Deep learning-based glomerular disease pathology classification model construction method

Abstract

Based on the computer artificial intelligence algorithm model, pathologists can accurately distinguish different primary glomerular diseases, and the diagnosis efficiency is improved. The invention provides a glomerular disease pathology classification model construction method based on deep learning. The method comprises the steps of constructing a data set, and dividing a training queue, an internal verification queue and an external verification queue. Constructing a classification model based on artificial intelligence, wherein the classification model comprises a glomerular segmentation model, a glomerular lesion classification model and a nephropathy diagnosis model; the glomerular segmentation model is used for realizing accurate positioning of the glomerulus and extracting a corresponding glomerular region, the glomerular lesion classification model is used for extracting glomerular lesion features, and the nephropathy diagnosis model is used for fusing the glomerular lesion features of all the glomerulus to obtain full-slide features of the primary glomerular disease. Four kinds of primary glomerular nephropathy are distinguished; and training and verifying the classification model by using the data set.

Description

Technical Field

[0001] The present invention belongs to the field of biomedical technology, and in particular relates to a method for constructing a glomerular disease pathology classification model based on deep learning. Background Art

[0002] Glomerulonephritis (GN) remains one of the leading causes of end-stage renal disease, especially in developing countries. Histopathological evaluation of renal biopsy tissue remains an irreplaceable tool for diagnosing GN, thereby guiding treatment and predicting prognosis. Currently, renal pathologists use visual classification of glomerular lesions to diagnose patients with GN, which requires considerable expertise, time, and effort. This manual and subjective approach is often compounded by interobserver variability between different pathologists. In addition, the pathologist workforce is decreasing, which is particularly evident in highly specialized fields such as renal pathology. Therefore, an automated and subjective method to address these challenges is urgently needed for an effective diagnostic tool that can provide reproducible and quantitative assessments.

[0003] In recent years, there has been a growing interest in the use of artificial intelligence (AI)-based pathological image analysis for medical diagnosis. This technology has been shown to be very useful in the detection and diagnosis of various diseases, including breast cancer and lung cancer. High-throughput digitization of renal tissue sections to generate whole-slice images (WSIs) has enabled the effective application of AI. Previous studies have shown that machine learning (ML) or deep learning (DL)-based techniques can facilitate glomerular detection and segmentation in WSIs, and furthermore, deep learning-based analysis tools have been developed for the identification of glomerular lesions in WSIs of patients with IgA nephropathy. However, there are no integrated AI-based diagnostic tools for glomerular segmentation, glomerular lesion classification, and nephropathy classification in the prior art. Summary of the invention

[0004] In view of this, the present invention aims to overcome the shortcomings of the above-mentioned problems in the prior art and proposes a method for constructing a glomerular disease pathology classification model based on deep learning, which is used to automatically diagnose primary glomerulonephritis using digital histological images, helping renal pathologists to accurately distinguish different glomerular diseases, reduce the time required for diagnostic reporting, and greatly improve the repeatability of traditional visual assessment.

[0005] To achieve the above object, the technical solution of the present invention is achieved as follows:

[0006] The first aspect of the present invention provides a method for constructing a glomerular disease pathology classification model based on deep learning, comprising:

[0007] Build the data set and divide it into training cohort, internal and external validation cohort;

[0008] Constructing an artificial intelligence-based classification model, including a glomerular segmentation model, a glomerular lesion classification model, and a kidney disease diagnosis model; the glomerular segmentation model is used to achieve accurate positioning of the glomerulus and extract the corresponding glomerular region, the glomerular lesion classification model is used to extract glomerular lesion features, and the kidney disease diagnosis model is used to fuse the glomerular lesion features of all glomeruli in WSI, obtain the full slide features for the diagnosis of primary glomerular nephropathy, and distinguish four types of primary glomerular nephropathy;

[0009] The dataset is used to train and validate the classification model.

[0010] Furthermore, the constructed data set also includes an external validation cohort.

