A multi-view pancreatic tumor subtype classification method based on super-evidence accumulation
By introducing medical prior knowledge into a multi-view learning method to generate hyperevidence and dynamically estimate view uncertainty, this approach addresses the performance degradation and insufficient robustness of existing methods in medical image classification, improves the accuracy and stability of pancreatic tumor subtype classification, and is applicable to the field of medical image processing.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- TONGJI UNIV
- Filing Date
- 2025-06-11
- Publication Date
- 2026-08-04
AI Technical Summary
Existing multi-view learning methods cannot effectively utilize prior medical knowledge in medical image classification, resulting in decreased classification performance and insufficient robustness. They are particularly ineffective when faced with data noise, and the uncertainty estimation is inaccurate, affecting the interpretability and robustness of the classification.
By introducing medical prior knowledge to generate super-evidence, including basic evidence and auxiliary evidence, dynamically estimating the uncertainty of each perspective, and calculating the perspective fusion weight based on the uncertainty, a multi-perspective pancreatic tumor subtype classification network framework based on super-evidence accumulation is constructed to improve feature fusion effect and robustness.
It improves the accuracy and robustness of pancreatic tumor subtype classification, reduces the upper bound of classification generalization error, ensures stability and reliability in the face of noise, and provides a reliable basis for clinical diagnosis and treatment.
Smart Images
Figure CN120612546B_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of computer medical image processing, specifically relating to a multi-view pancreatic tumor subtype classification method based on super-evidence accumulation. Background Technology
[0002] In recent years, multi-view learning methods have been widely applied to image classification tasks. In natural image scenarios, multi-view learning methods have demonstrated advantages in classification performance, computational performance, and scalability. However, due to the limitations of imaging techniques specific to medical images, image quality is easily affected by external factors during the shooting process, leading to a decline in image quality. Some low-quality data may contain noise and artifacts. Furthermore, in clinical practice, the diagnosis of pancreatic tumor subtypes often relies on prior medical knowledge. Pancreatic tumors are initially coarsely classified into high-risk, low-risk, and non-tumor types, followed by further subdivision into various subtypes. Existing multi-view learning methods cannot model this prior medical knowledge. Due to the scarcity of medical imaging data, existing multi-view learning methods often cannot acquire this prior knowledge through learning from large datasets. This results in poor performance of existing multi-view learning methods in pancreatic tumor subtype classification tasks. Moreover, when faced with data noise, existing uncertainty-aware multi-view learning methods suffer from poor multi-view information fusion due to their inability to accurately estimate uncertainty, leading to a significant decline in classification performance and a lack of robustness. The aforementioned problems result in poor interpretability and unreliable decision-making in existing multi-view learning methods, and their robustness in practical applications cannot be guaranteed. Summary of the Invention
[0003] To overcome the performance degradation and insufficient robustness of current multi-view learning methods in medical image classification, this invention proposes a multi-view pancreatic tumor subtype classification method based on super-evidence accumulation. By introducing medical prior knowledge to generate super-evidence including auxiliary evidence, the uncertainty of each viewpoint can be estimated more accurately, thereby effectively improving the effect of multi-view feature fusion and enhancing the robustness of the classification network.
[0004] Technical solution of the present invention:
[0005] A multi-perspective pancreatic tumor subtype classification method based on super-evidence accumulation includes the following steps:
[0006] Step 1: Dataset preparation: Collect multi-view medical images from the hospital, and after classification and tumor location annotation by professional clinicians, construct a multi-view pancreatic tumor subtype image dataset through data preprocessing.
[0007] Step 2: Construct a multi-perspective pancreatic tumor subtype classification network framework based on super-evidence accumulation, and train and optimize the multi-perspective pancreatic tumor subtype classification network framework based on super-evidence accumulation using the dataset from Step 1:
[0008] Step 3: Using the optimized multi-view pancreatic tumor subtype classification network framework trained in Step 2, classify the multi-view pancreatic tumor subtype images to be predicted.
[0009] Furthermore, in step 1, the specific steps of data preprocessing are as follows:
[0010] First, three consecutive CT slices are cut based on the largest cross-section of the tumor region, and the slice size is uniformly set to [512, 512].
[0011] Then, based on the tumor location annotations, a square region of interest (ROI) is cropped centered on the pancreatic tumor. Overlapping sliding windows are used within the ROI to obtain four sub-views of the pancreatic tumor. The entire ROI region is used as a global view. For multi-view extraction, the ROI, window, and stride sizes are set to 160, 128, and 32, respectively.
