Immunofluorescence image-based IgA nephropathy prognosis method, system and device

Abstract

The invention discloses an IgA nephropathy prognosis method, system and device based on an immunofluorescence image, and belongs to the technical field of image data processing, and the prognosis method comprises the following steps: obtaining a recognition model and a pathological model; extracting pathological features from the immunofluorescence image through the recognition model; predicting pathological characteristics through a pathological model to obtain the prognosis condition of the IgA nephropathy; the pathological feature extraction method comprises the following steps: segmenting an immunofluorescence image into a plurality of small blocks; establishing labels for the small blocks to obtain a second training set; based on a machine learning method, training the second training set to obtain an identification model; predicting the probability and the label of the small block based on the identification model; and according to the probability and the label, constructing a pathological feature. On the basis of a computer vision method, pathological features are extracted from an immunofluorescence image, and the prognosis of the IgA nephropathy is predicted in combination with a machine learning method, so that automatic prediction is facilitated, and mistakes and omissions caused by artificial naked eye recognition are avoided.

Description

Technical Field

[0001] The present invention relates to the technical field of image data processing, and particularly to a method, system and device for predicting the prognosis of IgA nephropathy based on immunofluorescence images. Background Art

[0002] IgA nephropathy (IgAN) is characterized by the presence of dominant or codominant IgA deposits in the glomerular mesangium. It is considered a primary glomerular disease and is the most common form of primary glomerulonephritis. Studies have shown that the deposition of IgA in the mesangial region and glomerular capillary loops is associated with more severe clinical and histopathological manifestations.

[0003] The core pathological hallmark of IgA nephropathy is the deposition of immune complexes mainly composed of IgA in the glomerular mesangium. This deposition usually shows a granular or clumpy diffuse distribution, and in some cases, it can extend along the capillary loops. Immunofluorescence examination can directly display the localization of IgA in the glomerulus through specific antibody labeling, which is an essential means for diagnosing this disease. The IgA immunofluorescence examination can directly display the localization of IgA in the glomerulus through specific antibody labeling, which is an essential means for diagnosing this disease. The IgA immunofluorescence image can show features such as the deposition site, brightness, and shape. Currently, these features are extracted from the images by the naked eye of doctors. However, there are problems such as subjectivity in reading the images, heterogeneity among different doctors, and key information may be missed by the naked eye. Summary of the Invention

[0004] In view of the above technical problems existing in the prior art, the present invention provides a method, system and device for predicting the prognosis of IgA nephropathy based on immunofluorescence images, extracts pathological features based on IgA immunofluorescence images, establishes a pathological model according to the pathological features, and predicts the prognosis of IgA nephropathy through the pathological model.

[0005] The present invention discloses a method for predicting the prognosis of IgA nephropathy based on immunofluorescence images, which includes the following steps: obtaining an identification model and a pathological model; identifying the immunofluorescence image through the identification model to extract pathological features; predicting the pathological features through the pathological model to obtain the prognosis of IgA nephropathy; wherein, the method for extracting pathological features includes: dividing the immunofluorescence image into multiple small pieces; establishing labels for the small pieces to obtain a second training set; training the second training set based on machine learning methods to obtain an identification model; predicting the probability and label of the small pieces based on the identification model; constructing pathological features according to the probability and label.

[0006] Preferably, the prediction method of the pathological model includes: Training the training set according to the pathological features based on machine learning methods to obtain a pathological model, and the pathological model is used to predict the prognosis of IgA nephropathy.

[0007] Preferably, the specific method for constructing pathological features includes the following steps: Based on the recognition model, predict the probability and label of small patches, and construct the third feature; Combined with histogram statistics, bin the third small patches to obtain bins; According to the probability and label frequency of the small patches in the bins, construct the fourth feature; Vectorize the third feature based on the bag-of-words model, and obtain the fifth feature based on the TF-IDF transformation; Fuse the fourth feature and the fifth feature to obtain the pathological feature.

