Ai diagnostic method for igan glomerulomegaly lesion based on pathological image

Abstract

The application discloses an AI diagnosis method for IgAN glomerular proliferative lesions based on pathological images and belongs to the technical field of pathological analysis. The method comprises the following steps: S1, collecting kidney biopsy samples of IgAN patients and obtaining digital pathological images; S2, constructing a cascade deep learning model, including a segmentation model adopting a CISeg framework and a classification model based on linear complexity sequence modeling, and extracting key pathological tissue features; S3, performing quantitative calculation and completing grading based on the extracted features; and S4, applying a trained artificial intelligence model to automatically perform tissue positioning segmentation and pathological degree grading prediction on unknown cases, and generating a standardized diagnosis report. Through the pathological tissue feature extraction and quantitative calculation method, the application realizes intelligent, objective and accurate evaluation of the pathological grading of IgAN glomerular proliferative lesions, and can provide reliable quantitative basis for clinical doctors to judge the severity of diseases and formulate diagnosis and treatment schemes.

Description

Technical Field

[0001] This invention belongs to the field of pathological analysis technology, specifically relating to an AI diagnostic method for IgAN glomerular proliferative lesions based on pathological images. Background Technology

[0002] Immunoglobulin A nephropathy (IgAN) is the most common primary glomerulonephritis in my country. Its clinical course is highly heterogeneous, with approximately 25% of patients progressing to end-stage renal disease within 20 years of diagnosis. Currently, there is no cure for this disease, and clinical diagnosis and prognosis heavily rely on pathological evaluation via renal biopsy. The Oxford classification system has become an internationally recognized standardized assessment tool.

[0003] However, the Oxford classification faces significant challenges in clinical application. This classification system relies heavily on semi-quantitative manual assessment by pathologists to determine key pathological indicators—particularly mesangial proliferation (M) and endocapillary proliferation (E). This highly subjective assessment method leads to significant discrepancies in diagnostic results between different medical institutions, and even among different pathologists within the same institution, severely impacting diagnostic consistency and comparability. Furthermore, traditional methods lack precise quantitative indicators of pathological changes, making it difficult to provide objective data support for individualized treatment and hindering the reliability of multi-center clinical studies and the standardization of treatment protocols.

[0004] With the rapid development of Artificial Intelligence (AI) and digital pathology, automated analysis based on digital pathological images has offered new possibilities for pathological assessment. However, existing technologies still face numerous challenges in addressing the specific task of IgAN pathological assessment. The kidney's complex structure, with its diverse morphologies and blurred boundaries of glomeruli and internal structures (such as capillary loops and mesangial areas), coupled with variations in tissue staining, places extremely high demands on the algorithm's accurate identification, segmentation, and quantitative analysis. Currently, a complete technical solution is lacking that can systematically address everything from tissue localization and substructure segmentation to quantitative analysis and automated typing.

[0005] To address the aforementioned clinical needs and technical bottlenecks, this invention develops an intelligent diagnostic method specifically for glomerular proliferative lesions, based on a large-scale, high-quality IgAN digital pathology slide dataset annotated by pathologists. This method, through the construction of a cascaded deep learning model architecture, achieves automatic quantitative calculation and Oxford typing of mesangial proliferation and intracapillary proliferation, from cortical region localization and glomerular identification and classification to the segmentation of capillary loops and internal substructures. Summary of the Invention

[0006] To overcome the shortcomings of existing technologies, this invention aims to provide an AI-based diagnostic method for IgAN glomerular proliferative lesions based on pathological images, enabling automatic quantitative calculation and Oxford classification of mesangial proliferation and intracapillary proliferation. This will promote the transformation of IgAN pathological assessment from an experience-based subjective model to a quantitative analysis model based on objective metrics, providing a more standardized, quantifiable, and reproducible auxiliary tool for clinical diagnosis and treatment.

[0007] To achieve the above objectives, the present invention may adopt the following specific technical solutions: The AI ​​diagnostic method for IgAN glomerular proliferative lesions based on pathological images includes the following steps: S1. Collect renal biopsy pathological samples from IgAN patients, obtain digital pathological images, and perform standardized preprocessing; S2. Construct a cascaded deep learning model, which includes a segmentation model using the CISeg framework and a classification model based on linear complexity sequence modeling; the segmentation model sequentially locates and segments the cortical region, glomeruli, capillary loops, and internal substructures; the classification model classifies the segmented glomeruli and extracts key pathological tissue features. S3. Based on the information obtained from the segmentation of the glomerulus, the mesangial area within the glomerulus, mesangial cells, capillary lumen, endothelial cells, and whether the glomerulus is sclerotic or empty, the mesangial proliferation M grade and capillary proliferation E grade are completed according to the preset quantitative calculation rules. S4. Output standardized diagnostic data containing grading results.

