IgA nephropathy survival analysis method and analysis system based on weak supervised learning

Abstract

The invention discloses an IgA nephropathy survival analysis method and system based on weak supervised learning, and belongs to the technical field of medical care information, and the method comprises the steps: extracting a label and a probability of a second region of interest from a full-slice image through a first model; obtaining a fusion feature group according to the label and the probability; and analyzing the fused feature group through a prediction model to obtain a survival analysis result. A second region of interest of the full-slice image subjected to PAS dyeing is identified by using cooperation of a MoE-based segmentation model and a gating network, and a label is established for the second region of interest through a weak supervision method, so that a high-risk region can be identified; and establishing a fusion feature group according to the probability and the label of the second region of interest identified by the first model, constructing an analysis model on the basis of the fusion feature group, and carrying out survival analysis on the IgA nephropathy, so that the analysis efficiency is improved.

Description

Technical Field

[0001] This invention relates to the field of healthcare information technology, specifically to a survival analysis method and system for IgA nephropathy based on weakly supervised learning. Background Technology

[0002] IgA nephropathy (IgAN) is characterized by dominant or co-dominant IgA deposition in the glomerular mesangium. It is considered a primary glomerular disease and is the most common form of primary glomerulonephritis. PAS staining (Periodic Acid-Schiffstain) is primarily used histologically to detect carbohydrates in tissues and can detect mild membranoproliferative IgA nephropathy, as well as glomerular capsule adhesions and segmental sclerosis. However, currently, PAS-stained slides or images are mainly interpreted manually, and the results depend on individual experience.

[0003] CN120852265A discloses a method, apparatus, electronic device, and medium for processing multimodal renal pathology images. It uses a first network model for anomaly classification and a second network model for pathological structure identification and labeling. This allows for the identification of multimodal renal pathology images of different kidney disease types and the addition of Oxford classification labels, providing auxiliary analytical data for kidney examinations. However, it has limitations: on the one hand, it cannot pinpoint high-risk areas; on the other hand, the auxiliary analytical data cannot be used for survival analysis. Summary of the Invention

[0004] To address the aforementioned technical problems in the existing technology, this invention provides a survival analysis method and system for IgA nephropathy based on weakly supervised learning, which analyzes PAS staining in IgA nephropathy, identifies high-risk areas, and performs survival analysis.

[0005] The first aspect of this invention discloses a survival analysis method for IgA nephropathy based on weakly supervised learning, comprising the following steps: acquiring PAS-stained whole-slice images; extracting labels and probabilities of a second region of interest from the whole-slice images using a first model; obtaining a fusion feature group based on the labels and probabilities; and analyzing the fusion feature group using a prediction model to obtain survival analysis results.

[0006] Preferably, the training method for the first model includes: The full-slice image is segmented using multiple segmentation models to obtain multiple regions of first interest. Based on a gated network, the multiple first regions of interest are fused to obtain a second region of interest; Determine whether the following first condition is met: whether the patient has end-stage renal disease, or whether the patient's eGFR has decreased beyond the first threshold; If satisfied, the label of the patient's second region of interest will be set to: high-risk group; If not, the label of the patient's second region of interest will be set to: low-risk group; The first model is trained based on the second region of interest and its label.

[0007] Preferably, the segmentation model is selected from: UNet, UNet++, and DeepLabV3; The training methods for the first model were selected from ResNet18, ResNet50, ResNet101, Inception_v3, and DenseNet121.

[0008] Preferably, the method for obtaining the second region of interest includes: The first dynamic weights of multiple regions of interest are obtained by using a lightweight CNN-based gating network. The first dynamic weight is normalized by Softmax to obtain the second weight; The second region of interest is obtained by summing the weighted sum of the first region of interest and the corresponding second weight.

[0009] Preferably, the fused feature group is generated based on multi-instance learning to aggregate probabilities and label values, specifically including the following steps: The probabilities are divided into 10 buckets at 0.1 intervals. The probability and number of labels of the second region of interest in each bucket are calculated to obtain the first feature and the second feature. Construct dictionaries based on the unique values ​​of the first and second features respectively; All features of the second region of interest for each patient are converted into a frequency vector of dictionary elements, and then the third and fourth features are obtained by TF-IDF transformation. The first, second, third, and fourth features are merged into a fused feature group using feature connectors.