[0008] Preferably, a more specific method for extracting pathological features includes: Preprocess the IgA immunofluorescence image; Cut the preprocessed IgA immunofluorescence image into multiple small patches; Screen the effective small patches from the multiple small patches; Normalize the effective small patches to obtain the third small patches; Establish labels for the third small patches to obtain the second training set; Based on machine learning methods, train the second training set to obtain a recognition model; Based on the recognition model, predict the probability of the third small patches; According to the prediction threshold and probability, obtain the label of the third small patches, and construct the third feature according to the probability and label; Combined with histogram statistics, obtain the probability distribution of the third small patches; Bin the third small patches to obtain bins; After calculating the probability and label frequency of the third small patches in the bins, perform normalization processing to construct the fourth feature of the bins; Based on the bag-of-words model, vectorize the third feature, and then based on the TF-IDF transformation, obtain the fifth feature; Fuse the fourth feature and the fifth feature to obtain the sixth feature vector and feature set; Screen the sixth feature vector of the feature set to obtain the pathological feature.

[0009] Preferably, the pathological feature or the sixth feature vector is expressed as:

[0010] Among them, is the feature connector, features fusion is expressed as the sixth feature vector or the pathological feature, Histo Prob and Histo Pred are the parameters of the fourth feature,Bow prob and Bow pred is the parameter of the fifth feature.

[0011] Preferably, the method for screening valid small blocks from multiple small blocks includes: Judging that the proportion of bright pixels in the small block exceeds the first threshold; If so, exclude the small block; If not, retain the small block.

[0012] Preferably, the method for screening the sixth feature vector of the feature set includes: The method for screening the sixth feature vector includes: Based on the correlation-based filtering method, retain the sixth feature vector whose Pearson correlation coefficient exceeds the fourth threshold; Use univariate Cox regression and retain the sixth feature vector whose p-value is lower than the fifth threshold; Select non-zero features to obtain pathological features.

[0013] Preferably, the machine learning method for training the recognition model includes any of the following models: ResNet18, ResNet50, Inception_v3, DenseNet121, and VGG19.

[0014] Preferably, the pathological module includes: Cox proportional hazards model or nomogram model; Based on the Min-Max normalization processing method, normalize the probability in the bucket and the frequency of the label.

[0015] The present invention also provides a system for implementing the above method, including a pathological feature recognition module and a prognosis module. The pathological feature recognition module is used to predict the immunofluorescence image through the recognition model and extract pathological features; the prognosis module is used to predict the pathological features through the pathological model to obtain the prognosis of IgA nephropathy.

[0016] The present invention also provides a device, including a processor and a memory, The memory stores code, and when the code is processed by the processor, the above-mentioned IgA nephropathy prognosis method is implemented.

[0017] Compared with the prior art, the beneficial effects of the present invention are: Based on the computer vision method, pathological features are extracted from IgA immunofluorescence images, and combined with the machine learning method to predict the prognosis of IgA nephropathy, which is conducive to automated prediction, avoids errors and omissions caused by manual visual recognition, and shows a robust prediction ability. Brief Description of the Drawings

[0018] Figure 1 It is the flow chart of the IgA nephropathy prognosis method based on multi-modal data of the present invention; Figure 2 It is the flow chart of the method for extracting pathological features; Figure 3 It is the ROC curve of the training set of the recognition model; Figure 4 It is the ROC curve of the test set of the recognition model; Figure 5 It is the KM curve graph of the training set of the pathological model; Figure 6 It is the KM curve graph of the test set of the pathological model; Figure 7 It is the display graph of Grad-CAM activation; Figure 8 It is the structural diagram of the COX univariate regression analysis of clinical features; Figure 9 It is the KM curve graph of the training set of the first model; Figure 10 It is the KM curve graph of the test set of the first model; Figure 11 It is the KM curve graph of the training set of the prognosis model; Figure 12 It is the KM curve graph of the test set of the prognosis model; Figure 13 It is the nomogram model of survival risk; Figure 14 It is the system logic block diagram of the present invention. Detailed implementation manners

[0019] To make the objectives, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are some but not all of the embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the scope of protection of the present invention.

[0020] The following further describes the present invention in detail with reference to the accompanying drawings: Overview: In IgA nephropathy (IgAN), IgA with or without C3 deposits in the glomerular mesangial region under immunofluorescence. The pathological changes under light microscopy are very extensive, and various types can occur, ranging from minimal change to crescentic nephritis (minimal change type, endocapillary proliferative type, FSGS type, membranoproliferative type, crescentic type, proliferative sclerosing type, combined with membranous nephropathy, combined with MCD, etc.). The most common pathological type is the proliferation of mesangial cells and increased matrix in the mesangial region, often accompanied by segmental sclerosis, crescents of varying sizes, and endothelial cell proliferation. The currently popular classification is the Oxford classification MEST proposed in 2009, which evaluates the pathological severity through M (mesangial cell proliferation), E (endothelial cell proliferation), S (segmental sclerosis), and T (the extent of renal cortical tubulointerstitial injury). This classification method mainly relies on subjective naked-eye identification.