[0011] Furthermore, the glomerulus segmentation model implementation process is as follows:

[0012] Split the entire slide image into multiple N×N low-resolution image blocks;

[0013] A glomerular segmentation network based on U-Net and Transformer is used to accurately segment the glomerular region within each patch and generate the corresponding glomerular mask map;

[0014] Reconstruct the glomerular masks of individual patches to create a comprehensive glomerular mask for the entire slide image;

[0015] The glomerular mask was mapped back to the original whole-slide image, and a post-processing algorithm based on morphological operations was applied to filter and refine the segmented regions;

[0016] Intact glomerular regions were extracted from whole slide images and these regions were further analyzed and characterized.

[0017] Furthermore, the glomerular lesion classification model includes a backbone network and a head network, the backbone network is used to extract semantic features from pathological plaques, and the head network is used for classification.

[0018] Furthermore, the backbone network includes three modules: Swin transformer block, patch embedding and patch fusion. The backbone network uses image block embedding to encode the input image into blocks to achieve image sequence; the shifted window transformer module performs global modeling to extract semantic features; in the subsequent stage, the image block merging module reduces the feature map resolution and adjusts the number of channels; the shifted window transformer module is applied again to extract semantic features, and a one-dimensional feature vector is obtained by averaging feature pooling.

[0019] Furthermore, the head network includes a fully connected (FC) layer responsible for feature aggregation and classification, adopts three layers of fully connected nodes, and is initialized using Xavier weights. The first two layers have 224 and 28 nodes respectively, and the last layer has 3 and 2 nodes for three-party and two-way tasks respectively; at the end of the network, a softmax function is applied to map the output probability value of each node to the range of [0,1], and ensure that the sum of the probabilities between the nodes is 1.

[0020] Furthermore, the WSI-level diagnostic model implementation process is as follows:

[0021] Each whole slide image contains N glomeruli, and each glomerulus undergoes a glomerular lesion classification network to produce a 13-dimensional probability vector. The three values ​​of this vector represent the probability of predicting the corresponding type of glomerular lesion. By combining the probability values ​​of each category, three 1*N vectors are obtained;

[0022] Using the kernel density estimation curve, 10 feature probability values ​​are uniformly sampled from each of the three 1*N vectors, and these feature probability values ​​representing the glomerular lesion features are connected to form a set of 30 feature probability values. Other glomerular lesion classification networks are used to derive 20 glomerular lesion features. The glomerular features are connected to obtain a total of 50 features, which are called WSI-level features.

[0023] Based on the extracted features, the extreme gradient boosting algorithm was used to classify the types of primary GN, including IgA nephropathy, membranous nephropathy, focal segmental glomerulosclerosis, and minimal change disease.

[0024] The second aspect of the present invention provides a device for constructing a glomerular disease classification model based on deep learning, comprising:

[0025] Data construction unit, used to construct the data set and divide it into training cohort and internal validation cohort;

[0026] A model building unit is used to build a classification model based on artificial intelligence, including a glomerular segmentation model, a glomerular lesion classification model, and a kidney disease diagnosis model; the glomerular segmentation model is used to achieve accurate positioning of the glomerulus and extract the corresponding glomerular area, the glomerular lesion classification model is used to extract glomerular lesion features, and the kidney disease diagnosis model is used to fuse the glomerular lesion features of all glomeruli in WSI, obtain the full slide features of primary glomerular kidney WSI-level diagnosis, and distinguish four primary glomerular nephropathy;

[0027] The model training unit is used to train and verify the classification model using the data set.

[0028] The third aspect of the present invention provides an electronic device, comprising a processor and a memory that is communicatively connected to the processor and is used to store executable instructions of the processor, wherein the processor is used to execute the above-mentioned method for constructing a glomerular disease pathology classification model based on deep learning.

[0029] A fourth aspect of the present invention provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the above-mentioned method for constructing a glomerular disease pathology classification model based on deep learning.

[0030] Compared with the prior art, the method for constructing a glomerular disease pathology classification model based on deep learning described in the present invention has the following advantages:

[0031] The classification model constructed in the present invention can quickly obtain the classification probability of primary glomerular diseases, help pathologists determine the type of disease, and thus improve diagnostic efficiency.