[0008] Furthermore, the processing of the segmentation model includes: extracting multi-scale feature embeddings of the input pathological image using a shared image encoder; generating spatial and temporal cues from the preliminary masks output by the semantic segmentation head and the instance segmentation head respectively using a spatiotemporal sequence cue encoder; fusing them into cross-task guidance signals through cross-attention; and achieving collaborative optimization of semantic segmentation and instance segmentation through the bidirectional cross-attention mechanism of the multi-task collaborative decoder. The processing steps of the classification model include: using an adaptive patch selection module to score and filter image patch features based on three criteria: relevance, diversity, and uncertainty; inputting the filtered feature sequences into an efficient sequence encoder for contextual modeling; and completing glomerular classification and key pathological feature extraction.

[0009] Furthermore, the preset quantitative calculation rules in step S3 specifically include: In the calculation of mesangial proliferation grading, first, we obtain r = the score of a single mesangial area within a single glomerulus. The scoring rules are: ≤3 mesangial cells in a single mesangial area = 0 points, 4-5 = 1 point, 6-7 = 2 points, ≥8 = 3 points. Then, we calculate R = the average score of the mesangial area within a single glomerulus. Finally, we calculate M = sum(R of effective glomeruli) / number of effective glomeruli. Mesangial proliferation is graded as M0 if M≤0.5 and M1 if M>0.5. In the calculation of capillary proliferation grading, when the number of cells in a single capillary lumen is ≥2, it is judged as cell proliferation. First, the number of capillary lumen with increased cells in a single glomerulus is calculated as p. Then, P is calculated as p / total number of capillary lumen. Finally, E is calculated as sum(P of effective glomeruli) / number of effective glomeruli. Capillary proliferation is graded as E0 if E≤0.5 and E1 if E>0.5.

[0010] Furthermore, the effective glomerulus count must exclude glomerulosclerotic glomeruli, segmental sclerotic glomeruli, peripheral loops, and empty glomeruli, and only when the effective glomerulus count is greater than 8 can the grading be performed.

[0011] Furthermore, the peripheral loop deduction criteria are as follows: after identifying the glomerulus containing the vascular pole, the relevant membrane area is removed along the direction of the vascular pole toward the capillary loop inside the glomerulus, and the removed part is not included in the M value calculation; the number of cells in the capillary lumen is to exclude the red blood cells in the capillary loop lumen.

[0012] Furthermore, the segmentation model employs a two-stage optimization strategy during training. In the first stage, a binary mask is generated and spatial consistency constraints are applied. In the second stage, the sum of Dice loss and cross-entropy loss is used for semantic tasks, and the sum of Dice loss, Focal loss, mean squared error loss, and multi-scale gradient error loss is used for instance tasks.

[0013] Furthermore, the classification of the segmented glomeruli using the classification model in step S2 specifically includes: The segmented glomeruli were divided into empty glomeruli and non-empty glomeruli using a classification model; The non-empty cystic globules were further classified into spherical sclerotic globules, segmental sclerotic globules, and non-sclerotic globules using a classification model. The capillary loops of non-sclerotic globules and segmentally sclerotic globules were localized and segmented using a segmentation model; The globules with capillary loops were segmented into mesangial region, mesangial cells, capillary lumen, and cells within the capillary lumen.

[0014] Furthermore, the classification process of the classification model is as follows: First, using linear complexity sequence modeling, the glomerulus and surrounding images are segmented to obtain Y image blocks; Then, a pre-trained ResNet50 network is used to extract features from each image patch to obtain Y corresponding d-dimensional feature vectors. The feature vectors are scored by the adaptive Patch selection module and a preset number of the most informative features are selected. Finally, the selected feature vectors are used to form an input sequence, which is then fed into an efficient sequence encoder. The sequence representation output by the encoder is subjected to global average pooling, and the probability distribution of each level is output through a linear classification head.

