A Chronic Kidney Disease Pathological Image Recognition Enhancement System

Abstract

The present invention discloses a system for enhancing the recognition of chronic kidney disease pathological images, which integrates data collection, kidney disease pathological modeling, pathological evolution enhancement, and image recognition enhancement modules. The present invention belongs to the technical field of pathological image enhancement, and specifically relates to a system for enhancing the recognition of chronic kidney disease pathological images. Through multi-modal data acquisition, adaptive light compensation, and image registration and reconstruction, a three-dimensional structure model of the renal unit is realized; combined with a three-dimensional spatio-temporal graph convolutional network, spatial structure, temporal dynamics, and clinical features are fused to improve the stability and interpretability of pathological evolution prediction; finally, based on a pre-trained recognition model and a visualization heat map, the accuracy and clinical usability of pathological image recognition are enhanced, effectively improving the intelligent and collaborative level of pathological recognition.

Description

Technical Field

[0001] The present invention belongs to the technical field of pathological image enhancement, and specifically relates to a chronic kidney disease pathological image recognition and enhancement system. Background Art

[0002] The chronic kidney disease pathological image recognition and enhancement system is an auxiliary diagnosis technology that combines artificial intelligence and medical image processing. It can accurately identify and highlight key lesion areas (such as glomerular sclerosis, interstitial fibrosis, etc.) in renal tissue section images through a deep learning model, thereby improving the judgment efficiency and accuracy of pathologists on the progression degree of chronic kidney disease, and plays an important role in assisting decision-making, standardized diagnosis, and reducing the burden of film reading.

[0003] However, in the existing chronic kidney disease pathological image recognition and enhancement systems, there are technical problems as follows: the traditional pathological image sources mainly rely on 2D slices from a single perspective, which are difficult to reflect the three-dimensional structural relationship of nephrons; image enhancement mainly relies on traditional image enhancement algorithms, lacking a deep enhancement mechanism based on the semantics of structure and pathological progression; and because the recognition module, structure modeling, and pathological evolution analysis are independent of each other, it is difficult to form a unified recognition and enhancement process; in the existing kidney disease pathological modeling process, there are technical problems such as insufficient registration accuracy of pathological section images resulting in obvious structural distortion in three-dimensional reconstruction; traditional modeling methods ignore the spatial relationship modeling of microscopic structures such as cell nuclei and nephrons, making it difficult to support subsequent pathological evolution analysis and spatial lesion tracking; in the existing pathological evolution enhancement process, there are technical problems such as the existing models not considering the spatio-temporal continuity of kidney disease evolution, lacking a mechanism to fuse structural changes and time dynamics, and lacking an effective modeling path when fusing clinical indicators; in the existing image recognition and enhancement process, there are technical problems such as the models generally adopting shallow feature extraction strategies, being difficult to obtain discriminative information of key lesion areas, lacking semantic fusion based on structural modeling and pathological evolution, resulting in the recognition task being insensitive to early changes, and at the same time having weak visualization ability, making it difficult to support the interpretable requirements of clinical auxiliary diagnosis. Summary of the Invention

[0004] In view of the above situation, to overcome the defects of the prior art, the present invention provides a system for enhancing the recognition of chronic kidney disease pathological images. In the existing system for enhancing the recognition of chronic kidney disease pathological images, the traditional pathological image sources mainly rely on 2D slices from a single perspective, which are difficult to reflect the three-dimensional structural relationship of nephrons. Image enhancement mainly relies on traditional image enhancement algorithms, lacking a deep enhancement mechanism based on the semantics of structure and pathological progression. Moreover, since the recognition module, structure modeling, and pathological evolution analysis are independent of each other, it is difficult to form a unified recognition and enhancement process. The present solution creatively adopts an overall recognition and enhancement idea that combines nephropathy pathological modeling, pathological evolution enhancement, and image recognition enhancement, achieving multi-level collaboration of structure modeling, evolution prediction, and image recognition, significantly improving the accuracy, stability, and clinical interpretability of pathological recognition. In the existing nephropathy pathological modeling process, there are technical problems such as insufficient registration accuracy of pathological slice images, resulting in obvious structural distortion in three-dimensional reconstruction; traditional modeling methods ignore the spatial relationship modeling of microscopic structures such as cell nuclei and nephrons, making it difficult to support subsequent pathological evolution analysis and spatial lesion tracking. The present solution creatively adopts a renal structure tracking modeling method combined with improved registration for nephropathy pathological modeling, thereby accurately expressing the spatial structure characteristics of nephrons. In the existing pathological evolution enhancement process, there are technical problems such as the existing models not considering the spatio-temporal continuity of nephropathy evolution, lacking a mechanism for fusing structural changes and temporal dynamics, and lacking an effective modeling path when fusing clinical indicators. The present solution creatively adopts a three-dimensional spatio-temporal graph convolutional network combined with non-deterministic quantization and multi-scale feature fusion enhancement for pathological evolution enhancement, achieving dynamic modeling of renal structure over time, improving prediction stability under uncertainty perception enhancement, and embedding and fusion of clinical-image bimodal features. In the existing image recognition enhancement process, there are technical problems such as the models generally adopting shallow feature extraction strategies, being difficult to obtain discriminative information of key lesion areas, lacking semantic fusion based on structure modeling and pathological evolution, resulting in the recognition task being insensitive to early changes, and having weak visualization ability, making it difficult to support the interpretable requirements of clinical auxiliary diagnosis. The present solution creatively adopts a pre-trained model recognition and analysis method for image recognition enhancement, achieving deep recognition that fuses local-global structural features, structure-evolution-recognition collaborative modeling driven by multi-source data, and interpretable image recognition output based on heatmaps and key region localization.

[0005] The technical solution adopted by the present invention is as follows: A system for enhancing the recognition of chronic kidney disease pathological images provided by the present invention includes a data collection module, a nephropathy pathological modeling module, a pathological evolution enhancement module, and an image recognition enhancement module;

[0006] The data collection module is used for data collection and optimization processing. Through data collection and optimization processing, an original dataset for enhanced pathological image recognition is obtained, and the original dataset for enhanced pathological image recognition is sent to the nephropathy pathological modeling module, the pathological evolution enhancement module, and the image recognition enhancement module;

[0007] The nephropathy pathological modeling module is used for nephropathy pathological modeling. Through nephropathy pathological modeling, renal topology modeling data is obtained, and the renal topology modeling data is sent to the pathological evolution enhancement module;

[0008] The pathological evolution enhancement module is used for pathological evolution enhancement. Through pathological evolution enhancement, comprehensive nephropathy pathological feature data is obtained and the comprehensive nephropathy pathological feature data is sent to the image recognition enhancement module;

[0009] The image recognition enhancement module is used for image recognition enhancement. Through image recognition enhancement, enhanced reference data for chronic nephropathy pathological image recognition is obtained.