[0012] Next, the pixel values of the CT image are truncated.
[0013] Specifically, the pixel values of the CT image are truncated using a window with a width of 200 and a position of 40. Pixel values less than -160 are set to -160, pixel values greater than 240 are set to 240, and the remaining pixel values remain unchanged.
[0014] Finally, the CT images were normalized using min-max scaling.
[0015] Furthermore, step 2:
[0016] Training of a multi-perspective pancreatic tumor subtype classification network framework based on super-evidence accumulation:
[0017] Step 2.1: Input the multi-view pancreatic tumor subtype images obtained in Step 1 into the view encoder respectively to encode the multi-view medical image features;
[0018] Step 2.2: Obtain hyper-evidence from various perspectives through the hyper-evidence generation module. The process includes:
[0019] The hyperevidence generation module accepts the image features generated in step 2.1 as input, generates hyperevidence after passing through an activation function, and calculates the uncertainty;
[0020] The super-evidence consists of two types of evidence: primary evidence and secondary evidence. Primary evidence represents the degree of support of image features for each pancreatic tumor subtype, while secondary evidence represents the degree of support of image features for the broader category of pancreatic tumor subtypes. Since secondary evidence directly measures the ambiguity between pancreatic tumor subtypes, the generation of secondary evidence can make uncertainty estimation more accurate.
[0021] It should be noted that, based on prior medical knowledge, pancreatic tumors are initially broadly classified into three subtypes: high-risk tumors, low-risk tumors, and non-tumor tumors, and then further subdivided into each subtype.
[0022] Step 2.3: Obtain fused hyper-evidence through the multi-view fusion module. The process includes:
[0023] The multi-view fusion module takes the hyperevidence and uncertainty of each view generated in step 2.2 as input, and uses the uncertainty of each view to generate the fusion weight of each view: the fusion weight obtained based on the uncertainty estimation characterizes the quality of feature extraction of each view; the uncertainty of the view where low-quality noisy data is located is higher, thus generating a lower fusion weight; the uncertainty of the view where high-quality data is located is lower, thus generating a higher fusion weight.
[0024] Based on the weights of the perspectives, the hyperevidence from each perspective is weighted and summed to obtain the fused hyperevidence: more accurate uncertainty estimation can generate more accurate fusion weights; more accurate fusion weights can reduce the upper bound of the generalization error of the fused classification and ensure the robustness of the classification model in the face of noise.
[0025] Step 2.4: Obtain fused multinomial evidence through the hyper-evidence projection module. The process includes:
[0026] The fused hyper-evidence obtained in step 2.3 is input into the hyper-evidence projection module to generate fused multinomial evidence; wherein the projection process is for the auxiliary evidence in the hyper-evidence, and according to the prior weights, the auxiliary evidence is assigned to the basic evidence for the classification of all pancreatic tumor subtypes;
[0027] Step 2.5: The fusion of multinomial evidence obtained in Step 2.4 and the medical image classification obtained in Step 1 are learned, and the logarithmic loss is used to calculate the loss; by training and optimizing the overall network composed of the above-mentioned perspective encoder, hyper-evidence generation module, multi-view fusion module and hyper-evidence projection module, a stable multi-view pancreatic tumor subtype classification network based on hyper-evidence accumulation is obtained.
[0028] Step 2.6: Repeat steps 2.1 to 2.5 above until the multi-perspective pancreatic tumor subtype classification network based on super-evidence accumulation is stable and save the network parameters after training.
[0029] Beneficial effects
[0030] This invention proposes a multi-view pancreatic tumor subtype classification method based on super-evidence accumulation to classify clinical multi-view pancreatic tumor subtype images, improving classification accuracy and robustness. The multi-view pancreatic tumor subtype classification network framework constructed in this invention generates super-evidence by introducing medical prior knowledge, thereby more accurately estimating the uncertainty of each viewpoint. Based on more accurate viewpoint uncertainty estimation, more accurate viewpoint fusion weights can be generated, effectively improving the effect of multi-view feature fusion, reducing the upper bound of the classification generalization error after fusion, and enhancing the robustness of the classification network. As a medical image assistance tool, it provides a basis for doctors' subsequent diagnosis and treatment. Attached Figure Description
[0031] Figure 1 This is a flowchart of the method of the present invention;
[0032] Figure 2 This is a flowchart of the training and optimization method according to an embodiment of the present invention;
[0033] Figure 3 This is a flowchart illustrating the multi-perspective pancreatic tumor subtype classification based on super-evidence accumulation, as described in an embodiment of the present invention.
