Pancreas segmentation model construction method based on enhanced CT and three-dimensional mask

Abstract

The invention provides a pancreas segmentation model construction method based on enhanced CT (Computed Tomography) and a three-dimensional mask. The pancreas segmentation model construction method comprises the following steps: S1, acquiring an enhanced CT image and preprocessing the enhanced CT image and data; s2, automatically segmenting the enhanced CT image to obtain a three-dimensional mask of the pancreas; s3, image omics features and semantic features are extracted from the three-dimensional mask; and S4, screening the features, and constructing a learning classification model based on the screened features. The pancreas segmentation is automatically carried out on the enhanced CT image through the deep learning technology, and compared with a traditional manual segmentation method, the pancreas boundary can be identified more accurately, and the segmentation accuracy is improved. The automatic process from image acquisition to pancreas segmentation, feature extraction and model construction is realized, the operation process is greatly simplified, the manual intervention degree is reduced, and the working efficiency is improved.

Description

Technical Field

[0001] The invention relates to the technical field of medical image processing, and in particular to a method for constructing a pancreas segmentation model based on enhanced CT and three-dimensional mask. Background Art

[0002] Chronic pancreatitis (CP) is a chronic inflammatory disease characterized by abdominal pain, recurrent inflammation, and pancreatic fibrosis. Imaging changes during the course of the disease include pancreatic parenchymal calcification, pancreatic duct stones, and obstruction. The global incidence of CP is approximately 10 / 100,000, and it is increasing year by year. Pancreatic exocrine insufficiency (PEI) is a common functional change in CP patients, which seriously affects the quality of life of patients and increases the risk of malnutrition-related complications.

[0003] At present, the diagnosis of PEI mainly relies on the detection of pancreatic exocrine function, including direct and indirect tests. Although direct tests such as the secretin-cholecystokinin test have high sensitivity and specificity, they are rarely used clinically due to their invasiveness, high cost and great pain for patients. The indirect test of fecal elastase-1 (FE-1) is a relatively stable, accurate and convenient method, but it is not widely used due to the limited availability of kits. Imaging examinations such as CT scans are widely used in the diagnosis of CP, but they are not helpful for the detection of PEI.

[0004] In recent years, deep learning technology has shown great potential in the field of medical image analysis, especially in tumor detection, disease differential diagnosis and prognosis prediction. However, no study has used deep learning technology to evaluate PEI. At the same time, the segmentation difficulty of the pancreas with chronic pancreatitis is much greater than that of the normal pancreas or the pancreas with other diseases. Deep learning models, such as convolutional neural networks (CNNs) and generative adversarial networks (GANs), can extract complex features from medical images and perform high-level data analysis, providing new perspectives for the diagnosis and treatment of diseases. Despite the increasing application of deep learning in medical image analysis, there is currently a lack of a deep learning model for pancreatic segmentation that can accurately segment the pancreas from enhanced CT images and further evaluate the exocrine function of the pancreas. Existing pancreatic segmentation methods often require manual intervention and their accuracy needs to be improved.

[0005] Therefore, developing an efficient and accurate deep learning model for pancreas segmentation is of great significance for the early detection and clinical diagnosis of PEI. Summary of the invention

[0006] The object of the present invention is to provide a method for constructing a pancreatic segmentation model for chronic pancreatitis, so as to evaluate the pancreatic exocrine function based on CT images.

[0007] The present invention provides a method for constructing a pancreas segmentation model based on enhanced CT and three-dimensional mask, comprising the steps of:

[0008] S1, acquiring enhanced CT images and preprocessing the enhanced CT images and data;

[0009] S2, automatically segmenting the enhanced CT image to obtain a three-dimensional mask of the pancreas;

[0010] S3, extracting radiomics features and semantic features from the three-dimensional mask;

[0011] S4. Screening the features and building a learning classification model based on the screened features.

[0012] Furthermore, in step S1, the enhanced CT image and data are preprocessed including normalization and denoising to improve the robustness of the model.

[0013] Furthermore, in step S2, a 3D ResUnet network is used for automatic segmentation, and a variety of loss functions are used to optimize the segmentation network to improve segmentation accuracy.

[0014] Furthermore, the loss function includes Dice loss, focal loss and weighted BCE loss.

[0015] Furthermore, in step S3, the radiomics features are extracted using the Pyradiomics software package, including first-order histogram features and high-order texture features; and the semantic features are obtained by training a segmentation network.