[0010] Furthermore, the data collection and optimization processing is used for collecting original data and performing optimization processing. Specifically, through multi-source data collection, original chronic nephropathy pathological data is obtained, and through image preprocessing and data alignment, an original dataset for enhanced pathological image recognition is obtained, including the following steps: multi-modal data acquisition, adaptive illumination enhancement, data cleaning and optimization, and overall data optimization;

[0011] The multi-modal data acquisition is specifically to obtain original chronic nephropathy pathological data through multi-source data collection. The original chronic nephropathy pathological data includes original pathological image data, clinical biochemical data, and pathological text data; by sequentially performing time series data alignment, maximum inter-class variance threshold segmentation optimization processing, and manual data annotation on the original chronic nephropathy pathological data, an optimized dataset for chronic nephropathy pathology is obtained;

[0012] The optimized dataset for chronic nephropathy pathology includes maximum inter-class variance threshold segmentation optimized images and time series alignment annotation data;

[0013] The adaptive illumination enhancement is specifically to construct an adaptive illumination compensation U-shaped network based on the optimized dataset for chronic nephropathy pathology, perform color correction and enhancement, and obtain a color correction dataset;

[0014] The adaptive illumination compensation U-shaped network includes an encoder structure, a decoder structure, and an optimized loss function;

[0015] The optimized loss function includes a reconstruction loss and a perceptual loss;

[0016] The reconstruction loss specifically uses the L1 norm and is calculated based on the pixel difference between the input and the output;

[0017] The perceptual loss is specifically obtained by constructing a pre-trained three-dimensional convolutional layer of VGG16, optimizing the images in the dataset according to the chronic kidney disease pathology, extracting perceptual features, and constructing the perceptual loss;

[0018] The data cleaning and optimization is specifically to perform data cleaning based on the color correction dataset through fuzzy slice detection and perform data sampling enhancement through geometric transformation to obtain a cleaned and optimized dataset;

[0019] The fuzzy slice detection calculates the clarity by setting a clarity metric threshold and a clarity metric matrix;

[0020] The geometric transformation includes rotation, scaling, and translation;

[0021] The overall data optimization is specifically to perform overall enhancement of the chronic kidney disease pathology optimized dataset through the adaptive illumination enhancement and the data cleaning and optimization to obtain a pathologically enhanced original dataset for image recognition;

[0022] The pathologically enhanced original dataset for image recognition includes enhanced image data, time-series alignment annotation data, unenhanced image data, and enhanced transformation record data.

[0023] Furthermore, the kidney disease pathology modeling is used to three-dimensionally reconstruct a nephron-level model. Specifically, according to the pathologically enhanced original dataset for image recognition, a kidney structure tracking and modeling method combined with improved registration is used to perform kidney disease pathology modeling to obtain kidney topology modeling data, including the following steps: slice feature matching, transformation matrix optimization, image deformation, nucleus center detection, node graph construction, and kidney three-dimensional reconstruction;

[0024] The slice feature matching is specifically to take every two adjacent pathological slice images in the pathologically enhanced original dataset for image recognition as the original data input, use the scale-invariant feature transform algorithm to extract slice feature points to obtain feature point feature descriptor data, use the standard fast robust binary descriptor algorithm to extract feature vectors based on the feature point feature descriptor data, and obtain candidate feature point matching data through two-way consistent matching;

[0025] The transformation matrix optimization is specifically to introduce the standard thin plate spline transformation algorithm to perform global non-linear deformation on the pathological slice images in the pathologically enhanced original dataset for image recognition, and construct a minimized objective function for improved registration based on the candidate feature point matching data to perform slice registration optimization to obtain slice transformation matrix data;

[0026] The calculation formula of the minimized objective function for the improved registration is:

[0027] ;

[0028] In the formula, F TRA is the minimization objective function for improved registration, T is the slice transformation matrix data output by the standard thin plate spline transformation algorithm, K is the total number of feature points in the candidate feature point matching data, k is the feature point index, FREAK(·) is the standard fast robust binary descriptor algorithm function, I i+1 (·) is the transformed pixel of the (i + 1)-th image, i is the pathological slice image index, T(·) is the standard thin plate spline transformation algorithm function, p k is the position parameter of the k-th feature point, SIFT(·) is the scale-invariant feature transform algorithm function, I i is the transformed pixel of the i-th image, ||·||2 is the L2 norm operator, is the regularization coefficient, is the gradient output by the standard thin plate spline transformation algorithm, ||·|| F is the sum of squares of partial derivatives operator;

[0029] The image deformation is specifically to perform deformation processing on the pathological slice images in the original dataset for pathological image recognition enhancement according to the slice transformation matrix data, so as to obtain three-dimensional reconstruction aligned pathological slice image data;

[0030] The nucleus center detection is specifically to construct a standard deep learning segmentation model, and perform nucleus center detection according to the three-dimensional reconstruction aligned pathological slice image data to obtain nucleus center coordinate detection data;

[0031] The standard deep learning segmentation model specifically uses the pre-trained StarDist deep learning model;

[0032] The node graph construction is specifically to perform graph structure construction according to the nucleus center coordinate detection data to obtain nephron node graph data;

[0033] The renal three-dimensional reconstruction is specifically to perform maximum weighted path search and three-dimensional curve fitting in sequence according to the nephron node graph data, and obtain renal topological modeling data through voxel volume construction;

[0034] The maximum weighted path search specifically uses the standard Dijkstra algorithm for path search;

[0035] The renal topological modeling data includes candidate feature point matching data, three-dimensional voxel modeling data of renal tubules, and nephron node graph data.