[0034] Figure 4 This is a schematic diagram of the processing module for classifying pancreatic tumor subtype images from multiple perspectives during the training phase of an embodiment of the present invention;
[0035] Figure 5 This is a schematic diagram of the processing module for classifying pancreatic tumor subtypes on multi-view images during the testing phase of an embodiment of the present invention;
[0036] Figure 6 This is the data preprocessing process in an embodiment of the present invention;
[0037] Figure 7 This is a comparison of the data metrics of this invention with other multi-view classification methods in the task of classifying pancreatic tumor subtypes. Detailed Implementation
[0038] The technical solutions provided in this application will be further described below with reference to specific embodiments and accompanying drawings. The advantages and features of this application will become clearer from the following description.
[0039] A multi-perspective pancreatic tumor subtype classification method based on super-evidence accumulation includes the following steps: (e.g.) Figure 1 )
[0040] Step 1: Dataset preparation: Collect multi-view medical images from the hospital, and after classification and tumor location annotation by professional clinicians, construct a multi-view pancreatic tumor subtype image dataset through data preprocessing.
[0041] Step 2: Construct a multi-perspective pancreatic tumor subtype classification network framework based on super-evidence accumulation, and train and optimize the multi-perspective pancreatic tumor subtype classification network framework based on super-evidence accumulation using the dataset from Step 1:
[0042] Step 3: Using the optimized multi-view pancreatic tumor subtype classification network framework trained in Step 2, classify the multi-view pancreatic tumor subtype images to be predicted.
[0043] Step 2 describes a multi-view pancreatic tumor subtype classification network framework based on super-evidence accumulation, which includes: a view encoder, a super-evidence generation module, a multi-view fusion module, and a super-evidence projection module, wherein:
[0044] Viewpoint Encoder: The viewpoint encoder takes a single-view image as input and extracts features from the single-view image. The viewpoint encoder uses a ResNet18 network to encode image features. The ResNet18 network consists of an input layer, a residual layer, a global average pooling layer, and a fully connected layer. The input layer consists of one convolutional layer and one pooling layer. The residual layer consists of 16 convolutional layers and eight residual layers. The fully connected layer is used to integrate features from the feature maps output by the previous layer, thereby achieving data classification.
[0045] Hyper-evidence Generation Module: This module is connected to the perspective encoder and the multi-perspective fusion module. It takes features from the last layer output of the perspective encoder as input and utilizes prior medical knowledge about pancreatic tumor subtypes to generate hyper-evidence and uncertainties for each perspective. The prior medical knowledge is a coarse classification strategy; based on the potential malignancy risk of pancreatic tumors, all pancreatic tumor subtypes can be divided into three categories: high-risk malignancy tumor subtypes, low-risk malignancy tumor subtypes, and non-tumor subtypes. The hyper-evidence generation module consists of neural network activation layers.
[0046] Multi-view fusion module: The multi-view fusion module is connected to the hyper-evidence generation module. The multi-view fusion module is used to fuse the hyper-evidence generated by each view into fused hyper-evidence. The hyper-evidence and uncertainty of each view output by the hyper-evidence generation module are used as inputs to the multi-view fusion module. The multi-view fusion module generates view weights based on the uncertainty of each input view and performs a weighted summation of the hyper-evidence of each input view to obtain the fused hyper-evidence.
[0047] Hyper-evidence projection module: The hyper-evidence projection module is connected to the multi-view fusion module and is used to project the fused hyper-evidence into fused multinomial evidence; wherein the fused multinomial opinion only contains basic evidence; since the fused hyper-evidence needs to be classified into pancreatic tumor subtypes, the auxiliary evidence in the fused hyper-evidence will be allocated to the basic evidence in the fused hyper-evidence according to the prior weight; the fused multinomial opinion will be used for pancreatic tumor subtype classification.