[0010] Preferably, based on the probability of the second region of interest, the corresponding positions of the full-slice image are marked to obtain a visualized image.

[0011] Preferably, the prediction model includes a second model, and the training method for the second model includes: The fusion feature set was deduplicated using the Pearson correlation coefficient, and pathological features with a correlation coefficient greater than the second threshold were retained. Significant pathological features with P less than the third threshold were screened using univariate Cox regression. Based on the selected significant pathological features, a second training set and a second test set were constructed, and a second model was trained using the Lasso-Cox method.

[0012] Preferably, the prediction model includes a third model, and the training method for the third model includes: Collect clinical data and screen clinical features; Screening for significant pathological features from fusion feature groups; A fourth training set was constructed based on clinical and significant pathological features. The third model is obtained by training the fourth training set using the linear Cox proportional hazards method.

[0013] Preferred clinical features include age, baseline eGFR, and pathological grade E.

[0014] A second aspect of the present invention provides an analysis system for implementing the above-described IgA nephropathy survival analysis method, comprising an acquisition module, a feature extraction module, and an identification module; The acquisition module is used to acquire PAS-stained whole-section images; The feature extraction module is used to extract the label and probability of the second region of interest from the whole slice image through the first model, and obtain the fused feature group based on the label and probability; The identification module is used to analyze the fused feature group through a prediction model to obtain survival analysis results.

[0015] Compared with existing technologies, the advantages of this invention are as follows: It uses a MoE-based segmentation model and a gated network to identify the second region of interest (ROI) in a PAS-stained whole-slice image, and employs a weakly supervised method to establish labels for the ROI, facilitating the identification of high-risk regions. Based on the probabilities and labels of the ROI identified by the first model, a fusion feature set is established, and an analysis model is constructed on the basis of the fusion feature set to perform survival analysis on IgA nephropathy, improving analysis efficiency and avoiding the influence of human experience on PAS staining identification. Attached Figure Description

[0016] Figure 1 This is a flowchart of the survival analysis method for IgA nephropathy based on weakly supervised learning according to the present invention; Figure 2 This is an example image of a PAC-stained whole slide; Figure 3 This is a Grad-CAM visualization of the probability of recognition by the first model; Figure 4 This is the training ROC curve for a 3-year survival analysis. Figure 5 This is the ROC curve for the 3-year survival analysis test; Figure 6 This is the training ROC curve for a 5-year survival analysis. Figure 7 It is the ROC curve of the 5-year survival analysis test; Figure 8 This is a logic block diagram of the analysis system of the present invention. Detailed Implementation

[0017] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] The present invention will now be described in further detail with reference to the accompanying drawings: The first aspect of this invention provides a survival analysis method for IgA nephropathy based on weakly supervised learning, such as... Figure 1 As shown, it includes the following steps: Step 101: Acquire whole slide images (WSIs) of the patient with PAS staining.

[0019] Step 102: Segment the whole slice image using multiple segmentation models to obtain multiple Regions of Interest (RoIs).

[0020] The segmentation model, also known as an expert model or expert network, can be UNet, UNet++, or DeepLabV3, etc. Step 103: Based on the gating network, the multiple first regions of interest are fused to obtain the second region of interest.

[0021] Step 104: Determine whether the following first condition is met: whether the patient has end-stage renal disease, or whether the patient's eGFR decreases by more than the first threshold, such as 50%.

[0022] If satisfied, proceed to step 105: set the label of the patient's second region of interest to: high-risk group.

[0023] If not satisfied, proceed to step 106: set the label of the patient's second region of interest to: low-risk group.

[0024] Step 107: Train the first model based on the second region of interest and its label.

[0025] Step 108: Identify the label and probability of the second region of interest to be tested using the first model.

[0026] Step 109: Based on the labels and probabilities, construct a fusion feature group and execute step 110 or 112.

[0027] Step 110: Construct a second training set based on the fused feature set; and train a second model using the second training set.

[0028] Step 111: Construct a fourth training set based on the fusion feature set and clinicopathological features; and train the third model using the fourth training set.

[0029] Step 112: Perform IgA nephropathy survival analysis on the patients to be tested using the second or third model.