[0015] Compared with the prior art, the present invention has the following advantages: (1) This invention, based on digital pathological image technology and an improved CISeg framework with an adaptive patch selection mechanism, develops a cascaded deep learning diagnostic model specifically for IgAN glomerular proliferative lesions. This method can automatically and precisely segment and quantitatively analyze the cortical region, glomeruli, capillary loops, and internal substructures in renal biopsy pathological images, achieving automated Oxford classification and grading of mesangial proliferation (M) and intracapillary proliferation (E). This provides clinicians with objective and accurate quantitative evidence for judging the severity of IgAN, assessing prognosis, and developing individualized treatment plans, significantly improving the accuracy, objectivity, and consistency of pathological assessment, and has significant clinical application value.

[0016] (2) This invention proposes a complete and systematic technical solution and quantitative calculation rules, covering the entire algorithm chain from effective region screening and glomerular classification to capillary lumen cell counting; by automatically calculating key quantitative indicators such as mesangial cell density and abnormal capillary lumen ratio, and performing M / E grading accordingly, it solves the problems of traditional Oxford classification relying on semi-quantitative scoring and strong subjectivity. This method not only transforms pathological diagnosis from experience-dependent to data-driven, but also ensures the repeatability of the evaluation process and the comparability of results through clear mathematical definitions and calculation standards, providing a reliable tool for the standardized evaluation of multi-center clinical research and treatment plans.

[0017] (3) The cascaded model design of this invention can efficiently handle complex renal tissue structures. Through a hierarchical analysis strategy of localization, segmentation, and classification, it effectively eliminates interference areas such as sclerotic glomeruli and peripheral loops that are not involved in the evaluation, and focuses computational resources on key lesions within the "effective glomeruli," thereby ensuring the pathological rationality and clinical relevance of the final grading results. This method improves the model's relevance and practicality in complex pathological scenarios.

[0018] (4) The present invention is based on a large-scale, high-quality expert-annotated dataset for model training, and the algorithm is highly targeted and stable. Compared with traditional manual assessment, this method avoids the influence of factors such as inter-observer differences and fatigue, and realizes standardized, high-throughput analysis. It can significantly reduce the heavy and repetitive workload of pathologists, and help promote the standardized IgAN pathological assessment process in medical institutions at all levels, thus promoting the development of precision medicine. Attached Figure Description

[0019] Figure 1 This is a flowchart of the method of the present invention; Figure 2 This is a diagram illustrating the segmentation model of the present invention; Figure 3 The diagram illustrates the classification model of the present invention; where (a) is a feature extractor, (b) is an adaptive patch selector, and (c) is a MIL regressor. Figure 4 This is a flowchart of the AI ​​diagnostic system for IgAN glomerular proliferative lesions of the present invention. Figure 5 This is a diagram illustrating the effective region segmentation effect of the present invention; Figure 6 This is a diagram illustrating the effect of glomerular segmentation within the effective area according to the present invention. Figure 7 This is a diagram illustrating the localization and segmentation of the glomerular substructures of the present invention: the mesangial region, mesangial cells, capillary lumen, and cells within the capillary lumen. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0021] This invention aims to provide a multi-task AI model for IgAN glomerular proliferative lesions, achieving a leap from traditional semi-quantitative assessment to fully quantitative analysis. It effectively improves the objectivity, consistency, and reproducibility of pathological assessment, providing a reliable tool for the clinical diagnosis, efficacy evaluation, and prognosis of IgAN glomerular proliferative lesions. Table 1 below shows the grading criteria for IgAN glomerular proliferative lesions.

[0022] Table 1. Grading criteria for IgAN glomerular proliferative lesions

[0023] Specifically, this invention first determines the effective area for glomerular analysis by segmenting the cortical region. Based on this, a segmentation model is used to achieve precise localization and segmentation of the glomeruli. Subsequently, a classification model is used to distinguish between empty and non-empty glomeruli, and non-empty glomeruli are further subdivided into glomerular sclerosis, segmental sclerosis, and non-sclerotic types. For non-sclerotic glomeruli, the localization and segmentation of capillary loops are further performed, and based on this, the mesangial region, mesangial cells, capillary lumen, and cells within the capillary lumen are precisely localized and segmented, thereby completing a systematic and hierarchical analysis of the glomerulus and its internal microstructure.