[0032] The present invention can improve the diagnostic consistency among pathologists, reduce the errors caused by differences among observers, and increase the objectivity, accuracy and repeatability of diagnosis. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] The accompanying drawings constituting a part of the present invention are used to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation of the present invention. In the accompanying drawings:

[0034] Figure 1 This is a flow chart of a method for constructing a glomerular disease classification model based on deep learning of the present invention;

[0035] Figure 2 A schematic diagram of a classification model of artificial intelligence of the present invention;

[0036] Figure 3 It is a schematic diagram of the WSI-level diagnostic model of the present invention;

[0037] Figure 4 A schematic diagram of visualization of glomerulus segmentation performance in a whole slide image of the present invention;

[0038] Figure 5 The typical results of the four types of primary glomerular segmentation of the present invention include IgAN, MN, FSGS and MCD;

[0039] Figure 6 The confusion matrix of the glomerular lesions of the present invention is divided into three categories and two categories, three-category classification in the internal validation cohort; (B) three-category classification in the internal validation cohort (standardized); (C) two-category classification in the internal validation cohort; (D) two-category classification in the internal validation cohort (standardized);

[0040] Figure 7 Figure 3 is the confusion matrix for multi-class diagnosis of primary GN in the internal and external validation cohorts of the present invention, including IgAN, MN, FSGS and MCD. (A) Four-class classification of the internal validation group; (B) Four-class classification of the internal validation group (normalized); (C) Four-class classification in the external validation cohort I; (D) Four-class classification in the external validation cohort I (standardized); (E) Four-class classification in the external validation cohort II; (F) Four-class classification in the external validation cohort II (standardized). DETAILED DESCRIPTION

[0041] It should be noted that, in the absence of conflict, the embodiments of the present invention and the features in the embodiments may be combined with each other.

[0042] In the description of the present invention, it should be understood that the terms "center", "longitudinal", "lateral", "up", "down", "front", "back", "left", "right", "vertical", "horizontal", "top", "bottom", "inside", "outside" and the like indicate positions or positional relationships based on the positions or positional relationships shown in the accompanying drawings, and are only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as limiting the present invention. In addition, the terms "first", "second", and the like are only used for descriptive purposes, and cannot be understood as indicating or implying relative importance or implicitly indicating the number of technical features indicated. Thus, features defined as "first", "second", and the like may explicitly or implicitly include one or more of the features. In the description of the present invention, unless otherwise specified, "multiple" means two or more.

[0043] In the description of the present invention, it should be noted that, unless otherwise clearly specified and limited, the terms "installed", "connected", and "connected" should be understood in a broad sense, for example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection, or it can be indirectly connected through an intermediate medium, or it can be the internal communication of two components. For ordinary technicians in this field, the specific meanings of the above terms in the present invention can be understood by specific circumstances.

[0044] The present invention will be described in detail below with reference to the accompanying drawings and in conjunction with embodiments.

[0045] Embodiment 1:

[0046] like Figure 1 As shown, the present invention provides a method for constructing a glomerular disease pathology classification model based on deep learning, comprising:

[0047] Build the dataset and divide it into training cohort and internal validation cohort;

[0048] Constructing an artificial intelligence-based classification model, including a glomerular segmentation model, a glomerular lesion classification model, and a kidney disease diagnosis model; the glomerular segmentation model is used to achieve accurate positioning of the glomerulus and extract the corresponding glomerular region, the glomerular lesion classification model is used to extract glomerular lesion features, and the kidney disease diagnosis model is used to fuse the glomerular lesion features of all glomeruli in WSI, obtain the full slide features of primary glomerular kidney WSI-level diagnosis, and distinguish four primary glomerular nephropathy;

[0049] The dataset is used to train and validate the classification model.