[0016] Furthermore, in step S4, features are selected using recursive feature elimination, feature selection algorithm and least absolute shrinkage and selection operator.

[0017] Furthermore, step S4 also includes deleting redundant features to reduce the correlation between remaining features.

[0018] Furthermore, the learning classification model uses support vector machine, random forest, K nearest neighbor algorithm and logistic regression for classification; the learning classification model also includes a deep learning network, and the deep learning network includes ConvLSTM and ResNet-50.

[0019] Furthermore, in step S4, features with a high correlation with the severity of PEI are screened.

[0020] Furthermore, after step S4, the method further includes: using sensitivity, specificity, accuracy and area under the ROC curve to evaluate the classification performance of the learning classification model.

[0021] Compared with the prior art, the present invention includes at least the following beneficial effects: by using deep learning technology to automatically segment the pancreas on enhanced CT images, the present invention can more accurately identify the pancreatic boundary and improve segmentation accuracy compared to traditional manual segmentation methods. It realizes the automation process from image acquisition to pancreatic segmentation, feature extraction and model construction, greatly simplifies the operation process, reduces the degree of manual intervention, and improves work efficiency. By extracting radiomics features and semantic features, the present invention can more comprehensively reflect the lesion information of the pancreas, which helps to assist clinicians in screening CT images. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying creative work.

[0023] Figure 1 is a flow chart of a method for constructing a deep learning model for pancreas segmentation in one embodiment of the present invention;

[0024] Figure 2 It is a schematic diagram of segmenting a CT image in a method for constructing a deep learning model for pancreas segmentation in one embodiment of the present invention;

[0025] Figure 3 The visualization results of ResUnet under different loss functions in one embodiment of the present invention are shown;

[0026] Figure 4a is the ROC curve of only the radiomics features in the test cohort in one embodiment of the present invention;

[0027] Figure 4b is a ROC curve of only semantic features in a test queue in one embodiment of the present invention;

[0028] Figure 4c is the ROC curve of radiomics + semantic features in the test cohort in one embodiment of the present invention;

[0029] Figure 4d is a ROC curve of a deep neural network in a test queue in one embodiment of the present invention;

[0030] Figure 5 Enhanced CT images and corresponding thermal maps randomly selected from a test queue in one embodiment of the present invention. DETAILED DESCRIPTION

[0031] A method for constructing a pancreas segmentation model based on enhanced CT and a three-dimensional mask of the present invention will be described in more detail below in conjunction with a schematic diagram, wherein a preferred embodiment of the present invention is shown, and it should be understood that those skilled in the art can modify the present invention described herein, while still achieving the advantageous effects of the present invention. Therefore, the following description should be understood as being widely known to those skilled in the art, and not as a limitation of the present invention.

[0032] It should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprises" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device.

[0033] The present invention is described in more detail in the following paragraphs by way of example with reference to the accompanying drawings. The advantages and features of the present invention will become more apparent from the following description. It should be noted that the accompanying drawings are in very simplified form and are not in exact proportions, and are only used to facilitate and clearly assist in illustrating the purpose of the embodiments of the present invention.

[0034] The present invention provides a method for constructing a deep learning model for pancreas segmentation. Figure 1 ,include:

[0035] S1, acquiring enhanced CT images and preprocessing the enhanced CT images and data;

[0036] S2, automatically segmenting the enhanced CT image to obtain a three-dimensional mask of the pancreas;

[0037] S3, extracting radiomics features and semantic features from the three-dimensional mask;

[0038] S4. Filter the features according to the conditions, and build a learning classification model based on the filtered features.

[0039] Specifically, in step S1, a computer tomography (CT) technique is used to obtain an image of the pancreas. Enhanced CT is a special CT scan that improves image quality by injecting contrast agents into the patient's veins, making the structure of organs such as the pancreas clearer. Preprocessing is a series of processing of these images to facilitate subsequent analysis and ensure the consistency of image quality and the input requirements of the model.

[0040] In step S2, the segmentation process is to separate the pancreas from the surrounding organs and tissues and create a mask on the image, which marks the location of the pancreas. Since it is a three-dimensional image, the segmentation result will be a three-dimensional mask that can show the exact location and shape of the pancreas in three-dimensional space.

[0041] In step S3, features that are helpful for diagnosis and analysis are extracted from the segmented 3D mask of the pancreas. Radiomics features refer to quantitative features extracted from medical images, such as shape, size, texture, etc. Semantic features refer to higher-level information related to the image content, such as the type and location of pancreatic lesions.