[0024] like Figure 1 As shown, the AI ​​diagnostic method for IgAN glomerular proliferative lesions based on pathological images includes the following steps: (I) S1. Collect renal biopsy pathological samples from IgAN patients, obtain digital pathological images and perform standardized preprocessing.

[0025] All pathological tissue sections were converted into digital pathological images, and standardized preprocessing such as format unification and color correction was performed to construct an original image database. All data were divided into training and test sets in an 8:2 ratio.

[0026] (ii) S2. Construct a cascaded deep learning model, which includes a segmentation model using the CISeg framework and a classification model based on linear complexity sequence modeling; the segmentation model is used to sequentially locate and segment the cortical region, glomeruli, capillary loops and internal substructures; the classification model is used to classify the segmented glomeruli and extract key pathological tissue features.

[0027] The algorithm implementation steps include: ① segmenting the effective region using a segmentation model; ② segmenting the glomeruli within the effective region using a segmentation model; ③ classifying the segmented glomeruli into empty glomeruli and non-empty glomeruli using a classification model; ④ further classifying non-empty glomeruli into glomerular sclerotic glomeruli, segmental sclerotic glomeruli, and non-sclerotic glomeruli using a classification model; ⑤ segmenting the capillary loops of non-sclerotic glomeruli and segmental sclerotic glomeruli using a segmentation model; and ⑥ segmenting the mesangial region, mesangial cells, capillary lumen, and intracapillary cells of glomeruli with capillary loops. Specifically, the cortical region segmentation in ①, the capillary loop segmentation in ⑤, and the mesangial region and capillary lumen segmentation in ⑥ all use a semantic segmentation head output; while the glomerular segmentation in ② and the mesangial cell and intracapillary cell segmentation in ⑥ all use an instance segmentation head output.

[0028] (1) Segmentation model The segmentation model employs the CISeg framework, a collaborative segmentation model specifically designed for medical images. Its core lies in achieving joint optimization of semantic and instance segmentation through a mechanism of "mutual prompting and guidance." It is particularly suitable for complex scenarios in IgAN pathological images where "regional structures (such as the mesangial area)" and "individual objects (such as cells)" coexist. Its specific structure includes a shared image encoder, a spatio-sequential prompt encoder (SSP-Encoder), and a multi-task collaborative decoder (MTC-Decoder), used for hierarchical, multi-task tissue structure segmentation of the input digital pathological image, as detailed below: ① Shared image encoder CISeg uses a pre-trained Hierarchical Vision Transformer (Hiera-L) as a shared image encoder to extract multi-scale feature embeddings from the input pathological images. Then, preliminary masks are generated using the semantic segmentation header and the instance segmentation header respectively. ,in To adapt to medical images, only a lightweight Adapter module is inserted into the FFN layer for fine-tuning, and the backbone weights are frozen to preserve general visual priors.

[0029] ② Spatiotemporal sequence cue encoder Spatial hint branches are used to create task-specific masks for shared semantic headers or instance header outputs. Processing: First, through 6 convolutions Layer normalization (LayerNorm) and the GELU activation function are used to convert the pixel mask into image patch markers; then... convolution Expanding the channel; finally, through multi-head self-attention. Capture the global spatial layout. The specific formula is as follows: , Output Representative task The spatial structure prior.

[0030] Timing hints branch to share image embedding As input: first, project it linearly. Dimensional Reduction and Re-engineering Convolution (Conv1D) and State-Space Model Modeling long-range dependencies, and finally passing Restoring the dimension. The specific formula is as follows: , Output Represents the global context dynamics of the image.

[0031] The two will achieve cross-attention and Fusion, generating tasks For the task guidance signal The details are as follows: , This signal will serve as a priori space constraint in the MTC-Decoder.

[0032] ③ Multi-task collaborative decoder The multi-task collaborative decoder includes a semantic segmentation head and an instance segmentation head, and achieves mutual guidance and optimization of the two types of tasks through a bidirectional cross-attention mechanism: to utilize the complementary effects between tasks, the tasks... Query First, enhance self-attention, then fuse it through cross-attention. Hints for dual tasks Receive detailed query As shown below: , in , This indicates element-wise addition.

[0033] by Perform inverse cross-attention for keys and values ​​to generate task-specific refined embeddings: , through 1 Convolution and The function generates a distribution and upsamples it to the original input resolution using a pixel decoder, outputting the final segmentation mask.