[0048] Preparation, training optimization, and classification process based on a multi-perspective pancreatic tumor subtype classification network framework built upon super-evidence accumulation:
[0049] Step 1, Dataset Preparation, the specific implementation process is as follows:
[0050] Data on pancreatic tumor subtypes from multiple perspectives, along with their corresponding classification labels, were collected from hospitals to create a dataset, denoted as [database name missing]. ,in: This represents a multi-view image set of pancreatic tumor subtypes. Indicates the number of viewpoints. Indicates data category labels;
[0051] Step 2, Training Optimization ( Figure 2 As shown in the figure, the specific implementation process is as follows:
[0052] 2.1: Randomly select samples from the dataset Multi-view images of pancreatic tumor subtypes The input view encoder extracts features from each viewpoint to obtain the features of each viewpoint. ;in, It is the first Features from a single perspective;
[0053] 2.2: Features from various perspectives Input hyperevidence generation module, where the neural network activation layer is Obtain super-evidence from various perspectives And the uncertainty of various perspectives .in, It is the first Super evidence from a different perspective It is the first Uncertainty from one perspective;
[0054] No. Hyperevidence from a Different Perspective Based on basic evidence and supporting evidence The composition includes: primary evidence, which represents the degree of support of image features for each pancreatic tumor subtype; and secondary evidence, which represents the degree of support of image features for three major pancreatic tumor subtype categories derived from prior medical knowledge. The two types of evidence for each perspective, along with the uncertainty calculation formula, are as follows:
[0055]
[0056] Where K represents the number of categories, that is, the number of classifications of pancreatic tumor subtypes.
[0057] Without the introduction of supporting evidence, the following constraints exist between uncertainty and evidence:
[0058]
[0059] When introducing supporting evidence, the following constraints exist between uncertainty and the evidence:
[0060]
[0061] As can be seen from the constraint equations above, since the sum of evidence and uncertainty remains unchanged, the introduced auxiliary evidence is obtained by separating it from uncertainty through explicit measurement. Uncertainty measures out-of-distribution uncertainty, while auxiliary evidence describes the ambiguity between categories rather than out-of-distribution uncertainty. Therefore, after introducing auxiliary evidence, the uncertainty estimate for each view will be more accurate.
[0062] 2.3: Obtaining Fusion Hyperevidence The process includes:
[0063] Uncertainty of each perspective obtained from 2.2 The fusion weights for each viewpoint are calculated. The calculation formula is:
[0064]
[0065] in The hyper-evidence obtained from various perspectives in section 2.2 The fusion super-evidence is obtained by weighted summation based on the fusion weights of each perspective. The calculation formula is:
[0066]
[0067] It can be proven that the classifier obtained through the above dynamic fusion method... The upper bound of the generalization error is calculated using the following formula:
[0068]
[0069] in This represents the classifier for each view. This represents the expected blending weight for each view. Representing Radmach complexity, This represents the classification loss for each view. The calculation formula shows that when the fusion weights and classification loss are negatively correlated, their covariance is negative, and the generalization error after fusion will decrease.
[0070] In this invention, since the method of generating super-evidence by introducing auxiliary evidence can more accurately estimate the uncertainty of each view, the generation of fusion weights for each view will be more accurate, thereby reducing the covariance of the fusion weights and classification loss of the view, and reducing the upper bound of the generalization error of the fused classifier.
[0071] Generalization error is the difference between a machine learning model's performance on unseen data and its performance on training data. It measures the model's predictive ability on new data and reflects its generalization performance. A lower upper bound on the generalization error ensures the robustness and reliability of this invention when faced with low-quality data.
[0072] 2.4: Obtaining Fusion Polynomial Evidence The process includes:
[0073] The fusion hyper-evidence obtained in 2.3 The fusion polynomial evidence was calculated based on the evidence projection formula. The projection formula is:
[0074]
[0075] in It is a size of The vector, It is a size of The matrix. Represents the prior weight, which is a function of size . The matrix is given by K, where K represents the number of pancreatic tumor subtypes. For the pancreatic tumor subtype classification task, The CCP's evidence includes evidence for three major categories and evidence for K minor categories. Therefore, when setting the projection weights, assume that the three major categories each contain x, y, and z minor categories, where x, y, and z satisfy... , Set it to the following form: (previous) Behavior An identity matrix, the last three rows of non-zero elements are respectively indivual , indivual , indivual )
[0076]
[0077] Because it requires extra-evidence The process of obtaining classification results and fusing projection involves fusing hyperevidence based on prior weights. The auxiliary evidence is assigned to the basic evidence; the fused polynomial evidence obtained after projection It contains only basic evidence;
[0078] 2.5: Obtain fused polynomial evidence from step 2.4 Compared with the data classification labels in the dataset of step 1 Log loss was calculated to train a multi-perspective pancreatic tumor subtype classification network based on super-evidence accumulation;
[0079] The logarithmic loss is:
[0080] ,
[0081] in This indicates the number of pancreatic tumor subtypes; express The The component, i.e., the first component Fusion polynomial evidence for several pancreatic tumor subtypes; Indicates the sample number One-hot classification labels for pancreatic tumor subtypes;
[0082] Unlike traditional evidence-based deep learning, this invention does not use the KL divergence regularization loss term during the learning process. Traditional evidence-based deep learning does use the KL divergence regularization loss term. The calculation method is as follows:
[0083]
[0084] The KL divergence regularization loss term is used to reduce the generation of incorrectly classified evidence. However, when auxiliary evidence is introduced, if KL divergence regularization loss is used, the optimization process will reduce the loss by not generating auxiliary evidence, causing the super-evidence to degenerate into basic evidence.