[0042] In step S4, the extracted features are screened to remove irrelevant or redundant features and retain the most valuable features for evaluating pancreatic exocrine function. The screening process is based on the importance of the features, their relevance to the disease, or other statistical criteria. Afterwards, the screened features are used to build a deep learning classification model that can predict whether the pancreatic exocrine function is normal based on the input CT image features, thereby assisting doctors in making diagnosis and treatment decisions.

[0043] Furthermore, in step S1, the enhanced CT image and data are preprocessed including normalization and denoising to improve the robustness of the model.

[0044] Specifically, normalization refers to adjusting the pixel values ​​of CT images to a uniform range, usually between 0 and 1. Since different CT scanning devices will produce different ranges of pixel values, and deep learning models usually expect the input data to be within a standard range. Normalization can eliminate these device differences and ensure the consistency of the model's performance on different data sets. In addition, normalization can also speed up the training of the model and improve convergence. Denoising refers to reducing the noise in the image. The noise comes from the scanning equipment or environmental factors. They will introduce undesirable random changes in the image, affecting the quality of the image and subsequent analysis. The denoising operation can use a variety of methods, such as filters, wavelet transforms, non-local mean denoising, etc. Through denoising, the clarity of the image can be improved, making the outline of the pancreas more obvious, thereby improving the accuracy of the deep learning model in segmentation and feature extraction.

[0045] Through normalization and denoising, the deep learning model can be made more stable and reliable, that is, the robustness of the model can be improved. This means that the model can maintain high performance when facing CT images of different quality and different sources, reducing the possibility of misclassification or segmentation, thereby providing more accurate results in clinical applications.

[0046] For further information, please refer to Figure 2In step S2, a 3D ResUnet network is used for automatic segmentation, and a variety of loss functions are used to optimize the segmentation network to improve segmentation accuracy.

[0047] Furthermore, the loss function includes Dice loss, focal loss and weighted BCE loss.

[0048] Specifically, 3D ResUnet is a deep learning network that is based on the Unet network structure and combines the characteristics of the residual network. The Unet network is a convolutional neural network widely used in medical image segmentation. It has two parts, an encoder and a decoder, which can effectively perform feature extraction and accurate image segmentation. Compared with 2D Unet, 3D ResUnet can better process three-dimensional image data because it simultaneously considers the image information in height, width and depth, which is crucial for the accurate segmentation of three-dimensional structures such as the pancreas. The introduction of the residual network solves the gradient vanishing problem in deep network training through jump connections, making the network deeper and having better performance.

[0049] In a possible embodiment of the present invention, please refer to Figure 2 , using a coarse-to-fine segmentation pipeline, the baseline network chooses the classic 3D ResUnet. The baseline network was retrained in both the coarse and fine segmentation steps. ResUnet is an evolution of the original U-Net, which contains an encoding path consisting of five convolutional pooling layers to extract multi-resolution semantic features. Afterwards, features of different scales are concatenated and upsampled to obtain a segmentation mask of the same size as the input image. ResUnet adds residual learning layers and jump connections in the encoder and decoder modules.

[0050] In order to further improve the accuracy of the segmentation network, a variety of loss functions were used to optimize the network. The loss function is an indicator that measures the difference between the model's predicted value and the true value. The network is trained by minimizing the loss function. That is, the hybrid loss function of Dice loss combined with BCE loss, the focal loss function, and the weighted BCE loss function. The network was evaluated for each patient using evaluation indicators such as Dice similarity coefficient (DSC), accuracy, sensitivity, and Hausdorff distance (HD).

[0051] Dice loss is a loss function commonly used in image segmentation. It is based on the Dice coefficient and is used to measure the overlap between the predicted segmentation and the true segmentation. Dice loss can effectively deal with the problem of class imbalance in segmentation tasks. Focal Loss is a loss function used to solve the problem of class imbalance. It adjusts the weight of each sample so that the model pays more attention to samples that are difficult to classify. In the segmentation task, focal loss can help the network better learn the boundaries of the pancreas. Weighted BCE loss (Weighted Binary Cross-Entropy Loss) is a variant of binary cross entropy loss. It solves the problem of class imbalance in the data set by assigning different weights to different categories. In pancreatic segmentation, weighted BCE loss can help the network better identify the pancreatic region, especially when the contrast between the pancreas and surrounding tissues is not high. Table 1 shows a comparison of 3D ResUnet segmentation results with different loss functions. Figure 3 The visualization results of ResUnet under different loss functions, where the arrows indicate the improved areas.