[0021] The pathogenesis of immunoglobulin A (IgA) nephropathy is the production of galactose-deficient IgA1 (Gd-IgA1) antibodies. Gd-IgA1 forms immune complexes that deposit in the renal mesangium and cause renal injury. Gd-IgA1 not only has a tendency to self-aggregate but also can amplify its pathogenic effect of renal deposition by abnormally binding to serum proteins such as fibronectin. The deposition of IgA in the mesangial region and glomerular capillary loops is associated with more severe clinical and histopathological manifestations.

[0022] The first aspect of the present invention provides a method for predicting the prognosis of IgA nephropathy based on immunofluorescence images, as Figure 1 shown, including the following steps: Step 101: Cut the immunofluorescence image into multiple small pieces.

[0023] Step 102: Establish labels for the small pieces to obtain a second training set.

[0024] Step 103: Based on a machine learning method, train the second training set to obtain an identification model.

[0025] Step 104: Predict the probability and its label of the small pieces through the identification model.

[0026] Step 105: Construct pathological features according to the probability and its label.

[0027] Step 106: Based on a machine learning method, train the training set according to the pathological features to obtain a pathological model, which is used to predict the prognosis of IgA nephropathy.

[0028] Visual recognition is performed through an identification model, and visual features are converted into pathological features, which is conducive to automated prediction, avoids errors and omissions caused by manual visual recognition, demonstrates robust prediction ability, realizes multi-modal data fusion, and multi-modal data fusion shows a gain effect on long-term prognosis prediction.

[0029] In a specific embodiment, historical data of patients diagnosed with IgA nephropathy by renal biopsy in the First Medical Center of a certain hospital and with a follow-up time of more than 6 months were received. A total of 488 subjects were included and randomly divided into a training set (n = 341) and a test set (n = 147). The endpoint event of the prognosis was set as end-stage renal disease or a 50% decrease in eGFR. The collection of historical data was approved by the medical committee and the patients.

[0030] Such as Figure 2 , the specific method for extracting pathological features includes: Step 201: Preprocess the IgA immunofluorescence image.

[0031] Preprocessing can adopt data augmentation and normalization, etc. In the image preparation stage, we applied Z-score normalization to the RGB channels to standardize the intensity distribution. During the training process, online data augmentation techniques were used, including random cropping and horizontal and vertical flipping, to increase the variability of the data. In model testing, only normalization can be adopted for small patches to ensure the consistency of data processing methods.

[0032] Step 202: Cut the preprocessed IgA immunofluorescence image into multiple small patches.

[0033] In a specific test, the size of the small patches is 512x512 pixels, and its large size can be processed at a magnification of 20 times.

[0034] Step 203: Screen out valid small patches from the multiple small patches.

[0035] For example, small patches mainly filled with bright pixels are excluded. The proportion of bright pixels in these small patches exceeds the first threshold and contains the least useful data. This can be achieved by eliminating all black backgrounds.

[0036] Step 204: Perform normalization processing on the valid small patches to obtain the third small patches.

[0037] Step 205: Establish labels for the third small patches Patch prob , and obtain the second training set.

[0038] Specifically, according to the situation of end-stage renal disease or the second threshold of eGFR decline (such as 50%), the IgA immunofluorescence images and patients are divided into high-risk and low-risk groups, and the corresponding small patches / third small patches are labeled with high-risk and low-risk group categories. However, the grouping method is not limited to this. The X-tile software can also be used to determine the optimal survival cut-off point, and the patients are divided into high-risk / low-risk groups, and the grouping labels can be marked on the corresponding small patches.

[0039] Step 206: Based on the machine learning method, train the second training set to obtain an identification model.

[0040] Specifically, the identification models include ResNet18, ResNet50, Inception_v3, and DenseNet121. The ResNet 18 model is preferred. The performance of each model in the training set is shown in Table 1, and the performance in the test set is shown in Table 2. Among them, DenseNet121 is the most complex, and ResNet50 is a widely used convolutional neural network model.