[0034] ④ Loss function and training strategy The model training employs a two-stage optimization strategy. The first stage generates a binary mask, subject to binary cross-entropy loss. Constraints are imposed, and spatial consistency constraint loss is applied. (Based on KL divergence), to force the outputs of semantic tasks and instance tasks to align spatially. The second stage optimizes the two types of tasks separately: the semantic task loss is... in, This indicates Dice's loss. Represents the cross-entropy loss, and the instance task loss is... ,in Used for difficult case mining Supervised distance graph regression, Constrain multi-scale boundary gradients. The total loss function is: , in, , For spatial consistency constraint loss, The loss is the binary mask loss. The loss for semantic segmentation is the sum of the Dice loss and the cross-entropy loss. The loss for instance segmentation is the sum of Dice loss, Focal loss, mean squared error loss (MSE), and multi-scale gradient error loss (MSGE).

[0035] like Figure 2 As shown, in the pathological image analysis framework, a shared image encoder based on Hierarchical VisionTransformer is used as the basic feature extractor, and a lightweight Adapter module is introduced for medical image domain adaptation. High-order semantic modeling of input pathological slices is achieved through multi-scale feature embedding and spatiotemporal sequence prompting guidance mechanism. At the same time, a multi-task collaborative decoder containing semantic segmentation head and instance segmentation head is used as the segmentation sub-model. Combined with the cross-task guidance signal generated by the spatiotemporal prompting encoder, joint segmentation of regional structures and independent objects in pathological images is completed. In addition, by utilizing the multi-scale feature representation output by the shared encoder, through the collaborative fusion of high, medium and low-level features, the overall morphology of the glomerulus and the microstructures such as mesangial cells are taken into account during the decoding process, ultimately realizing classification support, fine segmentation and multi-scale context parsing of pathological images.

[0036] (2) Classification model The classification model employs a multi-instance learning (MIL) architecture based on linear complexity sequence modeling: first, the glomerulus and surrounding images are segmented to obtain... Each image patch is used to extract features from a pre-trained ResNet50 algorithm, resulting in a feature set. Subsequently, the Adaptive Patch Selection (APS) module was introduced. The APS score is composed of a weighted average of three criteria: ① Relevance measures the confidence of a patch in classifying a target category, and is defined as: , in, For the number of categories, It is the sigmoid function in binary classification. The calculation is the softmax function in multi-class classification. Represents image blocks Category The predicted probability.

[0037] ② The Diversity Score encourages the selection of morphologically dissimilar image patches to ensure comprehensive tissue representativeness. It is calculated as follows: , in, yes The cosine similarity matrix, whose elements , It is an image block of Normalized eigenvectors, This is the total number of image blocks in the package.

[0038] ③ Uncertainty is identified by predictive entropy to identify fuzzy regions, and is expressed as: , in, It is an image block Category The predicted probability, a small constant Used to prevent when The values ​​sometimes become unstable.

[0039] The final selection score will combine the above three criteria with fixed weights: Among them, weight , .

[0040] Select the highest rated The input sequence consists of several patch features. The data is fed into a high-efficiency sequence encoder (selected from bidirectional GRU, bidirectional LSTM, or Mamba state-space model) to obtain a context-coded representation. Subsequently, residual connections and normalization layers are introduced to output... Then, global average pooling is used to obtain the global representation. The final prediction result is output by the linear classification head. ,in, and The weights and biases of the classification heads are respectively used, and the probability distribution of each category can be obtained by applying Softmax.

[0041] (III) S3. Based on the information obtained from the segmentation of the glomerulus, the mesangial area within the glomerulus, mesangial cells, capillary lumen, endothelial cells, and whether the glomerulus is sclerotic or empty, complete the M-grade of mesangial proliferation and the E-grade of intracapillary proliferation according to the preset quantitative calculation rules.

[0042] In the calculation of mesangial proliferation grading, first, we obtain r = the score of a single mesangial area within a single glomerulus. The scoring rules are as follows: ≤3 mesangial cells in a single mesangial area get 0 points, 4-5 get 1 point, 6-7 get 2 points, and ≥8 get 3 points. Then, we calculate R = the average score of the mesangial area within a single glomerulus. Finally, we calculate M = sum(R of effective glomeruli) / number of effective glomeruli. Mesangial proliferation is graded as M0 if M≤0.5 and M1 if M>0.5.