[0030] By combining a MoE-based segmentation model with a gated network, the second region of interest (ROI) in the whole-slice image of PAS staining is identified. A label is then created for the ROI using a weakly supervised method, which helps to identify high-risk areas. Based on the probabilities and labels of the ROI identified by the first model, a fusion feature set is established. On the basis of the fusion feature set, an analysis model is constructed to perform survival analysis on IgA nephropathy, improving analysis efficiency and avoiding the influence of human experience on PAS staining identification.

[0031] In one specific embodiment, the following steps are included: Step 301: Collect data.

[0032] The first cohort of PAS-stained whole-slice images and corresponding clinical data were randomly divided into training and test sets in a 7:3 ratio. Model training (including hyperparameter optimization) was performed only on the training set, and performance was evaluated based on the test set to avoid data leakage.

[0033] Specifically, the first cohort consisted of 405 patients, who were divided into a high-risk group (189 cases) and a low-risk group (216 cases) based on 3-year survival. The training set included 283 patients (age 35.47±11.74 years, baseline eGFR 83.25±32.06) and the test set included 122 patients (age 36.67±11.34 years, baseline eGFR 81.75±33.41). There were no statistically significant differences in clinical characteristics between the groups (P>0.05).

[0034] Step 302: Image processing.

[0035] After WSIs are digitized with 20x optical magnification, they are segmented using a multi-model collaborative approach involving UNet, UNet++, and DeepLabV3. The resulting output is dynamically fused using a hybrid expert system (MoE) to pinpoint key pathological regions in IgA nephropathy, such as glomerular segmentation, tubular atrophy, and interstitial fibrosis. The segments are divided into 512×512 pixel patches.

[0036] The UNet network consists of an encoder and a decoder. The encoder has four layers: 3×3 convolutional kernels, stride 1, and padding 1; the output channels are 64, 128, 256, and 512 respectively. The decoder has four layers (2×2 transposed convolutional kernels, stride 2), which use symmetrical skip connections to pass shallow detailed features.

[0037] The UNet++ network adds three levels of dense skip connections to the UNet network, fusing features at different scales across layers, such as 16×16, 32×32, and 64×64 resolution feature maps.

[0038] DeepLabV3 network: It uses dilated convolution to expand the receptive field, such as a dilation rate of 6 / 12 / 18, and combines the spatial pyramid pooling ASPP module to fuse multi-scale contextual information.

[0039] The gated network employs a lightweight CNN architecture, consisting of two 3×3 convolutional layers and one fully connected layer. The input is a low-resolution WSI thumbnail (512×512 pixels), and the output consists of the first dynamic weights of the three base models. These first dynamic weights are then normalized using Softmax to obtain the second weights, ensuring that the sum of all the second weights is 1.

[0040] Fusion result calculation: The second region of interest is the weighted sum of the corresponding first regions of interest.

[0041] Preprocessing is performed on the second region of interest, such as noise removal and color normalization. The second region of interest is also referred to as a patch below.

[0042] Noise removal: Patches with a white background greater than 90% were removed. Specifically, RGB values ​​greater than 240 were considered white. The effective patch retention rate reached 82%-88%, significantly higher than the 65%-75% of traditional non-segmentation methods. These patches contain very little pathological information.

[0043] Color standardization: The Macenko method is used to normalize the colors of effective patches, eliminating color deviations caused by different dyeing batches and scanners. After standardization, the mean error of the RGB channels of all patches is less than 5%.

[0044] Step 303: First model training.

[0045] The patient's label is inherited to their patch as a weak supervision label, eliminating the need to set labels for each patch separately.

[0046] The base models are selected from ResNet18, ResNet50, ResNet101, Inception_v3, and DenseNet121. Transfer learning can be used: the model parameters are initialized with ImageNet pre-trained weights, reducing the training data requirements.

[0047] Optimization strategy: The optimizer is stochastic gradient descent (SGD), the loss function is softmax cross-entropy, the learning rate adopts a cosine decay strategy (η_max=0.01, T_i=16 epochs), the batch size is 32, and 5-fold cross-validation is used.

[0048] Data augmentation: The training set patch was randomly cropped (224×224 pixels) and horizontally / vertically flipped (probability 0.5). The test set was only Z-score normalized (RGB channels were normalized separately), and no data augmentation was performed.