[0036] Furthermore, the pathological evolution enhancement is used to dynamically analyze the process of chronic kidney disease and perform feature fusion analysis. Specifically, based on the renal topological modeling data and the pathological images, an enhanced original dataset is identified. A three-dimensional spatio-temporal graph convolutional network combined with non-deterministic quantization and multi-scale feature fusion enhancement is used to perform pathological evolution enhancement, obtaining comprehensive pathological feature data of kidney disease, including the following steps: spatio-temporal graph modeling, dynamic feature extraction, non-deterministic quantization enhancement, multi-scale feature fusion, and pathological evolution enhancement;

[0037] The spatio-temporal graph modeling is specifically to perform temporal graph modeling based on the renal topological modeling data and the nephron node graph data, construct temporal nephron node graph data, and obtain temporal nephron node edge weight data through edge weight calculation;

[0038] The dynamic feature extraction is specifically to construct a three-dimensional graph convolutional network and a temporal convolutional network based on the temporal nephron node graph data and the temporal nephron node edge weight data, perform modeling of the spatial and temporal features of chronic kidney disease, and obtain spatial pathological feature data and temporal pathological feature data of kidney disease;

[0039] The non-deterministic quantization enhancement is specifically to perform pathological progression prediction based on the spatial pathological feature data and the temporal pathological feature data of kidney disease, and perform non-deterministic perception enhancement of pathological progression prediction by constructing an average prediction estimate and a non-deterministic prediction estimate, obtaining quantized enhanced pathological image feature data;

[0040] The multi-scale feature fusion is specifically to extract low-magnification features and high-magnification features from the quantized enhanced pathological image feature data, obtain spliced feature data by upsampling and splicing the low-magnification features, perform serum creatinine value feature embedding by constructing a standard multi-layer perceptron, obtain clinical embedding feature data, and perform feature fusion on the spliced feature data and the clinical embedding feature data to obtain multi-scale fused pathological feature data;

[0041] The calculation formula for the feature fusion is:

[0042] ;

[0043] In the formula, F final is the multi-scale fused pathological feature data, LayerNorm(·) is the layer normalization operation, Conv 1×1 (·) is a 1×1 feature fusion convolutional layer, Upsample(·) is the upsampling operation function, F 20x is the low-magnification feature, F 40x is the high-magnification feature, MLP(·) is the standard multi-layer perceptron function, sCr is the serum creatinine value feature, W cis the feature embedding weight;

[0044] The pathological evolution is enhanced. Specifically, based on the multi-scale fusion pathological feature data, a joint optimization loss function is constructed for model training to obtain a pathological evolution enhancement model, and the pathological evolution enhancement model is used to predict the pathological process to obtain comprehensive nephropathy pathological feature data.

[0045] Furthermore, the image recognition is enhanced to comprehensively recognize nephropathy pathological images by combining fusion features and three-dimensional reconstruction features. Specifically, based on the original dataset for enhanced pathological image recognition, the renal topological modeling data, and the comprehensive nephropathy pathological feature data, a pre-trained model recognition and analysis method is adopted for image recognition enhancement to obtain reference data for enhanced recognition of chronic nephropathy pathological images, including the following steps: multi-feature joint embedding, visualization optimization, and image recognition enhancement;

[0046] The multi-feature joint embedding is specifically to adopt a pre-trained image recognition model based on the original dataset for enhanced pathological image recognition, the renal topological modeling data, and the comprehensive nephropathy pathological feature data to perform multi-scale feature extraction to obtain standardized image feature vector data;

[0047] The pre-trained image recognition model specifically adopts a pre-trained EfficientNetV2-S model to obtain local features through the block3a layer and global features through the top_conv layer;

[0048] The visualization optimization is specifically to generate feature heatmap data and a list of key area coordinate data by adopting a data visualization method based on the standardized image feature vector data;

[0049] The image recognition enhancement is specifically to generate a structured image recognition report based on the feature heatmap data and the list of key area coordinate data to obtain reference data for enhanced recognition of chronic nephropathy pathological images.

[0050] The beneficial effects achieved by the present invention by adopting the above scheme are as follows:

[0051] (1) In the existing chronic kidney disease pathological image recognition and enhancement system, the traditional pathological image source mainly relies on 2D slices from a single perspective, which is difficult to reflect the three-dimensional structural relationship of nephrons. Image enhancement mainly depends on traditional image enhancement algorithms, lacking a deep enhancement mechanism based on the semantics of structure and pathological progression. Moreover, since the recognition module, structural modeling, and pathological evolution analysis are independent of each other, it is difficult to form a unified recognition and enhancement process. In view of these technical problems, this solution creatively adopts an overall recognition and enhancement idea that combines nephropathy pathological modeling, pathological evolution enhancement, and image recognition enhancement, realizing multi-level collaboration among structural modeling, evolution prediction, and image recognition, and significantly improving the accuracy, stability, and clinical interpretability of pathological recognition.

[0052] (2) In the existing nephropathy pathological modeling process, there are technical problems such as insufficient registration accuracy of pathological slice images, resulting in obvious structural distortion in three-dimensional reconstruction; traditional modeling methods ignore the spatial relationship modeling of microscopic structures such as cell nuclei and nephrons, making it difficult to support subsequent pathological evolution analysis and spatial lesion tracking. In view of these technical problems, this solution creatively adopts a renal structure tracking modeling method combined with improved registration to perform nephropathy pathological modeling, thereby accurately expressing the spatial structure characteristics of nephrons.

[0053] (3) In the existing pathological evolution enhancement process, there are technical problems such as the existing models not considering the spatio-temporal continuity of nephropathy evolution, lacking a mechanism to fuse structural changes and time dynamics, and lacking an effective modeling path when integrating clinical indicators. In view of these technical problems, this solution creatively adopts a three-dimensional spatio-temporal graph convolutional network combined with non-deterministic quantization and multi-scale feature fusion enhancement to perform pathological evolution enhancement, realizing dynamic modeling of renal structure over time, improving prediction stability under uncertainty perception enhancement, and embedding and fusing clinical-image bimodal features.