[0041] To ensure robust performance in different groups, transfer learning is adopted. The model parameters are initialized with the weights pre-trained on the ImageNet dataset, enabling the model to utilize general visual features. To promote better generalization ability, a cosine annealing learning rate scheduling strategy is used, which is defined as: η t = η min + 0.5 × (η max - η min ) × (1 + cos(π × T cur / T i )) Among them, η t is the current learning rate, η min represents the minimum learning rate, η max is the maximum learning rate, T cur is the current iteration number, T i corresponds to the total number of iteration cycles, such as taking 16.

[0042] The training process is optimized using Stochastic Gradient Descent (SGD), and the softmax cross-entropy is used as the loss function to improve the efficiency and performance of the model.

[0043] Table 1

[0044] Table 2

[0045] Among them, Acc (Accuracy): refers to the proportion of correct predictions among all prediction results. AUC (Area Under the Curve): is usually used to evaluate the performance of binary classification models and is the area under the ROC curve (Receiver Operating Characteristic Curve). Sensitivity: also known as the true positive rate, refers to the proportion of people actually suffering from the disease who are detected as positive. Specificity: also called the true negative rate, refers to the proportion of people not suffering from the disease who are detected as negative. PPV (Positive Predictive Value): refers to the proportion of people actually suffering from the disease among those with a positive test result. NPV (Negative Predictive Value): refers to the proportion of people not suffering from the disease among those with a negative test result.

[0046] The ROC curve of the training set of the recognition model is shown in Figure 3 , and the ROC curve of the test set is shown in Figure 4 . In the ResNet18 model, the AUC values in the training and test cohorts are 0.610 and 0.617 respectively. These AUC metrics represent the model's ability to effectively distinguish classes, indicating that its performance is at a medium level. The AUC of the training set falls within the 95% confidence interval (CI) of 0.5977 - 0.6213, while the AUC of the test set falls within the 95% CI of 0.5968 - 0.6372, indicating a certain degree of consistency in the model's performance on different datasets, although its discriminative ability is limited.

[0047] The analysis results show that although the ResNet18 model performs acceptably, it is not the highest-performance model available. However, this choice may be influenced by other factors, such as the simplicity of the model, computational efficiency, or generalization ability reflected in training and testing. In applications, emphasis should be placed on enhancing the robustness of the model and improving its discriminative ability, which may be achieved by exploring model adjustments, enhancement techniques, or adding more discriminative features during training.

[0048] After training the recognition model, it can be extended to predict the labels and probabilities of each small block. These probabilities are aggregated through histogram features and bag-of-words (BoW) features to facilitate whole immunofluorescence image-level prediction.

[0049] Step 207: Based on the recognition model, predict the probability of the third small block, and obtain the label and the third feature of the third small block according to the prediction threshold and the probability. Patch pred , such as when predicting the probability that the third small block belongs to the high-risk group, if the probability exceeds the third threshold, then it is determined that the third small block belongs to the high-risk group. The third feature of the third small block includes: probability

[0050] and label Patch pred . Patch prob .

[0051] Step 208: Combine histogram statistics to bin the third small block to obtain bins; obtain the fourth feature according to the probability and label frequency of the third small blocks in the bins.

[0052] More specifically, combine histogram statistics (such as the number of bins = 50, step size = 0.02, but not limited to this) to obtain the probability distribution of the third small block; bin the third small block to obtain bins; after calculating the frequency of the probability and label of the third small blocks in the bins, perform normalization processing to construct the fourth feature of the bins: Histo Prob and Histo Pred .

[0053] The Min-Max normalization method can be used to normalize the frequency of the probability and label in the bins.

[0054] Step 209: Based on the Bag-of-Words model (BOW), vectorize the third feature, and obtain the fifth feature based on the TF-IDF transformation: Bow prob and Bow pred .

[0055] Step 210: Fuse the fourth feature and the fifth feature to obtain the sixth feature vector features fusion and the feature set, that is, the pathological feature.

[0056] The feature set includes the sixth feature vectors of multiple small blocks, and the sixth feature vector features fusion is expressed as:

[0057] where is the feature connector.