[0085] 2.6: Repeat steps 2.1 to 2.5 above until the multi-perspective pancreatic tumor subtype classification network based on super-evidence accumulation is stable and save the network parameters after training.
[0086] Step 3: Predict the classification results of multi-view pancreatic tumor subtype images ( Figure 3 As shown in the figure, the specific implementation process is as follows:
[0087] 3.1 Input the multi-view pancreatic tumor subtype images to be predicted into step 2.6 to train a stable multi-view pancreatic tumor subtype classification network based on super-evidence accumulation, and obtain the predicted classification results.
[0088] Example
[0089] Following the illustrated process and model structure, a set of implementation examples are given.
[0090] During training, such as Figure 4 As shown, it is necessary to input multi-view pancreatic tumor subtype images from the collected dataset into a multi-view pancreatic tumor subtype classification network based on hyperevidence accumulation to model the feature encoding process of the multi-view pancreatic tumor subtype images. The multi-view pancreatic tumor subtype classification network based on hyperevidence accumulation consists of a view encoder, a hyperevidence generation module, a multi-view fusion module, and a hyperevidence projection module.
[0091] First, multi-view pancreatic tumor subtype images are input into a view encoder for feature extraction, yielding features for each viewpoint. Then, these features are input into a hyperevidence generation module to generate hyperevidence and uncertainty for each viewpoint. Next, the hyperevidence and uncertainty for each viewpoint are input into a multi-view fusion module. Fusion weights for each viewpoint are calculated based on their uncertainties. Based on these fusion weights, the hyperevidence from each viewpoint is weighted and summed to obtain fused hyperevidence. This fused hyperevidence is then input into a hyperevidence projection module to obtain fused multinomial evidence. Finally, the stacking loss is calculated using the multi-view pancreatic tumor subtype image classification labels, and backpropagation is performed to update the network parameters. This process is repeated several times until network training is complete.
[0092] During the testing phase, such as Figure 5 The network is input into multi-view images of pancreatic tumor subtypes to obtain predicted classification results.
[0093] Figure 6 This is a schematic diagram showing the results of the data preprocessing process in an embodiment of the present invention.
[0094] Comparing the method of this invention with existing typical methods, considering classification accuracy, AUROC (Area Under the ROC Curve), and F1 score, the results are as follows: Figure 7 As shown in the figure, the method of the present invention is significantly superior to the comparative method in all aspects.
[0095] The above description is merely a description of preferred embodiments of this application and is not intended to limit the scope of this application in any way. Any changes or modifications made by those skilled in the art based on the above-disclosed technical content should be considered as equivalent and valid embodiments and fall within the scope of protection of the technical solution of this application.