[0050] In the present invention, we retrospectively selected 1547 patients who underwent renal biopsy and were diagnosed with one of the following GN: IgAN (n = 401), MN (n = 389), MCD (n = 399), or FSGS (n = 358), Nanfang Hospital Affiliated to Southern Medical University, IgAN: IgA nephropathy, MN: membranous nephropathy, MCD: minimal change nephropathy, FSGS: focal segmental glomerulosclerosis. Among them, 1235 patients (79.8%) were randomly assigned to the training cohort, and the remaining 312 patients (20.2%) were assigned to the internal validation cohort. At the same time, 2652 patients who underwent renal biopsy and were diagnosed with one of the above glomerular diseases at Huayin Diagnostic Center from 2017 to 2019 were non-selectively included as the external validation cohort. The cohort included 891 IgAN patients, 1356 MN patients, 222 MCD patients, and 183 FSGS patients. In addition, 2483 patients who underwent renal biopsy and were diagnosed with one of the above glomerular diseases at KingMed Diagnostics were included in the study. The cohort included 914 IgAN patients, 1067 MN patients, 392 MCD patients, and 110 FSGS patients. Detailed inclusion and exclusion criteria are as follows Figure 1 shown.

[0051] All participating hospitals used paraffin-embedded renal tissue fixed in formalin or methylcanol solution. Sections 1–2 mm thick were stained with periodic acid–hexamethylenetetramine silver (PASM). All slides were digitized using a whole-slide scanner NanoZoomer HT2 (Hamamatsu Photonics, Hamamatsu, Japan) with a 320 objective or Aperio AT2 (Leica Biosystems, Wetzlar, Germany) with a 320 or 340 objective. Two renal pathologists reviewed the WSI and confirmed the diagnosis of these patients. The WSIs used for training, internal validation, and external validation cohorts were manually annotated as four types of major GN, including IgAN, MN, FSGS, and MCD. In addition, the detailed outlines and categories of glomeruli were thoroughly annotated in the WSIs for training and internal validation cohorts. Glomeruli diagnosed as IgAN, MN, or MCD in the WSI were annotated as IgAN, MN, or MCD lesions. Of note, some glomeruli in IgAN and MN cases were also synchronously annotated as FSGS lesions. In WSIs diagnosed as FSGS, glomeruli were annotated based on whether they exhibited FSGS lesions and were classified as FSGS or MCD lesions. A total of 6682 digital WSIs of PASM-stained sections were used for training (n = 1235) and validation (n = 5135) based on the classification model to distinguish the four most common primary GN (IgAN, MN, FSGS, and MCD).

[0052] In the present invention, in order to realize the diagnosis and analysis of chronic kidney disease using whole slide images (WSI), an artificial intelligence-based classification model was developed for automatic diagnosis of primary glomerulonephropathy. The classification model includes: a glomerular segmentation model, a glomerular lesion classification model and a WSI-level diagnosis model, such as Figure 2 As shown. The present invention developed the model in a training cohort and independently tested it in internal and external validation cohorts. The glomerular segmentation model achieves precise positioning of the glomerulus by identifying the position of the glomerulus in the entire slide image, and then extracts the corresponding glomerular region. The glomerular lesion classification model divides each glomerulus into three categories: IgAN, MN and MCD lesions, and also into two subcategories: FSGS and non-FSGS lesions. The classification model can be used to extract glomerular lesion features. The WSI-level diagnosis model fuses the glomerular lesion features of all glomeruli in the WSI to obtain the full slide features of the WSI-level diagnosis of primary glomerular nephropathy, that is, accurately distinguishing the four primary glomerular nephropathy.

[0053] Specifically, in order to facilitate the segmentation of glomeruli in the entire slide image, the glomerular segmentation model of the present invention proposes a method for accurate glomerular positioning. The complete positioning process consists of five steps. In the first step, the entire slide image is segmented into multiple N×N low-resolution image patches for efficient processing. Next, a glomerular segmentation network based on U-Net and Transformer is used to accurately segment the glomerular region within each patch and generate a corresponding glomerular mask map. In the third step, the glomerular masks of each patch are reconstructed to create a comprehensive glomerular mask for the entire slide image, the glomerular mask map is mapped back to the original entire slide image, and a post-processing algorithm based on morphological operations is applied to effectively filter and refine the segmented region. Finally, the complete glomerular region is extracted from the entire slide image, so that these regions can be further analyzed and characterized. In order to train the DL-based segmentation model, the present invention manually depicts the outline of each glomerulus in the WSI in the training set.