[0058] Step 211: Screen the sixth feature vectors of the feature set to obtain the pathological feature / pathological feature pathomics .

[0059] First, based on the correlation-based filtering method, features with Pearson correlation coefficients exceeding the fourth threshold (e.g., 0.9) are retained. Then, univariate Cox regression is used for further selection, and only features with p-values lower than the fifth threshold (e.g., 0.05) are retained. Non-zero features can also be selected.

[0060] In a specific implementation, after normalizing and blocking the IgA immunofluorescence images (512×512 pixels / block), the third feature vectors (including morphological parameters and spatial distribution features) of each image block are extracted through a pre-trained ResNet 18 model. That is, a weak supervision learning strategy is adopted to generate the prediction labels and their corresponding probabilities for each individual small block / third small block, and further two core feature sets are generated: 101 probability features: based on the high-risk / low-risk probability distribution of the image blocks, the proportion of small blocks in each risk level is calculated through histogram binning; 2 prediction label features.

[0061] Step 212: Train the training set according to the machine learning method and pathological features to obtain a pathological model. For example, based on the method of the Cox proportional hazards model (also known as the multivariate Cox model), train the pathological model.

[0062] Such as Figure 5 and Figure 6 , the survival analysis KM (Kaplan - Meier) curve of the pathological model shows that the C-index (Concordance Index, C-Index) of the training set and the test set are 0.757 and 0.755 respectively. The C-index, that is, the concordance index, is used to evaluate the prediction ability of the model.

[0063] Visualization can be achieved based on Grad-CAM, and the gradient-weighted class activation mapping (Grad-CAM) is used to visualize the recognition ability of the deep learning model on various samples. Figure 7 Shows the Grad-CAM activation of the final convolutional layer, identifies the key regions affecting prognosis prediction, and provides insights into the interpretability of the model.

[0064] The clinical features in historical data can be divided into: continuous variables, categorical variables, and Oxford pathological classification variables. The methods of univariate analysis and multivariate analysis can be used to analyze the clinical features. For example, clinical features are screened through Cox regression (P<0.05).

[0065] As shown in Table 3, the comparison of two groups of baselines between the training set and the test set shows that for continuous variables: age, proteinuria, mean arterial pressure (MAP), and baseline estimated glomerular filtration rate (eGFRbaseline) have no significant differences between the training set and the test set (p > 0.05). Specifically, the mean age of all subjects is 35.16 ± 11.01 years old, and the training set and the test set are 35.16 ± 11.19 and 35.18 ± 10.63 respectively (p = 0.645). The overall mean proteinuria level is 1.61 ± 1.53 g / d, and the distributions between the two groups are consistent (p = 0.176).

[0066] For categorical variables: the usage rates of RAS inhibitors (86.27% vs 87.76%, p = 0.63) and immunosuppressants (56.30% vs 51.02%, p = 0.329) are balanced between groups. The normality of continuous variables is verified by the Shapiro-Wilk test, and the t-test / Mann-Whitney U test is used for analysis respectively. The χ² test is used for categorical variables. In a specific embodiment, the analysis is completed based on Python 3.7.12 and scikit-learn 1.0.2, and the model training uses NVIDIA 4090 GPU and PyTorch 1.8.1.

[0067] There are no statistically significant differences in the distributions of each component (M / E / S / T) of the Oxford pathological classification (MEST-C score): the proportion of segmental sclerosis (S1) is 72.13% in all, 73.31% in the training set, and 69.39% in the test set (p = 0.437). The proportion of tubular atrophy (T1-2) is 73.18% in all, and there is no significant deviation between groups (p = 0.143).

[0068] Table 3

[0069] In Table 3, continuous variables are presented as mean ± standard deviation; categorical variables are presented as frequency (percentage). The p-value reflects the differences between the training set and the test set, and *p < 0.05 is considered statistically significant. RAS: renin-angiotensin system inhibitor; MEST-C: Oxford classification (mesangial proliferation M, endothelial proliferation E, segmental sclerosis S, tubular atrophy T, and cellular crescent C).

[0070] The method for screening clinical characteristics can be carried out by using the method based on COX univariate regression analysis or by using the method of Pearson correlation coefficient. See Figure 8, the Oxford classification lesions M, S, T lesions, immunosuppressant use, urinary protein, and mean arterial pressure are related to the prognosis of IgA nephropathy. Among them, the X-tile was used to determine the optimal risk stratification threshold, which was divided into low-risk and high-risk groups.