Claims
1. A multi-perspective pancreatic tumor subtype classification method based on super-evidence accumulation, characterized in that, Includes the following steps: Step 1: Dataset preparation: Collect multi-view medical images from the hospital, and after classification and tumor location annotation by professional clinicians, construct a multi-view pancreatic tumor subtype image dataset through data preprocessing. Step 2: Construct a multi-perspective pancreatic tumor subtype classification network framework based on super-evidence accumulation, and use the dataset from Step 1 to train and optimize the multi-perspective pancreatic tumor subtype classification network framework based on super-evidence accumulation. Step 3: Using the optimized multi-view pancreatic tumor subtype classification network framework trained in Step 2, classify the images of the multi-view pancreatic tumor subtypes to be predicted; In step 2, the multi-view pancreatic tumor subtype classification network framework based on super-evidence accumulation includes: a view encoder, a super-evidence generation module, a multi-view fusion module, and a super-evidence projection module, wherein: Viewpoint Encoder: The viewpoint encoder takes a single-view image as input and extracts features from the single-view image. The viewpoint encoder uses a ResNet18 network to encode image features. The ResNet18 network consists of an input layer, a residual layer, a global average pooling layer, and a fully connected layer. The input layer consists of one convolutional layer and one pooling layer. The residual layer consists of 16 convolutional layers and eight residual layers. The fully connected layer is used to integrate features from the feature maps output by the previous layer, thereby achieving data classification. Hyper-evidence Generation Module: This module is connected to the perspective encoder and the multi-perspective fusion module. It takes features from the last layer output of the perspective encoder as input and utilizes prior medical knowledge about pancreatic tumor subtypes to generate hyper-evidence and uncertainties for each perspective. The prior medical knowledge is a coarse classification strategy; based on the potential malignancy risk of pancreatic tumors, all pancreatic tumor subtypes can be divided into three categories: high-risk malignancy tumor subtypes, low-risk malignancy tumor subtypes, and non-tumor subtypes. The hyper-evidence generation module consists of neural network activation layers. Multi-view fusion module: The multi-view fusion module is connected to the hyper-evidence generation module. The multi-view fusion module is used to fuse the hyper-evidence generated by each view into fused hyper-evidence. The hyper-evidence and uncertainty of each view output by the hyper-evidence generation module are used as inputs to the multi-view fusion module. The multi-view fusion module generates view weights based on the uncertainty of each input view and performs a weighted summation of the hyper-evidence of each input view to obtain the fused hyper-evidence. Hyper-evidence projection module: The hyper-evidence projection module is connected to the multi-view fusion module and is used to project the fused hyper-evidence into fused multinomial evidence; wherein the fused multinomial opinion only contains basic evidence; since the fused hyper-evidence needs to be classified into pancreatic tumor subtypes, the auxiliary evidence in the fused hyper-evidence will be allocated to the basic evidence in the fused hyper-evidence according to the prior weight; the fused multinomial opinion will be used for pancreatic tumor subtype classification.
2. The multi-perspective pancreatic tumor subtype classification method based on super-evidence accumulation as described in claim 1, characterized in that, In step 1, the specific steps of data preprocessing are as follows: First, three consecutive CT slices are cut based on the largest cross-section of the tumor region, and the slice size is uniformly set to [512, 512]. Then, based on the tumor location annotation, a square region of interest (ROI) is cropped with the pancreatic tumor as the center; overlapping sliding windows are used in the ROI to obtain four sub-views of the pancreatic tumor; the image of the entire ROI region is used as a global view; for multi-view extraction, the ROI, window, and stride size are set to 160, 128, and 32, respectively. Next, the pixel values of the CT image are truncated; Finally, the CT images were normalized using minimum-maximum scaling.
3. The multi-perspective pancreatic tumor subtype classification method based on super-evidence accumulation as described in claim 1, characterized in that, Step 2 is as follows: Step 2.1: Input the multi-view pancreatic tumor subtype images obtained in Step 1 into the view encoder respectively to encode the multi-view medical image features; Step 2.2: Obtain hyper-evidence from various perspectives through the hyper-evidence generation module. The process includes: The hyperevidence generation module accepts the image features generated in step 2.1 as input, generates hyperevidence after passing through an activation function, and calculates the uncertainty; The super-evidence includes primary evidence and ancillary evidence; primary evidence represents the degree of support of image features for each pancreatic tumor subtype, while ancillary evidence represents the degree of support of image features for the broader category of pancreatic tumor subtypes. Step 2.3: Obtain fused hyper-evidence through the multi-view fusion module. The process includes: The multi-view fusion module takes the hyperevidence and uncertainty of each view generated in step 2.2 as input, and uses the uncertainty of each view to generate the fusion weight of each view: the fusion weight obtained based on the uncertainty estimation characterizes the quality of feature extraction of each view; the uncertainty of the view where low-quality noisy data is located is higher, thus generating a lower fusion weight; the uncertainty of the view where high-quality data is located is lower, thus generating a higher fusion weight. Based on the weights of the perspectives, the hyperevidence from each perspective is weighted and summed to obtain the fused hyperevidence: more accurate uncertainty estimation can generate more accurate fusion weights; more accurate fusion weights can reduce the upper bound of the generalization error of the fused classification and ensure the robustness of the classification model in the face of noise. Step 2.4: Obtain fused multinomial evidence through the hyper-evidence projection module. The process includes: The fused hyper-evidence obtained in step 2.3 is input into the hyper-evidence projection module to generate fused multinomial evidence; wherein the projection process is for the auxiliary evidence in the hyper-evidence, and according to the prior weights, the auxiliary evidence is assigned to the basic evidence for the classification of all pancreatic tumor subtypes; Step 2.5: The fusion of multinomial evidence obtained in Step 2.4 and the medical image classification obtained in Step 1 are learned, and the logarithmic loss is used to calculate the loss; by training and optimizing the overall network composed of the above-mentioned perspective encoder, hyper-evidence generation module, multi-view fusion module and hyper-evidence projection module, a stable multi-view pancreatic tumor subtype classification network based on hyper-evidence accumulation is obtained. Step 2.6: Repeat steps 2.1 to 2.5 above until the multi-perspective pancreatic tumor subtype classification network based on super-evidence accumulation is stable and save the network parameters after training.