[0054] (4) In the existing image recognition and enhancement process, the models generally adopt shallow feature extraction strategies, making it difficult to obtain discriminative information of key lesion areas, lacking semantic fusion based on structural modeling and pathological evolution, resulting in the recognition task being insensitive to early changes, and at the same time having weak visualization ability, making it difficult to support the interpretable requirements of clinical auxiliary diagnosis. In view of these technical problems, this solution creatively adopts a pre-trained model recognition and analysis method to perform image recognition and enhancement, realizing deep recognition that fuses local-global structural features, structure-evolution-recognition collaborative modeling driven by multi-source data, and interpretable image recognition output based on heat maps and key area localization. Brief Description of the Drawings

[0055] Figure 1 It is a schematic structural diagram of a chronic kidney disease pathological image recognition and enhancement system provided by the present invention;

[0056] Figure 2Flow diagram of the steps executed by the data collection module;

[0057] Figure 3 Flow diagram of the steps executed by the nephropathy pathology modeling module;

[0058] Figure 4 Flow diagram of the steps executed by the pathology evolution enhancement module;

[0059] Figure 5 Flow diagram of the steps executed by the image recognition enhancement module;

[0060] Figure 6 Example diagram of the feature heat map data obtained by the image recognition enhancement module.

[0061] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation to the present invention. Detailed implementation manners

[0062] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts belong to the scope of protection of the present invention.

[0063] Embodiment 1. Refer to Figure 1 , the technical solution adopted by the present invention is as follows: A chronic nephropathy pathology image recognition enhancement system provided by the present invention includes a data collection module, a nephropathy pathology modeling module, a pathology evolution enhancement module, and an image recognition enhancement module;

[0064] The data collection module is used for data collection and optimization processing. Through data collection and optimization processing, an original dataset for enhanced pathology image recognition is obtained, and the original dataset for enhanced pathology image recognition is sent to the nephropathy pathology modeling module, the pathology evolution enhancement module, and the image recognition enhancement module;

[0065] The nephropathy pathology modeling module is used for nephropathy pathology modeling. Through nephropathy pathology modeling, renal topology modeling data is obtained, and the renal topology modeling data is sent to the pathology evolution enhancement module;

[0066] The pathology evolution enhancement module is used for pathology evolution enhancement. Through pathology evolution enhancement, comprehensive nephropathy pathology feature data is obtained and the comprehensive nephropathy pathology feature data is sent to the image recognition enhancement module;

[0067] The image recognition enhancement module is used for image recognition enhancement. Through image recognition enhancement, reference data for the recognition enhancement of chronic kidney disease pathological images is obtained.

[0068] By performing the above operations, in the existing chronic kidney disease pathological image recognition enhancement system, the traditional pathological image sources mainly rely on 2D slices from a single perspective, which are difficult to reflect the three-dimensional structural relationship of nephrons. Image enhancement mainly relies on traditional image enhancement algorithms and lacks a deep enhancement mechanism based on the semantics of structure and pathological progression. Also, because the recognition module, structure modeling, and pathological evolution analysis are independent of each other, it is difficult to form a unified recognition enhancement process. This solution creatively adopts an overall recognition enhancement idea that combines nephropathy pathological modeling, pathological evolution enhancement, and image recognition enhancement, realizing multi-level collaboration of structure modeling, evolution prediction, and image recognition, and significantly improving the accuracy, stability, and clinical interpretability of pathological recognition.

[0069] Example 2. This example is based on the above example. Refer to Figure 2 , the data collection and optimization processing is used to collect raw data and perform optimization processing. Specifically, through multi-source data collection, raw data of chronic kidney disease pathology is obtained, and through image preprocessing and data alignment, a raw dataset for the recognition enhancement of pathological images is obtained, including the following steps: multi-modal data acquisition, adaptive illumination enhancement, data cleaning and optimization, and overall data optimization;

[0070] The multi-modal data acquisition is specifically to obtain raw data of chronic kidney disease pathology through multi-source data collection. The raw data of chronic kidney disease pathology includes raw pathological image data, clinical biochemical data, and pathological text data. By sequentially performing time series data alignment, maximum inter-class variance threshold segmentation optimization processing, and manual data annotation on the raw data of chronic kidney disease pathology, an optimized dataset of chronic kidney disease pathology is obtained;

[0071] The optimized dataset of chronic kidney disease pathology includes maximum inter-class variance threshold segmentation optimized images and time series alignment annotation data;

[0072] The adaptive illumination enhancement is specifically to construct an adaptive illumination compensation U-shaped network based on the optimized dataset of chronic kidney disease pathology, perform color correction enhancement, and obtain a color correction dataset;

[0073] The adaptive illumination compensation U-shaped network includes an encoder structure, a decoder structure, and an optimized loss function;

[0074] The encoder adopts a 4-layer convolutional structure, sets the stride to 2, and the number of channels to [32, 64, 128, 256];

[0075] The decoder adopts a 4-layer transposed convolution structure, specifically using a bilinear upsampling and skip connection structure;

[0076] The optimized loss function includes a reconstruction loss and a perceptual loss;

[0077] The reconstruction loss specifically uses the L1 norm and is calculated based on the pixel difference between the input and output;

[0078] The perceptual loss is specifically obtained by constructing a pre-trained 3D convolutional layer of VGG16, optimizing the images in the chronic kidney disease pathology optimization dataset, extracting perceptual features, and constructing the perceptual loss;

[0079] The calculation formula of the optimized loss function is:

[0080] ;

[0081] In the formula, L is the optimized loss function, y is the input pixel value, specifically extracted from the chronic kidney disease pathology optimization dataset, f(x) is the output pixel value of the adaptive illumination compensation U-shaped network, x is the original image input, is a regulation factor, is the feature data extracted from the input pixel value through the pre-trained 3D convolutional layer of VGG16, is the feature data extracted from the output pixel value of the adaptive illumination compensation U-shaped network through the pre-trained 3D convolutional layer of VGG16, ||·||1 is the L1 norm operator, and ||·||2 is the L2 norm operator;

[0082] The data cleaning and optimization is specifically based on the color correction dataset, using data cleaning through fuzzy slice detection and data sampling enhancement through geometric transformation to obtain a cleaned and optimized dataset;

[0083] The fuzzy slice detection calculates the clarity by setting a clarity metric threshold and a clarity metric matrix, and the calculation formula is:

[0084] ;

[0085] In the formula, FS is the clarity calculation value, var(·) is the matrix element variance calculation function, C is the convolution kernel matrix, I is the pixel value of the case image slice in the color correction dataset, and C(·) is the convolution operation representation function;

[0086] The geometric transformation includes rotation, scaling, and translation;

[0087] The overall data optimization is specifically to perform the overall enhancement of the chronic kidney disease pathology optimization dataset through the adaptive illumination enhancement and the data cleaning and optimization to obtain the original dataset for enhanced pathological image recognition;

[0088] The enhanced original dataset for pathological image recognition includes enhanced image data, time series alignment annotation data, unenhanced image data, and enhanced transformation record data.