[0049] Table 1 shows the comparative data of the first model based on various CNNs. Among them, DenseNet121 achieved an AUC of 0.625 (95% CI: 0.6122–0.6372) and an accuracy (Acc) of 0.656 on the test set, indicating that the weakly supervised label design can effectively capture pathological patterns. ResNet50 achieved an AUC of 0.862 (95% CI: 0.8566–0.8680) on the training set, with Acc=0.748, sensitivity=0.808, specificity=0.740, PPV=0.299, and NPV=0.966; and an AUC of 0.570 (95% CI: 0.5576–0.5832) on the test set, with Acc=0.560, sensitivity=0.539, specificity=0.563, PPV=0.170, and NPV=0.880. The first model can effectively extract patch-level pathological features, providing high-quality input for subsequent feature fusion.

[0050] Table 1

[0051] Among them, Accuracy (Accuracy) refers to the proportion of correct predictions out of all predicted results. AUC (Area Under the Curve) is typically used to evaluate the performance of binary classification models; it is the area under the ROC (Receiver Operating Characteristic Curve). Sensitivity, also known as the true positive rate, refers to the proportion of people who actually have the disease but test positive. Specificity, also called the true negative rate, refers to the proportion of people who actually do not have the disease but test negative. PPV (Positive Predictive Value) refers to the proportion of people who actually have the disease among those who test positive. NPV (Negative Predictive Value) refers to the proportion of people who actually do not have the disease among those who test negative.

[0052] Step 303: Construct fusion feature groups.

[0053] Using the first model, the probabilities of the second region of interest to be tested are identified, and the label value is obtained based on the probability threshold.

[0054] Grad-CAM visualization allows for clear localization of critical pathological regions. See also... Figure 2 and Figure 3 The probability of each second region of interest is set in Figure 3 The images are marked in corresponding locations, with red areas indicating high-probability zones and blue areas indicating low-probability zones. During diagnosis and treatment, this provides doctors with risk assessment information, allowing them to focus on high-probability areas and improve the efficiency of image interpretation.

[0055] Based on Multiple Instance Learning (MIL), probabilities and label values ​​are aggregated to generate fused feature groups. Specifically, derived features are generated separately using histogram statistics and bag-of-words model aggregation methods, and then the derived features are fused. This includes steps 331-333: Step 331: Histogram statistics. The probabilities are divided into 10 buckets at 0.1 intervals; the probability and number of labels of each small block in the bucket are calculated to obtain the first feature Histo_pred and the second feature Histo_prob.

[0056] Step 332: Bag-of-Words (BoW) aggregation. Construct dictionaries with a length of approximately 120-150 based on the unique values ​​of the first feature Histo_pred and the second feature Histo_prob. Convert all patch features for each patient into frequency vectors of dictionary elements, and then perform TF-IDF transformation to obtain the third feature Bow_prob and the fourth feature Bow_pred.

[0057] Step 333: Feature Fusion. Using feature connectors, the first, second, third, and fourth features are fused into a fused feature group, which includes 206 dimensions of pathological features.

[0058] Step 304: Training the second model.

[0059] First, the fused feature group was deduplicated using the Pearson correlation coefficient. Specifically, pathological features with a correlation coefficient greater than the second threshold, such as 0.9, were retained, and a total of 42 redundant features were removed.

[0060] Then, significant pathological features with P < 0.05 were screened using univariate Cox regression, and 38 significant features were retained.

[0061] Finally, based on the selected significant pathological features, a second training set and a second test set were constructed, and a Lasso-Cox model was trained. The regularization parameter λ was determined through 5-fold cross-validation. Fifteen features with non-zero coefficients were extracted from the model, and a multivariate Cox model, referred to below as the pathology group model, was finally constructed. The training set consisted of 283 cases, and the test set consisted of 122 cases.

[0062] Step 305: Clinical model training.

[0063] Clinical data were collected and clinical features were screened: univariate Cox regression was performed on clinical indicators such as age, MAP, baseline eGFR, RAS usage, and Oxford classification to screen clinical features with P < 0.05.

[0064] Based on the selected clinical characteristics, a third training set and a third test set were constructed, and a clinical model was trained, specifically a multivariate Cox proportional hazards model.

[0065] Step 306: Training the third model.

[0066] By integrating clinical features (P<0.05) and fusion feature groups (P<0.05), a fourth training set and a fourth test set were constructed.

[0067] The third model, specifically a linear Cox proportional hazards model, is trained using the fourth training set; it will also be referred to as the fusion model below.