[0052] Table 1

[0053]

[0054] According to Table 1, the 3D ResUnet network that combines Dice loss, Focal loss and weighted BCE loss performs well in reducing HD and significantly improves the performance. Figure 3 The pancreatic masks obtained by the above models are visualized in Figure 2. The best segmentation network has the closest boundaries to the manually annotated boundaries in the pancreatic head and tail regions, thereby improving HD. In addition, the best segmentation network shows robust performance in both patients with and without CP. Based on accurate segmentation, the robustness and reliability of the extracted radiomics and semantic features can be ensured, thereby improving the accuracy of classification.

[0055] Furthermore, in step S3, the radiomics features are extracted using the Pyradiomics software package, including first-order histogram features and high-order texture features; and the semantic features are obtained by training a segmentation network.

[0056] Pyradiomics is an open-source Python library for extracting radiomics features from medical images. Radiomics is a method of converting medical images into quantifiable features that can be used for disease diagnosis, prognosis assessment, and treatment response prediction. First-order histogram features describe the distribution of voxel intensities in an image, including mean, standard deviation, skewness, kurtosis, entropy, energy, uniformity, and coefficient of variation. First-order histogram features provide statistical information about the grayscale of an image, which helps to understand the contrast and uniformity of the image. Higher-order texture features describe the relationship between voxels in an image, including the roughness, contrast, uniformity, and complexity of the texture. Higher-order texture features can capture microstructural information in images and help to distinguish different types of tissues.

[0057] Semantic features usually refer to features that are related to the image content and have clear meaning, such as the shape, size, and position of the organ. In the context of pancreas segmentation, semantic features include the outline of the pancreas, its relative position to the surrounding tissue, etc. This semantic information can be extracted through the feature maps of the intermediate layers of the network, or indirectly obtained through the output of the segmentation network (such as the 3D mask of the pancreas). These features are very valuable for subsequent classification models and help improve the prediction ability of the model.

[0058] In one possible embodiment of the present invention, CT-based radiomics features were extracted from the automatic segmentation mask using the Pyradiomics package (version 2.2.0) in Python (version 3.6.5), in accordance with the guidelines of the Image Biomarker Standardization Initiative. First-order histogram features and high-order texture features, grayscale co-occurrence matrix features, and grayscale run length matrix features were extracted. These features describe the pattern of grayscale pixel value distribution within the region of interest (ROI), and the histogram features obtained from the pancreatic ROI were averaged as representative values ​​for each patient.

[0059] Furthermore, in step S4, recursive feature elimination (RFE), a feature selection algorithm (ReliefF) and a least absolute shrinkage and selection operator (LASSO) are used to select features.

[0060] Recursive feature elimination is a feature selection method that works by recursively considering a smaller set of features. In RFE, the features are first scored using a classifier and then the least important features are removed. This process is repeated until the desired number of features is reached. RFE is able to effectively reduce the number of features while retaining the features that are most important for the model's predictions.

[0061] ReliefF is an instance-based feature selection algorithm that evaluates the importance of features by comparing feature values ​​between samples of the same and different classes. The ReliefF algorithm considers the ability of features to distinguish adjacent heterogeneous samples, so it is particularly suitable for processing data sets with noisy and redundant features. With ReliefF, the most discriminative features can be selected for distinguishing between healthy and diseased pancreas.

[0062] LASSO is a penalty function for linear regression, which implements feature selection and regularization by adding an absolute value penalty term (L1 penalty) to the loss function. LASSO can compress the coefficients of some features to zero, thereby achieving sparse selection of features, that is, automatically selecting the most important features for predicting the target. LASSO is particularly useful when dealing with high-dimensional data because it can help identify truly relevant features while removing unimportant features.

[0063] By combining these three feature selection methods, the most predictive features can be effectively selected from a large number of radiomics features and semantic features, thereby building a more accurate and efficient learning classification model. Such a model can not only improve the diagnostic accuracy of chronic pancreatitis, but also provide strong support for the assessment of pancreatic exocrine insufficiency.

[0064] Furthermore, step S4 also includes deleting redundant features to reduce the correlation between remaining features.

[0065] Redundant features are those that are highly correlated with other features and contain the same or very similar information. In model training, redundant features not only do not provide additional information gain, but may also increase the complexity of calculations and cause model overfitting. Excessive correlation between features can lead to multicollinearity problems, which will affect the stability and predictive ability of the model. By reducing the correlation between features, collinearity problems can be reduced so that each feature can provide unique and non-overlapping information to the model. Removing redundancy can improve the interpretability of the model and help improve the predictive accuracy and robustness of the model.