[0089] Example 3 is based on the above example. Refer to Figure 3 , the nephropathy pathological modeling is used to reconstruct the renal unit-level model in three dimensions. Specifically, based on the enhanced original dataset for pathological image recognition, a renal structure tracking and modeling method combining improved registration is adopted to perform nephropathy pathological modeling, and renal topological modeling data is obtained, including the following steps: slice feature matching, transformation matrix optimization, image deformation, nucleus center detection, node graph construction, and renal three-dimensional reconstruction;

[0090] The slice feature matching is specifically as follows: based on the enhanced original dataset for pathological image recognition, every two adjacent pathological slice images are used as the original data input, and the scale-invariant feature transform algorithm is used to extract slice feature points to obtain feature point feature descriptor data. Based on the feature point feature descriptor data, the standard fast robust binary descriptor algorithm is used to extract feature vectors, and candidate feature point matching data is obtained through two-way consistent matching;

[0091] The transformation matrix optimization is specifically as follows: the standard thin plate spline transformation algorithm is introduced to perform global non-linear deformation on the pathological slice images in the enhanced original dataset for pathological image recognition, and an improved registration minimization objective function is constructed based on the candidate feature point matching data to perform slice registration optimization, and slice transformation matrix data is obtained;

[0092] The calculation formula of the improved registration minimization objective function is:

[0093] ;

[0094] In the formula, F TRA is the improved registration minimization objective function, T is the slice transformation matrix data output by the standard thin plate spline transformation algorithm, K is the total number of feature points in the candidate feature point matching data, k is the feature point index, FREAK(·) is the standard fast robust binary descriptor algorithm function, I i+1 (·) is the pixel after transformation of the (i + 1)-th image, i is the pathological slice image index, T(·) is the standard thin plate spline transformation algorithm function, p k is the position parameter of the k-th feature point, SIFT(·) is the scale-invariant feature transform algorithm function, I i is the pixel after transformation of the i-th image, ||·||2 is the L2 norm operator, is the regularization coefficient, is the gradient output by the standard thin plate spline transformation algorithm, ||·|| F is the sum of squares of partial derivatives operator;

[0095] The image deformation is specifically to perform deformation processing on the pathological slice images in the original dataset of pathological image recognition enhancement according to the slice transformation matrix data to obtain three-dimensional reconstructed aligned pathological slice image data;

[0096] The nucleus center detection is specifically to construct a standard deep learning segmentation model and perform nucleus center detection according to the three-dimensional reconstructed aligned pathological slice image data to obtain nucleus center coordinate detection data;

[0097] The standard deep learning segmentation model specifically uses a pre-trained StarDist deep learning model;

[0098] The node graph construction is specifically to perform graph structure construction according to the nucleus center coordinate detection data to obtain nephron node graph data;

[0099] The renal three-dimensional reconstruction is specifically to perform maximum weighted path search and three-dimensional curve fitting in sequence according to the nephron node graph data, and through voxel volume construction, to obtain renal topological modeling data;

[0100] The maximum weighted path search specifically uses the standard Dijkstra algorithm for path search;

[0101] The renal topological modeling data includes candidate feature point matching data, three-dimensional voxel modeling data of renal tubules, and nephron node graph data.

[0102] By performing the above operations, in the existing pathological modeling process of kidney diseases, there are technical problems such as insufficient registration accuracy of pathological slice images leading to obvious structural distortion in three-dimensional reconstruction; traditional modeling methods ignore the spatial relationship modeling of microscopic structures such as cell nuclei and nephrons, and it is difficult to support subsequent pathological evolution analysis and spatial lesion tracking. This solution creatively adopts a kidney structure tracking modeling method combined with improved registration to perform pathological modeling of kidney diseases, so as to accurately express the spatial structure characteristics of nephrons.

[0103] Example 4, this example is based on the above example, refer to Figure 4 The pathological evolution enhancement is used to dynamically analyze the process of chronic kidney disease and perform feature fusion analysis. Specifically, according to the renal topological modeling data and the original dataset of pathological image recognition enhancement, a three-dimensional spatio-temporal graph convolutional network combined with non-deterministic quantization and multi-scale feature fusion enhancement is used to perform pathological evolution enhancement to obtain comprehensive pathological feature data of kidney diseases, including the following steps: spatio-temporal graph modeling, dynamic feature extraction, non-deterministic quantization enhancement, multi-scale feature fusion, and pathological evolution enhancement;

[0104] The spatio-temporal graph modeling specifically includes: based on the renal topology modeling data and the nephron node graph data, performing temporal graph modeling to construct temporal nephron node graph data, and through edge weight calculation, obtaining temporal nephron node edge weight data;

[0105] The dynamic feature extraction specifically includes: based on the temporal nephron node graph data and the temporal nephron node edge weight data, constructing a three-dimensional graph convolutional network and a temporal convolutional network to perform chronic kidney disease spatial feature and temporal feature modeling, obtaining kidney disease pathological spatial feature data and kidney disease pathological temporal feature data;

[0106] The calculation formula of the three-dimensional graph convolutional network is:

[0107] ;

[0108] In the formula, is the feature vector representation of the temporal nephron node v in the (l + 1)-th layer of the three-dimensional graph convolutional layer, used to represent the kidney disease pathological spatial feature data, ReLU(·) is a non-linear activation function, v is the temporal nephron node index, u is the adjacent node index of the temporal nephron node v, is the set of adjacent nodes, is the edge weight at time t, d v is the normalized degree matrix of the temporal nephron node v, d u is the normalized degree matrix of the adjacent node u of the temporal nephron node v, is the spatial weight of the l-th layer of the three-dimensional graph convolutional layer, is the feature vector representation of the adjacent node u of the temporal nephron node in the l-th layer of the three-dimensional graph convolutional layer;

[0109] The calculation formula of the temporal convolutional network is:

[0110] ;

[0111] In the formula, H t+1 is the kidney disease pathological temporal feature data at time t + 1, is the size of the temporal convolutional kernel, is the temporal convolutional offset index, W t is the temporal convolutional weight, is the node feature before the -th temporal convolutional offset interval, where d is the temporal dilation step parameter;

