fat index VFI, adjuvant therapy, lymphatic vessel invasion LVI, tumor grade and TNM stage were combined into the radiomics model to obtain the clinical-radiomics combined model; The performance of the clinical-radiomics combined model was evaluated by ROC index, AUC index and DCA index.

8. The method of claim 7, wherein the method is based on an interpretable machine model. The method for performing SHAP analysis on the clinical-radiomics combined model comprises: Global explanation of the clinical-radiomics combined model was performed at the feature level by SHAP visualization analysis, and the contribution of intratumoral-peritumoral radiomics features and clinical features to the output of the clinical-radiomics combined model was obtained, wherein the contribution of intratumoral-peritumoral radiomics features was the largest, followed by clinical features, and high radiomics feature values were closely related to increased risk of early recurrence, high visceral fat index VFI, CA12-5 ≥ 35U / ml, positive lymphatic vessel invasion LVI, TNM stage ≥IIA and no adjuvant therapy were all related to increased risk of early recurrence, and intratumoral-peritumoral radiomics features and VFI had an interaction of increasing the risk of early recurrence. The SHAP visual analysis is applied to the individual level of the clinical-imaging-omics combined model to obtain the decision logic matching the contribution of the intratumor-peritumoral imaging-omics features and the clinical features to the output of the clinical-imaging-omics combined model for the cases without early recurrence of pancreatic ductal adenocarcinoma and the cases with early recurrence of pancreatic ductal adenocarcinoma during the follow-up period. 9.A system for predicting early recurrence of pancreatic cancer based on an interpretable machine model, the system comprising: The system is applied to the pancreatic cancer early recurrence prediction method based on the interpretable machine model according to any one of claims 1-8, and the system comprises: a data acquisition unit configured to acquire CT image data, clinical pathology data, and postoperative follow-up data of a case; a feature processing unit configured to segment intratumor and peritumor regions in the CT image, determine VOIs, and extract intratumor and peritumor imaging-omics features in the VOIs; perform quantitative analysis on body composition in the CT image to obtain body composition parameters, and perform single factor analysis and multivariate logistic regression analysis on the body composition parameters and the clinical pathology data to obtain clinical features; a model construction unit configured to construct six groups of classifier models based on the intratumor and peritumor imaging-omics features and the postoperative follow-up data by using six machine learning algorithms, obtain an imaging-omics model according to performance comparison of the six groups of classifier models, construct a clinical-imaging-omics combined model for pancreatic cancer early recurrence prediction by combining the clinical features and the imaging-omics model, and perform SHAP analysis on the clinical-imaging-omics combined model.

Images