[0054] The segmentation results were quantitatively evaluated using pixel-level and glomerular-level evaluation metrics in an internal validation cohort. F1 score, precision, and recall are reported as “mean ± standard deviation”, while mean absolute error (MAE) and root mean square error (RMSE) are presented as single values.

[0055] Specifically, the glomerular lesion classification model consists of two parts: the backbone network and the head network. The backbone network is responsible for extracting semantic features from pathological plaques, and it utilizes the SwinT-B network without fully connected layers. The backbone network includes three modules: Swin transformer blocks, patch embedding, and patch fusion. It follows a hierarchical structure of four stages to gradually reduce the feature map size and expand the receptive field layer by layer. In the first stage, the input image is encoded into blocks using image patch embedding to achieve image serialization. Then, the shifted window transformer module performs global modeling to extract semantic features. In the subsequent stage, the image patch merging module reduces the feature map resolution and adjusts the number of channels. The shifted window transformer module is applied again to extract semantic features. Finally, a one-dimensional feature vector is obtained by averaging feature pooling, which is then used by the head network for classification. The head network consists of fully connected (FC) layers responsible for feature aggregation and classification. The head network adopts three layers of fully connected nodes and is initialized with Xavier weights. The first two layers have 224 and 28 nodes, respectively, and the last layer has 3 and 2 nodes for three-way and two-way tasks, respectively. At the end of the network, a softmax function is applied to map the output probability value of each node to the range of [0,1] and ensure that the sum of the probabilities between nodes is 1. The class of the glomerulus is based on the node with the highest probability value.

[0056] To train the DL-based glomerular lesion classification model, all glomeruli in a WSI were automatically assigned the disease label of the WSI. For example, if a WSI was diagnosed as IgAN, the label of all glomeruli in that WSI would be IgAN. The same process was applied to MN and MCD. However, each glomerulus in a WSI diagnosed as FSGS was specifically examined to confirm whether the glomerulus exhibited FSGS lesions; if FSGS lesions were not present in the glomerulus, it was labeled as an MCD glomerulus.

[0057] For the three-category glomerulus classification, macroscopic F1 score, macroscopic precision, macroscopic recall, accuracy, and F1 score for IgAN, MN, and MCD / FSGS were used as evaluation metrics, and ROC, AUC, accuracy, precision, sensitivity, specificity, Youden index, and F1 score for the two-category glomerulus classification in the internal validation cohort.

[0058] Specifically, the WSI-level diagnosis model proposes a WSI-level diagnosis method using kernel density estimation, such as Figure 3 As shown. Each whole slide image contains N glomeruli, and each glomerulus undergoes a glomerular lesion classification network to produce a 13-dimensional probability vector. The three values ​​of this vector represent the probability of predicting the corresponding type of glomerular lesion. By combining the probability values ​​of each category, three 1*N vectors are obtained. Using the kernel density estimation curve, 10 feature probability values ​​are uniformly sampled from each of the three 1*N vectors. These feature probability values ​​representing the characteristics of glomerular lesions are then connected to form a set of 30 feature probability values. Similarly, other glomerular lesion classification networks are used to derive 20 glomerular lesion features. The glomerular features are connected to obtain a total of 50 features, which are called WSI-level features. Finally, based on the extracted features, the extreme gradient boosting (XGBoost) algorithm is used to classify the types of primary GN, including IgAN, MN, FSGS, and MCD.

[0059] We used two manually annotated validation datasets to evaluate the accuracy of the DL model in predicting primary GN. The performance of the model was analyzed using metrics such as F1 score, precision, recall, and accuracy.