[0071] Based on a machine learning method, a first model was trained according to clinical characteristics. Such as Figure 9 and Figure 10 The survival analysis KM (Kaplan-Meier) curves of

[0072] The prognostic model was constructed on the basis of integrating immunofluorescence omics features and clinical characteristics. Such as Figure 11 and Figure 12 , and the survival analysis KM curves showed that the C-index (Concordance Index, C-Index) of the training set and the test set of the model was 0.882 and 0.888 respectively. The performance was better than that of the first model based only on clinical characteristics. Among them, the X-tile was used to determine the optimal risk stratification threshold, which was divided into low-risk and high-risk groups.

[0073] Time-dependent survival analysis: Refer to Table 4. In the 5-year survival prediction analysis, the AUC values of different models showed significant differences in predictive efficacy. The first model, also known as the Clinical Model, reached AUC values of 0.884 and 0.927 in the training set and the test set respectively, showing a robust predictive ability; the prognostic model, also known as the Combined Model, by integrating clinical and pathological features, the AUC value was significantly increased to 0.931 (training set) and 0.940 (test set), verifying the advantage of multi-modal data collaboration (p<0.001).

[0074] In the 7-year survival prediction, the AUC values of each model were overall lower than those in the 5-year analysis, but still maintained a consistent pattern: the AUC of the training set of the first model was 0.908, and the test set slightly decreased to 0.871; the AUC of the training set and the test set of the prognostic model were 0.917 and 0.901 respectively, continuing to lead, further confirming the gain effect of cross-modal integration on the prediction of long-term prognosis (ΔAUC=+4.7% vs the clinical model).

[0075] Table 4

[0076] The present invention also constructed nomogram models for 3-year and 5-year survival risks according to pathological features and clinical characteristics. Such as Figure 13, corresponding survival risks can be found according to corresponding features. A nomogram is a graphical tool used to predict or evaluate the probability or value of an event occurring. It integrates multiple related variables for comprehensive analysis.

[0077] The second aspect of the present invention provides a system for implementing the above method, such as Figure 14 , including a pathological feature recognition module 1 and a prognosis module 2, The pathological feature recognition module 1 is used to predict an immunofluorescence image through a recognition model and extract pathological features; The prognosis module 2 is used to predict the pathological features through a pathological model to obtain the prognosis of IgA nephropathy.

[0078] The prognosis module can also be used to predict historical data and IgA immunofluorescence images through a prognosis model or a first model to obtain the prognosis.

[0079] The system also provides a model training module 3, which is used to train a training set based on machine learning methods according to pathological features to obtain a pathological model, and the pathological model is used for prognosis prediction of IgA nephropathy.

[0080] The third aspect of the present invention provides a device, which includes a processor and a memory. The memory stores code, and when the code is processed by the processor, the above prognosis method is implemented.

[0081] Based on the method of computer vision, the present invention extracts pathological features from IgA immunofluorescence images, which can automatically extract pathological features and avoid errors and omissions caused by manual / visual recognition; it also combines clinical features to predict the prognosis of IgA nephropathy, which is conducive to automatic prediction, has a robust prediction ability, realizes multi-modal data fusion, and multi-modal data fusion shows a gain effect on long-term prognosis prediction.

[0082] A multi-modal survival prediction model is constructed through feature fusion, and finally a Cox proportional hazards model is established to complete survival analysis. Experiments prove that this method can effectively capture potential pathological feature patterns. The multi-modal model shows a C-index of 0.888 in the test set, providing a reliable prognosis estimate for medical treatment or patients, and can be used as a reliable tool for individualized medical evaluation. It should be noted that the prognosis method of the present invention is used to provide suggestions for medical staff or patients and does not belong to the method of disease treatment or diagnosis.

[0083] The above are only the preferred embodiments of the present invention and are not used to limit the present invention. For those skilled in the art, the present invention can have various changes and modifications. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included within the protection scope of the present invention.