[0112] The non-deterministic quantization enhancement is specifically to perform pathological progression prediction based on the nephropathy pathological spatial feature data and nephropathy pathological time feature data, and enhance the non-deterministic perception of the pathological progression prediction by constructing an average prediction estimate and a non-deterministic prediction estimate, so as to obtain quantized enhanced pathological image feature data;

[0113] The calculation formula of the average prediction estimate is:

[0114] ;

[0115] In the formula, is the average prediction estimate, M is the total number of predictions, m is the prediction number index, is the result of the m-th prediction;

[0116] Preferably, the total number of predictions M is set to 50;

[0117] The calculation formula of the non-deterministic prediction estimate is:

[0118] ;

[0119] In the formula, U is the non-deterministic prediction estimate;

[0120] The multi-scale feature fusion is specifically to extract low-magnification features and high-magnification features from the quantized enhanced pathological image feature data, and obtain spliced feature data by upsampling and splicing the low-magnification features, and perform serum creatinine value feature embedding by constructing a standard multi-layer perceptron to obtain clinical embedded feature data, and perform feature fusion on the spliced feature data and the clinical embedded feature data to obtain multi-scale fusion pathological feature data;

[0121] The calculation formula of the feature fusion is:

[0122] ;

[0123] In the formula, F final is the multi-scale fusion pathological feature data, LayerNorm(·) is the layer normalization operation, Conv 1×1 (·) is the 1×1 feature fusion convolutional layer, Upsample(·) is the upsampling operation function, F 20x is the low-magnification feature, F 40x is the high-magnification feature, MLP(·) is the standard multi-layer perceptron function, sCr is the serum creatinine value feature, W c is the feature embedding weight;

[0124] The pathological evolution is enhanced. Specifically, based on the multi-scale fusion pathological feature data, a joint optimization loss function is constructed for model training to obtain a pathological evolution enhancement model, and the pathological evolution enhancement model is used to predict the pathological process to obtain comprehensive pathological feature data of kidney disease;

[0125] The calculation formula of the joint optimization loss function is:

[0126] ;

[0127] In the formula, L c is the joint optimization loss function, Y is the true pathological progression label value, is the pathological progression value predicted by the model, is the balance factor, KL(·) is the KL divergence calculation function, is the encoded posterior distribution for the given input X, p(z) is the standard normal distribution, and X is the original data input, which is used to represent the multi-scale fusion pathological feature data.

[0128] By performing the above operations, aiming at the technical problems existing in the existing pathological evolution enhancement process, that is, the existing model does not consider the spatio-temporal continuity of kidney disease evolution, lacks a mechanism for fusing structural changes and time dynamics, and lacks an effective modeling path when fusing clinical indicators, this solution creatively adopts a three-dimensional spatio-temporal graph convolutional network combined with non-deterministic quantization and multi-scale feature fusion enhancement for pathological evolution enhancement, realizing dynamic modeling of the renal structure over time, improving the prediction stability under uncertainty perception enhancement, and embedding and fusing clinical-image bimodal features.

[0129] Example 5. This example is based on the above example. Referring to Figure 5 and Figure 6 The image recognition is enhanced to comprehensively recognize the pathological images of kidney disease by combining the fusion features and three-dimensional reconstruction features. Specifically, based on the original dataset for enhanced pathological image recognition, the renal topological modeling data, and the comprehensive pathological feature data of kidney disease, a pre-trained model recognition and analysis method is used for image recognition enhancement to obtain reference data for enhanced recognition of pathological images of chronic kidney disease, including the following steps: multi-feature joint embedding, visualization optimization, and image recognition enhancement;

[0130] The multi-feature joint embedding is specifically to use a pre-trained image recognition model to perform multi-scale feature extraction on the original dataset for enhanced pathological image recognition, the renal topological modeling data, and the comprehensive pathological feature data of kidney disease to obtain standardized image feature vector data;

[0131] The pre-trained image recognition model specifically uses the pre-trained EfficientNetV2-S model to obtain local features through the block3a layer and global features through the top_conv layer;

[0132] The visualization optimization specifically generates feature heatmap data and a list of key region coordinate data by using a data visualization method based on the standardized image feature vector data;

[0133] The image recognition enhancement specifically generates a structured image recognition report based on the feature heatmap data and the list of key region coordinate data to obtain reference data for enhanced recognition of chronic kidney disease pathological images.

[0134] By performing the above operations, in view of the technical problems existing in the existing image recognition enhancement process, such as the model generally adopting a shallow feature extraction strategy, being difficult to obtain discriminative information of key lesion areas, lacking semantic fusion based on structural modeling and pathological evolution, resulting in the recognition task being insensitive to early changes, and at the same time having weak visualization ability and being difficult to support the interpretable requirements of clinical auxiliary diagnosis, this solution creatively adopts a pre-trained model recognition and analysis method for image recognition enhancement, realizing deep recognition that fuses local-global structural features, collaborative modeling of structure-evolution-recognition driven by multi-source data, and interpretable image recognition output based on heatmaps and key region localization.

[0135] It should be noted that in this article, relational terms such as first and second are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the term "comprising", "including" or any other variant thereof is intended to cover non-exclusive inclusion, so that a process and method comprising a series of elements not only includes those elements, but also includes other elements not expressly listed, or elements inherent to such a process and method.

[0136] Although the embodiments of the present invention have been shown and described, those of ordinary skill in the art can understand that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principle and spirit of the present invention.

[0137] The above describes the present invention and its implementation manners. Such a description is not restrictive. What is shown in the drawings is only one of the implementation manners of the present invention, and the actual structure is not limited thereto. Generally speaking, if those of ordinary skill in the art are inspired by it and design similar structural manners and embodiments without creative efforts without departing from the purpose of the present invention, they shall fall within the protection scope of the present invention.