[0066] In a possible embodiment of the present invention, redundant features are deleted to ensure that the correlation coefficients between the remaining features are higher than 0.85.

[0067] Furthermore, the learning classification model uses support vector machine (SVM), random forest (RF), K nearest neighbor algorithm (KNN) and logistic regression (LR) for classification. The learning classification model also includes a deep learning network, and the deep learning network includes ConvLSTM and ResNet-50.

[0068] When building a learning classification model, a variety of traditional machine learning algorithms are used. These algorithms each have different characteristics and capabilities, and their combination can improve the generalization ability and prediction accuracy of the model.

[0069] Support vector machine (SVM) is an effective binary classification model that can find the optimal hyperplane that separates the data set. It is suitable for small and medium-sized complex data sets. Random forest (RF) is an ensemble learning method that improves the accuracy and stability of predictions by building multiple decision trees and voting. It is particularly suitable for processing data with a large number of features. K-nearest neighbor algorithm (KNN) is an instance-based learning method that makes predictions by finding the most similar known instances. It is suitable for classification and regression problems. Logistic regression (LR) is a widely used classification method that predicts probabilities through maximum likelihood estimation. It is particularly suitable for processing binary classification problems.

[0070] In addition to traditional machine learning algorithms, deep learning networks are also incorporated when building learning classification models. ConvLSTM is a convolutional long short-term memory network that combines the characteristics of convolutional neural networks (CNN) and long short-term memory networks (LSTM), and is particularly suitable for processing data with spatial and temporal dependencies, such as video sequences or medical image sequences. ResNet-50 is a deep residual network that solves the gradient vanishing problem in deep network training by introducing residual learning. It can effectively extract deep features of images and is suitable for image recognition and classification tasks. Table 2 shows the performance evaluation of machine learning models and deep learning networks, including different combinations of feature selection methods and classifiers. Sen, Spec, ACC and AUC represent sensitivity, specificity, accuracy and area under the ROC curve, respectively, and the results are expressed in percentages.

[0071] Table 2

[0072]

[0073] By combining traditional machine learning algorithms and deep learning networks, we can fully utilize the advantages of different models and improve the prediction performance and robustness of the models. This multi-model fusion strategy helps to better process complex medical image data, thereby achieving better results in pancreatic segmentation and classification tasks.

[0074] Furthermore, in step S4, features with a high correlation with the severity of PEI are screened.

[0075] When selecting features, priority will be given to those with a high statistical correlation with the severity of PEI. The selected features should be able to well reflect the severity of PEI, thereby providing a strong predictive basis for the subsequent classification model.

[0076] Specifically, the correlation coefficient between each feature and the severity of PEI (such as the Pearson correlation coefficient or the Spearman rank correlation coefficient) is calculated to evaluate the correlation between the feature and the severity of PEI. Secondly, according to the preset correlation threshold, those features that are highly correlated with the severity of PEI are retained, while those with low correlation are eliminated. Finally, in order to ensure the effectiveness of the selected features, cross-validation is also required to verify the predictive ability of these features on independent data sets.

[0077] Through such screening conditions, it can be ensured that the features that are ultimately used to construct the learning classification model are those that are most valuable for PEI diagnosis, thereby improving the accuracy and practicality of the model in predicting the severity of PEI.

[0078] Furthermore, after step S4, the method further includes: evaluating the classification performance of the learning classification model using indicators such as sensitivity, specificity, accuracy and area under the ROC curve (AUC).

[0079] Specifically, sensitivity refers to the proportion of patients who actually have PEI that the model correctly identifies, and is used to evaluate the model's ability to identify positive cases. Specificity refers to the proportion of patients who actually do not have PEI that the model correctly identifies, and is used to evaluate the model's ability to exclude negative cases. Accuracy refers to the proportion of samples correctly classified by the model to the total number of samples. It is used to evaluate the overall classification accuracy of the model.

[0080] The ROC (Receiver Operating Characteristic) curve is a graphical tool used to show the sensitivity and specificity of a model at different thresholds. AUC is the area under the ROC curve, with a value range of 0.5 to 1. The higher the value, the better the model performance. AUC provides a comprehensive indicator for evaluating the model's ability to distinguish between positive and negative samples, and is one of the important indicators of classification model performance. For specific ROC curves of each evaluation model, please refer to Figure 4a , Figure 4b , Figure 4c and Figure 4d shown.