[0068] The features of the fourth training set included 15 significant features (P<0.05) in the pathology model and 3 significant features (age, baseline eGFR, and pathological grade E, P<0.05) in the clinical model. Pathological grade E is the Oxford classification of IgA.

[0069] Step 307: Survival analysis and statistical validation.

[0070] Risk stratification: The optimal risk cutoff value was determined using X-tile software, and patients were divided into high-risk and low-risk groups. The cutoff value is a threshold used to distinguish between high-risk and low-risk groups.

[0071] Model evaluation: The C-index (consistency index) was used to evaluate the model's discriminative ability, and the Kaplan-Meier (KM) curve combined with the log-rank test (P<0.05 was considered significant) was used to evaluate the survival differences between the high-risk and low-risk groups; Statistical tools: All analyses were performed using Python 3.7.12, Onekey 3.3.5, and scikit-learn 1.0.2. Model training relied on an NVIDIA 4090 GPU and PyTorch 1.8.1.

[0072] Table 2 shows a comparison of the consistency indices (C-Index) of each model.

[0073] Table 2

[0074] The second model, the pathology group model, has a training set consistency index of 0.879 and a non-parametric log-rank test P<0.001 for the high and low risk groups of the KM curve; the test set consistency index is 0.870 and the non-parametric log-rank test P<0.01 for the high and low risk groups of the KM curve. This demonstrates that the weakly supervised framework maintains higher feature extraction capabilities while reducing annotation costs and can independently achieve high-precision survival prediction.

[0075] The third model, the fusion model, achieved a C-index of 0.901 on the test set, which is 2.27% higher than the clinical model (0.881) and 3.56% higher than the pathomic model (0.870), validating the synergistic advantages of multimodal fusion.

[0076] See Table 3 and Figures 4-7In the prediction of 3-year and 5-year survival risks, the fusion model still maintains its leading position: 3-year survival prediction: test set AUC 0.925 (95% CI: 0.8642-0.9864), Acc 0.879, with a specificity of 100% and no false positive high-risk predictions; 5-year survival prediction: test set AUC 0.931 (95% CI: 0.8196-1.0000), Acc 0.929, sensitivity 0.936, which can effectively identify long-term high-risk patients.

[0077] Table 3

[0078] The training set AUC of the third model for 3-year survival prediction was 0.961 (95% CI: 0.9328–0.9885), Acc was 0.936, sensitivity was 0.935, and specificity was 100%, with P < 0.001 in the high- and low-risk groups; the test set AUC was 0.925 (95% CI: 0.8642–0.9864), Acc was 0.879, sensitivity was 0.874, and specificity was 100%, with P < 0.01 in the high- and low-risk groups. The fusion model showed no false positives in 3-year survival prediction and could accurately identify short-term high-risk patients.

[0079] The training set AUC of the third model for 5-year survival prediction was 0.921 (95% CI: 0.8745–0.9685), Acc was 0.806, sensitivity was 0.788, and specificity was 100%. The test set AUC was 0.931 (95% CI: 0.8196–1.0000), Acc was 0.929, sensitivity was 0.936, and specificity was 88.9%. P < 0.001 was observed between the high-risk and low-risk groups. The fusion model maintained high accuracy in long-term survival prediction and can assist in the formulation of long-term treatment plans.

[0080] A second aspect of the present invention provides an analytical system for implementing the above-described analytical method for IgA nephropathy, such as... Figure 8 As shown, it includes a data acquisition module 1, a feature extraction module 2, and a recognition module 3; The acquisition module 1 is used to acquire PAS-stained whole-section images and clinical data; The feature extraction module 2 is used to extract the label and probability of the second region of interest from the whole slice image through the first model, and obtain the fused feature group according to the label and probability, and also extract clinical features from clinical data; The identification module 3 is used to analyze the fused feature group through a prediction model to obtain survival analysis results; the prediction model includes a second model or a third model.

[0081] The analysis system further includes a first training module, a second training module, and a third training module. The first training module is used to train a first model; the second training module is used to train a second model; and the third training module is used to train a third model.

[0082] The analysis system also includes a visualization module for generating visualized key pathological regions of PAS based on the labels and probabilities.