Claims

1. An enhanced system for identifying pathological images of chronic kidney disease, characterized in that: A data collection module, a nephropathy pathology modeling module, a pathology evolution enhancement module, and an image recognition enhancement module; The data collection module is used for data collection and optimization processing. Through data collection and optimization processing, an original dataset for enhanced pathological image recognition is obtained, and the original dataset for enhanced pathological image recognition is sent to the nephropathy pathology modeling module, the pathology evolution enhancement module, and the image recognition enhancement module; The nephropathy pathology modeling module is used for nephropathy pathology modeling. Adopting a kidney structure tracking modeling method combined with improved registration, nephropathy pathology modeling is carried out to obtain kidney topology modeling data, and the kidney topology modeling data is sent to the pathology evolution enhancement module, including the following steps: slice feature matching, transformation matrix optimization, image deformation, nucleus center detection, node graph construction, and three-dimensional kidney reconstruction; The transformation matrix optimization is specifically to introduce a standard thin plate spline transformation algorithm to perform full-image non-linear deformation on the pathological slice images in the original dataset for enhanced pathological image recognition, and based on the candidate feature point matching data, construct a minimized objective function for improved registration to perform slice registration optimization; The calculation formula of the minimized objective function for the improved registration is: Where, F TRA is the minimization objective function for improved registration, T is the slice transformation matrix data output by the standard thin plate spline transformation algorithm, K is the total number of feature points in the candidate feature point matching data, k is the feature point index, FREAK(·) is the standard fast robust binary descriptor algorithm function, I i+1 (·) is the transformed pixel of the (i + 1)-th image, i is the pathological slice image index, T(·) is the standard thin plate spline transformation algorithm function, p k is the position parameter of the k-th feature point, SIFT(·) is the scale-invariant feature transform algorithm function, I i is the transformed pixel of the i-th image, ||·||2 is the L2 norm operator, λ is the regularization coefficient, is the gradient output by the standard thin plate spline transformation algorithm, ||·|| F is the sum of squares of partial derivatives operator; The pathology evolution enhancement module is used for pathology evolution enhancement. Adopting a three-dimensional spatio-temporal graph convolutional network combined with non-deterministic quantization and multi-scale feature fusion enhancement, pathology evolution enhancement is carried out to obtain comprehensive nephropathy pathology feature data, and the comprehensive nephropathy pathology feature data is sent to the image recognition enhancement module, including the following steps: spatio-temporal graph modeling, dynamic feature extraction, non-deterministic quantization enhancement, multi-scale feature fusion, and pathology evolution enhancement; For the non-deterministic quantization enhancement, based on the nephropathy pathology spatial feature data and the nephropathy pathology time feature data, pathological progression prediction is carried out, and by constructing an average prediction estimate and a non-deterministic prediction estimate, non-deterministic perception enhancement of the pathological progression prediction is carried out; for the multi-scale feature fusion, low-magnification features and high-magnification features are extracted, and by upsampling and splicing the low-magnification features, a standard multi-layer perceptron is constructed for serum creatinine value feature embedding, and through feature fusion, multi-scale fusion pathological feature data is obtained; The image recognition enhancement module is used for image recognition enhancement. Through image recognition enhancement, enhanced reference data for chronic nephropathy pathological image recognition is obtained.

2. The enhanced system for identifying chronic kidney disease pathological images according to claim 1, wherein: The data collection and optimization processing is used for collecting original data and performing optimization processing. Specifically, through multi-source data collection, original chronic nephropathy pathology data is obtained, and through image preprocessing and data alignment, an original dataset for enhanced pathological image recognition is obtained, including the following steps: multi-modal data acquisition, adaptive illumination enhancement, data cleaning optimization, and overall data optimization.

3. The chronic kidney disease pathological image recognition enhancement system according to claim 2, wherein: The multi-modal data acquisition is specifically to obtain original chronic nephropathy pathology data through multi-source data collection. The original chronic nephropathy pathology data includes original pathological image data, clinical biochemical data, and pathological text data; by sequentially performing time series data alignment, maximum inter-class variance threshold segmentation optimization processing, and manual data annotation on the original chronic nephropathy pathology data, an optimized dataset for chronic nephropathy pathology is obtained; The chronic kidney disease pathology optimization dataset includes the maximum between-class variance threshold segmentation optimization image and the time-series alignment annotation data; The adaptive light enhancement is specifically to construct an adaptive light compensation U-shaped network based on the chronic kidney disease pathology optimization dataset, perform color correction enhancement, and obtain a color correction dataset; The adaptive light compensation U-shaped network includes an encoder structure, a decoder structure, and an optimization loss function; The encoder adopts a 4-layer convolutional structure with a stride of 2 and the number of channels set to [32, 64, 128, 256]; The decoder adopts a 4-layer transposed convolutional structure, specifically using a bilinear upsampling and skip connection structure; The optimization loss function includes a reconstruction loss and a perceptual loss; The reconstruction loss specifically uses the L1 norm and is calculated based on the difference between the input and output pixels; The perceptual loss is specifically obtained by constructing a pre-trained VGG16 three-dimensional convolutional layer, extracting perceptual features based on the images in the chronic kidney disease pathology optimization dataset, and constructing a perceptual loss; The calculation formula of the optimization loss function is: In the formula, L is the optimization loss function, y is the input pixel value, specifically extracted from the chronic kidney disease pathology optimization dataset, f(x) is the output pixel value of the adaptive light compensation U-shaped network, x is the original image input, λ′ is the adjustment factor, φ(y) is the feature data extracted from the input pixel value through the pre-trained VGG16 three-dimensional convolutional layer, φ(f(x)) is the feature data extracted from the output pixel value of the adaptive light compensation U-shaped network through the pre-trained VGG16 three-dimensional convolutional layer, ||·||1 is the L1 norm operator, and ||·||2 is the L2 norm operator; The data cleaning optimization is specifically to perform data cleaning through fuzzy slice detection and perform data sampling enhancement through geometric transformation based on the color correction dataset to obtain a cleaned and optimized dataset; The fuzzy slice detection calculates the clarity by setting a clarity metric threshold and a clarity metric matrix; The geometric transformation includes rotation, scaling, and translation; The overall data optimization is specifically to perform overall enhancement of the chronic kidney disease pathology optimization dataset through the adaptive light enhancement and the data cleaning optimization to obtain a pathology image recognition enhanced original dataset; The pathology image recognition enhanced original dataset includes enhanced image data, time-series alignment annotation data, unenhanced image data, and enhanced transformation record data.

4. The enhanced system for identifying chronic kidney disease pathological images according to claim 3, wherein: The kidney disease pathology modeling is used to three-dimensionally reconstruct the nephron-level model. Specifically, based on the pathology image recognition enhanced original dataset, a kidney structure tracking modeling method combined with improved registration is used to perform kidney disease pathology modeling to obtain kidney topology modeling data, including the following steps: slice feature matching, transformation matrix optimization, image deformation, nucleus center detection, node graph construction, and kidney three-dimensional reconstruction.