[0081] Specifically, the data set is divided into a training set and a test set, where the training set is used to build the model and the test set is used to evaluate the model performance. The classification model is trained using the training set data. The above indicators are calculated using the test set data, and the performance of the model is evaluated by these indicators. Based on the calculated indicator values, the advantages and disadvantages of the model are analyzed, and the model is adjusted or optimized as needed.

[0082] Figure 5The input CT images randomly selected from the test cohort and the heat maps of the corresponding locations are shown. Despite the interference of the closely surrounding splenic artery and vein, the deep learning model showed good attention to the structure of the pancreas, which helped to correct the difference between the ROI and the easily confused areas of similar contrast and improve the recognition performance. At the same time, the heat map paid more attention to the stone area, indicating that the model believed that stones were significantly associated with PEI. In addition, different deep learning models have different focuses on the adjacent areas of the pancreas of the same patient. This difference explains the differences in the predictive performance of these models. Specifically, the heat map of the ResNet-50 model contains more gradient attention on specific areas related to the pancreas itself, in addition to the duodenum and splenic vessels.

[0083] Through these evaluation indicators, we can fully understand the performance of the learning classification model in predicting the severity of PEI, thereby providing reliable support for clinical decision-making. If the performance of the model does not meet the requirements, it is necessary to go back to the previous steps and adjust the feature selection, model architecture or training process. Table 3 compares the classification results of clinicians with and without AI assistance, and all indicators are expressed in percentages.

[0084] Table 3

[0085]

[0086] The above specific examples are used to illustrate the present invention, which is only used to help understand the present invention and is not intended to limit the present invention. For those skilled in the art, according to the concept of the present invention, some simple deductions, modifications or substitutions can be made.

Claims

1. A method for constructing a pancreatic segmentation model based on enhanced CT and three-dimensional mask, characterized in that: Includes steps: S1, acquiring enhanced CT images and preprocessing the enhanced CT images and data; S2, automatically segmenting the enhanced CT image to obtain a three-dimensional mask of the pancreas; S3, extracting radiomics features and semantic features from the three-dimensional mask; S4. Screening the features and building a learning classification model based on the screened features.

2. The method for constructing a pancreatic segmentation model based on enhanced CT and three-dimensional mask according to claim 1, characterized in that: In the step S1, the enhanced CT image and data are preprocessed including normalization and denoising to improve the robustness of the model.

3. The method for constructing a pancreatic segmentation model based on enhanced CT and three-dimensional mask according to claim 1, characterized in that: In step S2, a 3D ResUnet network is used for automatic segmentation, and a variety of loss functions are used to optimize the segmentation network to improve segmentation accuracy.

4. The method for constructing a pancreatic segmentation model based on enhanced CT and three-dimensional mask according to claim 3, characterized in that: The loss functions include Dice loss, focal loss and weighted BCE loss.

5. The method for constructing a pancreatic segmentation model based on enhanced CT and three-dimensional mask according to claim 1, characterized in that: In step S3, the radiomics features are extracted using the Pyradiomics software package, including first-order histogram features and high-order texture features; the semantic features are obtained by training a segmentation network.

6. The method for constructing a pancreatic segmentation model based on enhanced CT and three-dimensional mask according to claim 1, characterized in that: In step S4, features are selected using recursive feature elimination, feature selection algorithm and least absolute shrinkage and selection operator.

7. The method for constructing a pancreatic segmentation model based on enhanced CT and three-dimensional mask according to claim 1, characterized in that: The step S4 also includes deleting redundant features and reducing the correlation between the remaining features.

8. The method for constructing a pancreas segmentation model based on enhanced CT and three-dimensional mask according to claim 1, characterized in that: The learning classification model uses support vector machine, random forest, K nearest neighbor algorithm and logistic regression for classification; the learning classification model also includes a deep learning network, and the deep learning network includes ConvLSTM and ResNet-50.

9. The method for constructing a pancreatic segmentation model based on enhanced CT and three-dimensional mask according to claim 1, characterized in that: In step S4, features with a high correlation with the severity of PEI are screened.

10. The method for constructing a pancreas segmentation model based on enhanced CT and three-dimensional mask according to claim 1, characterized in that: After step S4, the method further includes: evaluating the classification performance of the learning classification model using sensitivity, specificity, accuracy and area under the ROC curve (AUC).

Images