4. The multi-perspective pancreatic tumor subtype classification method based on super-evidence accumulation as described in claim 3, characterized in that, Step 2.1 specifically involves randomly selecting samples from the dataset. Multi-view images of pancreatic tumor subtypes The input view encoder extracts features from each viewpoint to obtain the features of each viewpoint. ;in, It is the first Features from a single perspective.
5. The multi-perspective pancreatic tumor subtype classification method based on super-evidence accumulation as described in claim 3, characterized in that, Step 2.2 specifically involves: extracting features from each perspective. Input hyperevidence generation module, where the neural network activation layer is Obtain super-evidence from various perspectives And the uncertainty of various perspectives ;in, It is the first Evidence from multiple perspectives It is the first Uncertainty from one perspective; No. Hyperevidence from a Different Perspective Based on basic evidence and supporting evidence The composition includes: primary evidence, which represents the degree of support of image features for each pancreatic tumor subtype; and secondary evidence, which represents the degree of support of image features for three major pancreatic tumor subtype categories derived from prior medical knowledge. The two types of evidence for each perspective, along with the uncertainty calculation formula, are as follows: Where K represents the number of categories, that is, the number of classifications of pancreatic tumor subtypes.
6. The multi-perspective pancreatic tumor subtype classification method based on super-evidence accumulation as described in claim 3, characterized in that, Step 2.3 specifically involves: obtaining fusion hyper-evidence. The process includes: Uncertainty of each perspective obtained in step 2.2 The fusion weights for each viewpoint are calculated. The calculation formula is: in The hyper-evidence obtained from various perspectives in section 2.2 The fusion super-evidence is obtained by weighted summation based on the fusion weights of each perspective. The calculation formula is: 。 7. The multi-perspective pancreatic tumor subtype classification method based on super-evidence accumulation as described in claim 3, characterized in that, Step 2.4 specifically involves obtaining fused polynomial evidence. The process includes: The fused hyper-evidence obtained in step 2.3 The fusion polynomial evidence was calculated based on the evidence projection formula. The projection formula is: in It is a size of The vector, It is a size of Matrix; Represents the prior weight, which is a function of size . The matrix is given by K, where K represents the number of pancreatic tumor subtypes; for the pancreatic tumor subtype classification task, The evidence from the CCP comprises three major categories and K minor categories; therefore, when setting the projection weights, assume that the three major categories each contain x, y, and z minor categories, where x, y, and z should satisfy... , Set it to the following format: Because it requires extra-evidence The process of obtaining classification results and fusion projection involves fusing hyperevidence based on prior weights. The auxiliary evidence is assigned to the basic evidence; the fused polynomial evidence obtained after projection It contains only basic evidence.
8. The multi-perspective pancreatic tumor subtype classification method based on super-evidence accumulation as described in claim 3, characterized in that, Step 2.5 specifically involves obtaining the fused polynomial evidence from step 2.
4. Compared with the data classification labels in the dataset of step 1 Log loss was calculated to train a multi-perspective pancreatic tumor subtype classification network based on super-evidence accumulation; The logarithmic loss is: ， in This indicates the number of pancreatic tumor subtypes; express The The component, i.e., the first component Fusion polynomial evidence for several pancreatic tumor subtypes; Indicates the sample number One-hot classification labels for pancreatic tumor subtypes.