[0043] In the calculation of capillary proliferation grading, when the number of cells in a single capillary lumen is ≥2, it is judged as cell proliferation. First, the number of capillary lumen with increased cells in a single glomerulus is calculated as p. Then, P is calculated as p / total number of capillary lumen. Finally, E is calculated as sum(P of effective glomeruli) / number of effective glomeruli. Capillary proliferation is graded as E0 if E≤0.5 and E1 if E>0.5.

[0044] The effective glomerular count requires excluding glomerulosclerotic glomeruli, segmental sclerotic portions, peripheral loops, and empty glomeruli. Grading is only performed if the effective glomerular count is greater than 8; otherwise, grading is not performed. The peripheral loop deduction criteria are as follows: after identifying the glomerulus containing the vascular pole, the relevant membrane area is removed along the vascular pole towards the capillary loop within the glomerulus. The removed portion is not included in the M-value calculation. The capillary lumen cell count requires excluding red blood cells within the capillary loop lumen.

[0045] (iv) S4. Output standardized diagnostic data containing grading results.

[0046] The fully trained artificial intelligence model is used to automatically perform tissue localization and segmentation and pathological degree classification prediction for unknown cases, and generate standardized diagnostic reports.

[0047] like Figure 3 As shown, in the training phase of this pathological image analysis, the glomerular and surrounding images are first divided into multiple image patches. A pre-trained ResNet50 is used as the basic feature extractor to obtain the feature vectors of each image patch. Subsequently, an Adaptive Patch Selection (APS) module is introduced. This module calculates a comprehensive score for each patch based on three criteria: relevance, diversity, and uncertainty. And select the one with the highest score. Each patch constitutes a simplified input sequence; this sequence is then fed into an efficient sequence encoder (selectable from bidirectional GRU, bidirectional LSTM, or Mamba state-space model) to generate a context-aware representation by modeling long-range dependencies between image patches; after residual connections and layer normalization, the sequence is subjected to global average pooling to obtain a global representation, and the probability distribution of each category is output by a linear classification head. During the inference phase, the system performs the same patch feature extraction and adaptive selection process on the glomerular and surrounding images, and then outputs the final diagnostic category determination through the multi-instance learning (MIL) model.

[0048] like Figure 4 The diagram illustrates in detail the algorithm processing flow employed in this invention. Based on an in-depth analysis of the key points for detecting IgAN glomerular proliferative lesions, this invention constructs a clear algorithmic chain by optimizing core algorithm modules one by one. The process begins with effective region localization and segmentation, and then gradually refines into a progressive task from glomerular localization to substructure analysis: first, effective region localization is completed ( Figure 5 ), to locate the glomerulus within the effective area ( Figure 6 Next, glomerular classification distinguishes between empty and non-empty glomeruli. Then, non-empty glomeruli are classified for glomerular sclerosis. After screening out non-sclerotic and segmentally sclerotic glomeruli, further segmentation of glomerular substructures, including the mesangial area, mesangial cells, capillary lumen, and endothelial cells, is performed. Figure 7 The entire diagram clearly expresses the logical connections and process evolution between modules through differentiated colors and connecting lines, achieving precise analysis from tissue regions to glomerular substructures.

[0049] In summary, this invention provides a reliable tool for the clinical diagnosis, efficacy evaluation, and prognosis of IgAN glomerular proliferative lesions.

[0050] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. An AI-based diagnostic method for IgAN glomerular proliferative lesions based on pathological images, characterized in that, Includes the following steps: S1. Collect renal biopsy pathological samples from IgAN patients, obtain digital pathological images, and perform standardized preprocessing. S2. Construct a cascaded deep learning model, which includes a segmentation model using the CISeg framework and a classification model based on linear complexity sequence modeling; the segmentation model sequentially locates and segments the cortical region, glomeruli, capillary loops, and internal substructures; the classification model classifies the segmented glomeruli and extracts key pathological tissue features. S3. Based on the segmented glomeruli, mesangial area within the glomeruli, mesangial cells, capillary lumen, endothelial cells, and information on whether the glomeruli are sclerotic or empty, complete the M-grade of mesangial proliferation and the E-grade of intracapillary proliferation according to the preset quantitative calculation rules. S4. Output standardized diagnostic data containing grading results.