5. An enhanced system for identifying chronic kidney disease pathological images according to claim 4, characterized in that: The slice feature matching is specifically as follows: based on the pathological images, the enhanced original dataset is recognized, and every two adjacent pathological slice images therein are used as the original data input. The scale-invariant feature transform algorithm is adopted to extract slice feature points, and the feature point feature descriptor data is obtained. Based on the feature point feature descriptor data, the standard fast robust binary descriptor algorithm is adopted to extract feature vectors, and through two-way consistent matching, the candidate feature point matching data is obtained; The transformation matrix optimization is specifically as follows: the standard thin plate spline transformation algorithm is introduced to perform global non-linear deformation on the pathological slice images in the enhanced original dataset of pathological image recognition. Based on the candidate feature point matching data, the minimized objective function for improved registration is constructed to optimize slice registration, and the slice transformation matrix data is obtained; The image deformation is specifically as follows: based on the slice transformation matrix data, the pathological slice images in the enhanced original dataset of pathological image recognition are deformed to obtain the three-dimensional reconstructed and aligned pathological slice image data; The nucleus center detection is specifically as follows: a standard deep learning segmentation model is constructed. Based on the three-dimensional reconstructed and aligned pathological slice image data, the nucleus center detection is performed to obtain the nucleus center coordinate detection data; The standard deep learning segmentation model specifically adopts the pre-trained StarDist deep learning model; The node graph construction is specifically as follows: based on the nucleus center coordinate detection data, the graph structure is constructed to obtain the nephron node graph data; The renal three-dimensional reconstruction is specifically as follows: based on the nephron node graph data, the maximum weighted path search and three-dimensional curve fitting are sequentially performed, and through voxel volume construction, the renal topological modeling data is obtained; The maximum weighted path search specifically adopts the standard Dijkstra algorithm for path search; The renal topological modeling data includes candidate feature point matching data, three-dimensional voxel modeling data of renal tubules, and nephron node graph data.

6. The enhanced system for identifying chronic kidney disease pathological images according to claim 5, wherein: The pathological evolution enhancement is used for dynamically analyzing the process of chronic kidney disease and performing feature fusion analysis. Specifically, based on the renal topological modeling data and the enhanced original dataset of pathological image recognition, the three-dimensional spatio-temporal graph convolutional network combined with non-deterministic quantization and multi-scale feature fusion enhancement is adopted to perform pathological evolution enhancement to obtain the comprehensive pathological feature data of kidney disease, including the following steps: spatio-temporal graph modeling, dynamic feature extraction, non-deterministic quantization enhancement, multi-scale feature fusion, and pathological evolution enhancement; The spatio-temporal graph modeling is specifically as follows: based on the renal topological modeling data and the nephron node graph data, the temporal graph modeling is performed to construct the temporal nephron node graph data, and through edge weight calculation, the temporal nephron node edge weight data is obtained; The dynamic feature extraction is specifically as follows: based on the temporal nephron node graph data and the temporal nephron node edge weight data, a three-dimensional graph convolutional network and a temporal convolutional network are constructed to perform the spatial feature and temporal feature modeling of chronic kidney disease, and the spatial pathological feature data and the temporal pathological feature data of kidney disease are obtained; The non-deterministic quantization enhancement specifically performs pathological progression prediction based on the nephropathy pathological spatial feature data and nephropathy pathological temporal feature data, and enhances the non-deterministic perception of pathological progression prediction by constructing an average prediction estimate and a non-deterministic prediction estimate, thereby obtaining quantized enhanced pathological image feature data; The multi-scale feature fusion specifically extracts low-magnification features and high-magnification features from the quantized enhanced pathological image feature data, obtains spliced feature data by upsampling and splicing the low-magnification features, and embeds serum creatinine value features by constructing a standard multi-layer perceptron to obtain clinically embedded feature data. Feature fusion is performed on the spliced feature data and the clinically embedded feature data to obtain multi-scale fused pathological feature data; The calculation formula for the feature fusion is: F final = LayerNorm(Conv 1×1 ([Upsample(F 20x );F 40x ) + MLP(sCr)W c ); where, F final is the multi-scale fusion pathological feature data, LayerNorm(·) is the layer normalization operation, Conv 1×1 (·) is the 1×1 feature fusion convolutional layer, Upsample(·) is the upsampling operation function, F 20x is the low magnification feature, F 40x is the high magnification feature, MLP(·) is the standard multi-layer perceptron function, sCr is the serum creatinine value feature, W c is the feature embedding weight; The pathological evolution enhancement specifically constructs a joint optimization loss function based on the multi-scale fused pathological feature data for model training to obtain a pathological evolution enhancement model, and uses the pathological evolution enhancement model to predict the pathological process to obtain comprehensive nephropathy pathological feature data.

7. An enhanced system for recognizing chronic kidney disease pathological images according to claim 6, characterized in that: The image recognition enhancement is used to comprehensively identify nephropathy pathological images by combining fused features and three-dimensional reconstruction features. Specifically, based on the original dataset for pathological image recognition enhancement, the renal topological modeling data, and the comprehensive nephropathy pathological feature data, a pre-trained model recognition analysis method is adopted to perform image recognition enhancement to obtain reference data for enhanced recognition of chronic nephropathy pathological images, including the following steps: multi-feature joint embedding, visualization optimization, and image recognition enhancement.

8. An enhanced system for recognizing pathological images of chronic kidney disease according to claim 7, characterized in that: The multi-feature joint embedding specifically uses a pre-trained image recognition model to perform multi-scale feature extraction based on the original dataset for pathological image recognition enhancement, the renal topological modeling data, and the comprehensive nephropathy pathological feature data to obtain standardized image feature vector data; The pre-trained image recognition model specifically uses a pre-trained EfficientNetV2-S model to obtain local features through the block3a layer and global features through the top_conv layer; The visualization optimization specifically uses a data visualization method to generate feature heatmap data and a list of key area coordinate data based on the standardized image feature vector data; The image recognition enhancement specifically generates a structured image recognition report based on the feature heatmap data and the list of key area coordinate data to obtain reference data for enhanced recognition of chronic nephropathy pathological images.

Images