[0060] Table 1 lists the detailed clinicopathological characteristics of patients in the training group (n = 1235), internal validation group (n = 312), and external validation group (n = 2652). Most of the characteristics were similar and balanced between the different groups. In all three cohorts. Among the 1547 patients included in this study, 918 patients (59.34%) were males, and the median age was 34 years (IQR 23-49) years. Thirty-five (6.40%) patients had diabetes, and the median serum creatinine (Scr) and estimated glomerular filtration rate (eGFR) were 80.00 (IQR 62.00-106.50) μmol / L and 95.10 (IQR 63.07-115.84) ml / min / 1.73m2, respectively. In the training cohort, 286 (23.16%), 320 (25.91%), 319 (25.83%), 310 (25.10%), and 72 (23.08%) patients were pathologically diagnosed with FSGS, IgAN, MCD, and MN, respectively. 81 (25.96%), 80 (25.64%), and 79 (25.32%) patients were included in the internal validation cohort. The median number of glomeruli per patient in each WSI was 26 (IQR 20-32). Among the 2652 patients included in the external validation cohort, the median age was 43 (IQR, 32-54) years, of which 54.8% were male. The present invention trained an artificial intelligence-based diagnostic model for automatic diagnosis of primary glomerulonephritis using whole-slide images, including a glomerular segmentation model, a glomerular lesion classification model, and a WSI-level diagnostic model.

[0061] Table 1

[0062]

[0063]

[0064] The WSI-level glomerular mask map was reconstructed using multiple glomerular-level mask maps generated by the glomerular segmentation model. Typical results of the segmentation of four types of primary glomeruli are shown in Figure 2. Figure 4 and Figure 5 The segmentation results were quantitatively evaluated using pixel-level and glomerular-level evaluation indicators, including F1 score, precision, recall, mean absolute error (MAE), and root mean square error (RMSE), as shown in Tables 2 and 3. Table 2 shows the pixel-level performance of various glomerular segmentation models in the internal validation cohort, and Table 3 shows the glomerular-level performance of various glomerular segmentation models in the internal validation cohort.

[0065] Table 2

[0066]

[0067] Table 3

[0068]

[0069]

[0070] The proposed GloSNet was compared with state-of-the-art segmentation methods, including U-Net, U-Net++, TransUNet, and SwinUNet. In terms of pixel-level metrics, GloSNet demonstrated the best performance, with an F1 score, precision, and recall of 93.01%, 93.87%, and 92.42%, respectively. The performance of the other networks was slightly lower than that of GloSNet. U-Net achieved a high recall of 93.46%, but showed the lowest F1 score and precision, indicating potential over-discrimination, misclassifying non-glomerular areas as glomeruli (Table 2). In terms of glomerular-level metrics, GloSNet achieved the best F1 score, precision, and recall, which increased significantly from 94.56%, 96.55%, and 94.26% to 96.10%, 97.13%, and 95.41%, respectively, compared to the second-best SwinUNet. The MAE and RMSE of GloSNet were only 1.29 and 1.84, indicating that there were few misclassified or missed glomeruli, and the number of glomeruli predicted in WSI was very consistent with the actual number of glomeruli, thus achieving accurate glomerulus identification (Table 3).

[0071] Two glomerular lesion classification models were trained to classify each glomerulus into three categories and two subcategories. The glomerular classification results for the three categories of IgAN, MN, and MCD / FSGS lesions are shown in Figure 2. Figure 6 As shown in Table 4, the macro F1 score, macro precision, macro recall, accuracy and F1 score of IgAN, MN and MCD / FSGS lesions in the internal validation cohort reached 86.43%, 86.04%, 86.89%, 86.80%, 79.26%, 92.31% and 87.71%, respectively. The results of glomerular classification of FSGS and non-FSGS lesions are shown in Table 5, indicating that the ROC AUC, accuracy, precision, sensitivity, specificity, Youden index and F1 score in the internal validation cohort reached 90.49%, 95.54%, 61.61%, 63.41%, 97.54%, 60.96% and 62.50%, respectively. Table 4 shows the performance of the glomerular classification model in distinguishing the three categories of IgAN, MN and FSGS / MCD lesions in the internal validation cohort. Table 5 shows the performance of the glomerular classification model in distinguishing the two categories of FSGS and non-FSGS lesions in the internal validation cohort.