2. The AI ​​diagnostic method for IgAN glomerular proliferative lesions based on pathological images according to claim 1, characterized in that, The segmentation model's processing includes: extracting multi-scale feature embeddings from the input pathological image using a shared image encoder; generating spatial and temporal cues from the preliminary masks output by the semantic segmentation head and instance segmentation head respectively using a spatiotemporal sequence cue encoder; fusing them into cross-task guidance signals through cross-attention; and achieving collaborative optimization of semantic segmentation and instance segmentation through the bidirectional cross-attention mechanism of a multi-task collaborative decoder. The processing steps of the classification model include: using an adaptive patch selection module to score and filter image patch features based on three criteria: relevance, diversity, and uncertainty; inputting the filtered feature sequences into an efficient sequence encoder for contextual modeling; and completing glomerular classification and key pathological feature extraction.

3. The AI ​​diagnostic method for IgAN glomerular proliferative lesions based on pathological images according to claim 1, characterized in that, The preset quantitative calculation rules in step S3 specifically include: In the calculation of mesangial proliferation grading, first, we obtain r = the score of a single mesangial area within a single glomerulus. The scoring rules are: ≤3 mesangial cells in a single mesangial area = 0 points, 4-5 = 1 point, 6-7 = 2 points, ≥8 = 3 points. Then, we calculate R = the average score of the mesangial area within a single glomerulus. Finally, we calculate M = sum(R of effective glomeruli) / number of effective glomeruli. Mesangial proliferation is graded as M0 if M≤0.5 and M1 if M>0.

5. In the calculation of capillary proliferation grading, when the number of cells in a single capillary lumen is ≥2, it is judged as cell proliferation. First, the number of capillary lumen with increased cells in a single glomerulus is calculated as p. Then, P is calculated as p / total number of capillary lumen. Finally, E is calculated as sum(P of effective glomeruli) / number of effective glomeruli. Capillary proliferation is graded as E0 if E≤0.5 and E1 if E>0.

5.

4. The AI ​​diagnostic method for IgAN glomerular proliferative lesions based on pathological images according to claim 3, characterized in that, The effective glomerulus count must exclude glomerulosclerotic glomeruli, segmental sclerotic glomeruli, peripheral loops, and empty glomeruli, and only when the effective glomerulus count is greater than 8 can grading be performed.

5. The AI ​​diagnostic method for IgAN glomerular proliferative lesions based on pathological images according to claim 4, characterized in that, The peripheral loop deduction criteria are as follows: after identifying the glomerulus containing the vascular pole, the relevant membrane area is removed along the direction of the vascular pole toward the capillary loop inside the glomerulus. The removed part is not included in the M value calculation; the number of cells in the capillary lumen is reduced by removing the red blood cells in the capillary lumen.

6. The AI ​​diagnostic method for IgAN glomerular proliferative lesions based on pathological images according to claim 1, characterized in that, The segmentation model employs a two-stage optimization strategy during training. In the first stage, a binary mask is generated and spatial consistency constraints are applied. In the second stage, the sum of Dice loss and cross-entropy loss is used for semantic tasks, while the sum of Dice loss, Focal loss, mean squared error loss, and multi-scale gradient error loss is used for instance tasks.

7. The AI ​​diagnostic method for IgAN glomerular proliferative lesions based on pathological images according to claim 1, characterized in that, The classification of the segmented glomeruli using the classification model in step S2 specifically includes: The segmented glomeruli were divided into empty glomeruli and non-empty glomeruli using a classification model; The non-empty cystic globules were further classified into spherical sclerotic globules, segmental sclerotic globules, and non-sclerotic globules using a classification model. The capillary loops of non-sclerotic globules and segmentally sclerotic globules were localized and segmented using a segmentation model; The globules with capillary loops were segmented into mesangial region, mesangial cells, capillary lumen, and cells within the capillary lumen.

8. The AI ​​diagnostic method for IgAN glomerular proliferative lesions based on pathological images according to claim 1, characterized in that, The classification process of the classification model is as follows: First, using linear complexity sequence modeling, the glomerulus and surrounding images are segmented to obtain Y image blocks; Then, a pre-trained ResNet50 network is used to extract features from each image patch to obtain Y corresponding d-dimensional feature vectors. The feature vectors are scored by an adaptive Patch selection module and a preset number of the most informative features are selected. Finally, the selected feature vectors are used to form an input sequence, which is then fed into an efficient sequence encoder. The sequence representation output by the encoder is subjected to global average pooling, and the probability distribution of each category is output through a linear classification head.

Images