Claims

1. A method for predicting the prognosis of IgA nephropathy based on immunofluorescence images, characterized in that, It includes the following steps: Obtain an identification model and a pathological model; Predict the immunofluorescence image through the identification model and extract pathological features; Predict the pathological features through the pathological model to obtain the prognosis of IgA nephropathy; Among them, the method for extracting pathological features includes: Cut the immunofluorescence image into multiple small pieces; Establish labels for the small pieces to obtain a second training set; Based on machine learning methods, train the second training set to obtain an identification model; Based on the identification model, predict the probability and label of the small pieces; Construct pathological features according to the probability and the label; 2. The IgA nephropathy prognosis method according to claim 1, wherein, The prediction method of the pathological model includes: Based on machine learning methods, train the training set according to the pathological features to obtain a pathological model, and the pathological model is used to predict the prognosis of IgA nephropathy; The specific method for constructing pathological features includes the following steps: Based on the identification model, predict the probability and label of the small pieces and construct a third feature; Combine histogram statistics, bin the third small piece to obtain bins; Construct a fourth feature according to the probability and label frequency of the small pieces in the bin; Vectorize the third feature based on the bag-of-words model and obtain a fifth feature based on the TF-IDF transformation; Fuse the fourth feature and the fifth feature to obtain pathological features; 3. The IgA nephropathy prognosis method according to claim 1, characterized in that, The specific method for extracting pathological features includes: Preprocess the IgA immunofluorescence image; Cut the preprocessed IgA immunofluorescence image into multiple small pieces; Screen out valid small pieces from the multiple small pieces; Normalize the valid small pieces to obtain a third small piece; Establish labels for the third small piece to obtain a second training set; Based on machine learning methods, train the second training set to obtain an identification model; Based on the identification model, predict the probability of the third small piece; According to the prediction threshold and probability, obtain the label of the third small piece, and construct a third feature according to the probability and the label; Combine histogram statistics to obtain the probability distribution of the third small piece; Bin the third small piece to obtain bins; After calculating the probability and label frequency of the third small pieces in the bin, perform normalization processing to construct the fourth feature of the bin; Based on the bag-of-words model, vectorize the third feature, and then based on the TF-IDF transformation, obtain a fifth feature; Fuse the fourth feature and the fifth feature to obtain a sixth feature vector and a feature set; Screen the sixth feature vector of the feature set to obtain pathological features; 4. The method for predicting the prognosis of IgA nephropathy according to claim 2 or 3, characterized in that, The pathological features or the sixth feature vector are represented as: ； Among them, is a feature connector, features fusion is represented as the sixth eigenvector or pathological feature, Histo Prob and Histo Pred are the parameters of the fourth feature, Bow prob and Bow pred are the parameters of the fifth feature.

5. The IgA nephropathy prognosis method according to claim 3, wherein The method for screening valid small pieces from multiple small pieces includes: Judge whether the proportion of bright pixels in the small piece exceeds the first threshold; If so, exclude the small piece; If not, retain the small piece; 6. The IgA nephropathy prognosis method according to claim 3, characterized in that, The method for screening the sixth feature vector of the feature set includes: The method for screening the sixth feature vector includes: Based on the correlation-based filtering method, retain the sixth feature vector with a Pearson correlation coefficient exceeding the fourth threshold; Use univariate Cox regression to retain the sixth feature vector with a p-value lower than the fifth threshold; Select non-zero features to obtain pathological features; 7. The method for predicting the prognosis of IgA nephropathy according to claim 1, characterized in that, The machine learning methods for training the identification model include any of the following models: ResNet18, ResNet50, Inception_v3, and DenseNet121.

8. The IgA nephropathy prognosis method according to claim 1, characterized in that, The pathological module includes: a Cox proportional hazards model or a nomogram model; Based on the Min-Max normalization method, normalize the probability in the bucket and the frequency of labels.

9. An IgA nephropathy prognosis system based on immunofluorescence images, characterized in that, Used to implement the IgA nephropathy prognosis method described in any one of claims 1-8, including a pathological feature recognition module and a prognosis module. The pathological feature recognition module is used to predict the immunofluorescence image through a recognition model and extract pathological features. The prognosis module is used to predict the pathological features through a pathological model to obtain the prognosis of IgA nephropathy.

10. An IgA nephropathy prognosis device based on immunofluorescence images, characterized in that, Includes a processor and a memory. The memory stores code, which when processed by the processor, implements the IgA nephropathy prognosis method described in any one of claims 1-8.

Images