[0072] Table 4

[0073]

[0074] Table 5

[0075]

[0076] The WSI-level diagnostic method extracts glomerular lesion features and generates whole-slide global features for classification of primary GN types, including IgAN, MN, FSGS, and MCD in internal and external validation cohorts. Figure 7 As shown in Table 6, in the internal validation cohort, the macro F1 score, macro precision, macro recall, accuracy, and IgAN, MN, FSGS, and MCD reached 85.48%, 85.48%, 85.52%, 87.12%, 87.12%, 97.47%, 71.43%, and 85.89%, respectively. Among them, MN had the highest recognition accuracy, reaching 97.5%. Only two cases were misidentified as IgAN and MCD. The recognition performance of IgAN and MCD showed similar accuracy, 87.7% and 87.5%, respectively. However, the recognition performance of FSGS was relatively low, at 69.4%, and it was more easily misclassified as MCD and IgAN. In the external validation cohort, the corresponding indicators reached 85.45%, 83.12%, 88.94%, 92.91%, 92.45%, 99.11%, 69.14%, and 81.08%, respectively. The recognition performance of IgAN, MN, and MCD was similar to that of the internal validation set, reaching 88.0%, 98.5%, and 87.8%, respectively. The recognition accuracy of FSGS was as high as 81.4%, but it showed significant confusion with MCD. Table 6 shows the performance of the diagnostic model for distinguishing the four types of primary GN.

[0077] Table 6

[0078]

[0079]

[0080] Considering the challenges associated with differentiating FSGS from MCD, a strategy was implemented to merge the two categories into one, referred to as FSGS&MCD, as shown in Table 7. In the internal validation cohort, the macro F1 score, macro precision, macro recall, accuracy, and F1 score of IgAN, MN, and FSGS&MCD showed commendable performance, reaching 92.33%, 92.26%, 92.41%, 92.31%, 87.12%, 97.47%, and 92.41%, respectively. In the external validation cohort, the corresponding metrics reached 92.22%, 91.21%, 94.11%, 94.57%, 92.45%, 99.11%, and 85.09%, respectively. These results reflect the effectiveness of the proposed method in accurately classifying the main types of GN. Table 7 shows the performance of the diagnostic model for distinguishing the three types of primary GN (optical microscopy).

[0081] Table 7

[0082]

[0083] Embodiment 2:

[0084] A device for constructing a glomerular disease classification model based on deep learning, comprising:

[0085] Data construction unit, used to construct the data set and divide it into training cohort and internal validation cohort;

[0086] A model building unit, for building a classification model based on artificial intelligence, including a glomerular segmentation model, a glomerular lesion classification model, and a WSI-level diagnosis model; the glomerular segmentation model is used to achieve accurate positioning of the glomerulus and extract the corresponding glomerular region, the glomerular lesion classification model is used to extract glomerular lesion features, and the WSI-level diagnosis model is used to fuse the glomerular lesion features of all glomeruli in the WSI, obtain the full-slide features of the WSI-level diagnosis of primary glomerular nephropathy, and distinguish four primary glomerular nephropathy;

[0087] The model training unit is used to train and verify the classification model using the data set.

[0088] Embodiment three:

[0089] An electronic device includes a processor and a memory that is communicatively connected to the processor and is used to store instructions executable by the processor, wherein the processor is used to execute the above-mentioned method for constructing a glomerular disease pathology classification model based on deep learning.

[0090] Embodiment 4:

[0091] A computer-readable storage medium stores a computer program, which, when executed by a processor, implements the above-mentioned method for constructing a glomerular disease pathology classification model based on deep learning.

[0092] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. A method for constructing a glomerular disease pathology classification model based on deep learning, characterized in that: include: Build the data set and divide it into training cohort, internal and external validation cohort; Constructing an artificial intelligence-based classification model, including a glomerular segmentation model, a glomerular lesion classification model, and a kidney disease diagnosis model; the glomerular segmentation model is used to achieve accurate positioning of the glomerulus and extract the corresponding glomerular region, the glomerular lesion classification model is used to extract glomerular lesion features, and the kidney disease diagnosis model is used to fuse the glomerular lesion features of all glomeruli, obtain the full slide features for the diagnosis of primary glomerular nephropathy, and distinguish four types of primary glomerular nephropathy; The dataset is used to train and validate the classification model.

2. The method for constructing a glomerular disease pathology classification model based on deep learning according to claim 1, characterized in that: The construction dataset also included an external validation cohort.