[0083] A third aspect of the present invention provides an analysis device, including a processor and a storage medium, wherein the storage medium stores code for implementing the above-described IgA nephropathy survival analysis method, and the processor, when executing the code, implements the IgA nephropathy survival analysis method.

[0084] This invention offers significant clinical advantages. The weakly supervised framework eliminates the need for manual annotation of pathological images, reducing data processing time for a single patient from hours in traditional strongly supervised methods to minutes. The output can be directly mapped to "high / low risk" groupings, intuitively demonstrating the survival differences between the two groups, aligning with clinicians' decision-making habits. Furthermore, Grad-CAM can provide pathological evidence, enhancing clinical trust.

[0085] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for IgA nephropathy survival analysis based on weakly supervised learning, characterized in that, The method comprises the following steps: Collecting PAS-stained whole slide images; Extracting labels and probabilities of the second region of interest from the whole slide images by a first model; Obtaining a fusion feature group according to the labels and probabilities; Analyzing the fusion feature group by a prediction model to obtain a survival analysis result.

2. The IgA nephropathy survival analysis method of claim 1, wherein, The training method of the first model comprises: Segmenting the whole slide images by a plurality of segmentation models to obtain a plurality of first regions of interest; Fusing the plurality of first regions of interest based on a gating network to obtain a second region of interest; Judging whether the following first condition is met: whether the patient has end-stage renal disease or whether the patient's eGFR has decreased by more than a first threshold; If the condition is met, setting the label of the second region of interest of the patient as: high-risk group; If the condition is not met, setting the label of the second region of interest of the patient as: low-risk group; Training the first model according to the second region of interest and its label.

3. The IgA nephropathy survival analysis method of claim 2, wherein, The segmentation model is selected from: UNet, UNet++, and DeepLabV3; The training method of the first model is selected from: ResNet18, ResNet50, ResNet101, Inception_v3, and DenseNet121.

4. The IgA nephropathy survival analysis method of claim 2, wherein, The method for obtaining the second region of interest comprises: Obtaining first dynamic weights of the plurality of first regions of interest by a lightweight CNN gating network; After the first dynamic weights are normalized by Softmax, second weights are obtained; Summing the first regions of interest and the corresponding second weights to obtain the second region of interest.

5. The IgA nephropathy survival analysis method of claim 1, wherein, Based on multi-instance learning, the aggregation probability and label value are used to generate a fusion feature group, which comprises the following steps: The probability is divided into 10 buckets at intervals of 0.1, and the probability and the number of labels of the second region of interest in each bucket are calculated to obtain first and second features; Dictionaries are constructed according to the unique values of the first and second features, respectively; The features of all second regions of interest of each patient are converted into dictionary element frequency vectors, and third and fourth features are obtained by TF-IDF transformation; The first, second, third, and fourth features are fused into a fusion feature group by a feature connector.

6. The IgA nephropathy survival analysis method of claim 1, wherein, According to the probability of the second region of interest, the corresponding positions of the whole slide images are marked to obtain a visualized image.

7. The IgA nephropathy survival analysis method of claim 1, wherein, The prediction model comprises a second model, and the training method of the second model comprises: Removing duplicates from the fusion feature group by a Pearson correlation coefficient, and retaining pathological features with a correlation coefficient greater than a second threshold; Screening significant pathological features with P less than a third threshold by univariate Cox regression; Constructing a second training set and a second test set according to the screened significant pathological features, and training the second model based on the Lasso-Cox method.

8. The IgA nephropathy survival analysis method of claim 1, wherein, The prediction model comprises a third model, and the training method of the third model comprises: Collecting clinical data and screening clinical features; Screening significant pathological features from the fusion feature group; Constructing a fourth training set according to the clinical features and the significant pathological features; Training the fourth training set based on the linear Cox proportional risk method to obtain the third model.

9. The IgA nephropathy survival analysis method of claim 8, wherein, Clinical features include age, baseline eGFR, and pathological grade E.

10. An analysis system, characterized by For realizing the IgA nephropathy survival analysis method as claimed in any one of claims 1-9, the analysis system comprises a collection module, a feature extraction module, and an identification module; The collection module is used for collecting PAS-stained whole section images; The feature extraction module is used for extracting a label and a probability of a second region of interest from the whole section images through a first model, and obtaining a fusion feature group according to the label and the probability; The identification module is used for analyzing the fusion feature group through a prediction model to obtain a survival analysis result.

Images