3. The method for constructing a glomerular disease pathology classification model based on deep learning according to claim 1, characterized in that: The glomerular segmentation model implementation process is as follows: Split the entire slide image into multiple N×N low-resolution image blocks; A glomerular segmentation network based on U-Net and Transformer is used to accurately segment the glomerular region within each patch and generate the corresponding glomerular mask map; Reconstruct the glomerular masks of individual patches to create a comprehensive glomerular mask for the entire slide image; The glomerular mask was mapped back to the original whole-slide image, and a post-processing algorithm based on morphological operations was applied to filter and refine the segmented regions; Intact glomerular regions were extracted from whole slide images and these regions were further analyzed and characterized.

4. The method for constructing a glomerular disease pathology classification model based on deep learning according to claim 1, characterized in that: The glomerular lesion classification model includes a backbone network and a head network, the backbone network is used to extract semantic features from pathological plaques, and the head network is used for classification.

5. The method for constructing a glomerular disease pathology classification model based on deep learning according to claim 4, characterized in that: The backbone network includes three modules: Swin transformer block, patch embedding and patch fusion. The backbone network uses image block embedding to encode the input image into blocks, thereby realizing image sequences; The shifted window transformer module performs global modeling to extract semantic features; in the subsequent stage, the image block merging module reduces the feature map resolution and adjusts the number of channels; the shifted window transformer module is applied again to extract semantic features, and a one-dimensional feature vector is obtained by averaging feature pooling.

6. The method for constructing a glomerular disease pathology classification model based on deep learning according to claim 4, characterized in that: The head network includes a fully connected (FC) layer responsible for feature aggregation and classification, adopts three layers of fully connected nodes and is initialized with Xavier weights. The first two layers have 224 and 28 nodes respectively, and the last layer has 3 and 2 nodes for three-way and two-way tasks respectively; at the end of the network, the softmax function is applied to map the output probability value of each node to the range of [0,1], and ensure that the sum of the probabilities between the nodes is 1.

7. The method for constructing a glomerular disease pathology classification model based on deep learning according to claim 1, characterized in that: The WSI-level diagnostic model implementation process is as follows: Each whole slide image contains N glomeruli, and each glomerulus undergoes a glomerular lesion classification network to produce a 13-dimensional probability vector. The three values ​​of this vector represent the probability of predicting the corresponding type of glomerular lesion. By combining the probability values ​​of each category, three 1*N vectors are obtained; Using the kernel density estimation curve, 10 feature probability values ​​are uniformly sampled from each of the three 1*N vectors, and these feature probability values ​​representing the glomerular lesion features are connected to form a set of 30 feature probability values. Other glomerular lesion classification networks are used to derive 20 glomerular lesion features. The glomerular features are connected to obtain a total of 50 features, which are called WSI-level features. Based on the extracted features, the extreme gradient boosting algorithm was used to classify the types of primary GN, including IgA nephropathy, membranous nephropathy, focal segmental glomerulosclerosis, and minimal change nephropathy.

8. A device for constructing a glomerular disease classification model based on deep learning, characterized in that: include: Data construction unit, used to construct the data set and divide it into training cohort and internal validation cohort; A model building unit, used to build a classification model based on artificial intelligence, including a glomerular segmentation model, a glomerular lesion classification model, and a kidney disease diagnosis model; the glomerular segmentation model is used to achieve accurate positioning of the glomerulus and extract the corresponding glomerular area, the glomerular lesion classification model is used to extract glomerular lesion features, and the kidney disease diagnosis model is used to fuse the glomerular lesion features of all glomeruli, obtain the full slide features for the diagnosis of primary glomerular nephropathy, and distinguish four types of primary glomerular nephropathy; The model training unit is used to train and verify the classification model using the data set.

9. An electronic device, comprising a processor and a memory connected to the processor for storing instructions executable by the processor, characterized in that: The processor is used to execute a method for constructing a glomerular disease pathology classification model based on deep learning as described in any one of claims 1 to 7.

10. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, it implements the method for constructing a glomerular disease pathology classification model based on deep learning as described in any one of claims 1 to 7.

Images