Methods and Systems for Quantitative Assessment of Renal Pathological Section Fibrosis Degree

Abstract

This invention discloses a method and system for quantitatively assessing the degree of fibrosis in kidney pathological sections, belonging to the field of medical image processing technology. The method includes: performing color normalization and adaptive contrast enhancement preprocessing on Masson stained section images; classifying pixel-level tissue components using a multi-component semantic segmentation network containing a channel-space dual-path attention module; calculating the fibrosis area ratio and outputting a CI score based on the Banff grading standard; generating a fibrosis spatial distribution heatmap; and calculating the annual fibrosis progression rate through longitudinal follow-up comparison.

Description

Technical Field

[0001] This invention belongs to the field of medical image processing technology, specifically relating to a method and system for quantitatively evaluating the degree of fibrosis in kidney pathological sections. Background Technology

[0002] Chronic kidney disease (CHD) has become a major public health problem affecting human health worldwide, with its incidence rate rising annually. It is estimated that approximately 850 million people globally suffer from varying degrees of kidney disease. In the pathological progression of CHD, renal interstitial fibrosis is widely recognized as a key pathological indicator for predicting renal function deterioration and disease prognosis, with a significant positive correlation between the degree of fibrosis and the rate of decline in glomerular filtration rate (GFR). Accurate assessment of the degree of renal interstitial fibrosis is crucial for developing clinical treatment plans, monitoring the efficacy of anti-fibrotic drugs, and predicting patient prognosis. In clinical practice, Masson's trichrome staining method is the most commonly used histochemical staining method for assessing renal interstitial fibrosis. The blue-marked areas of collagen fiber deposition represent fibrotic tissue, while the red areas represent normal renal parenchyma. Currently, the clinical assessment of the degree of renal interstitial fibrosis mainly relies on pathologists' manual microscopic interpretation of renal biopsy tissue sections. Pathologists need to visually estimate the percentage of the renal cortex covered by blue collagen fiber areas under a microscope or in a digital slide viewer, and convert it into a CI score according to the international Banff classification standard. However, this manual assessment method has inherent drawbacks such as high subjectivity and poor repeatability. Studies have shown that different pathologists can have assessments of the percentage of fibrosis on the same slide with differences of 15% to 20%. Significant differences often exist between different pathologists, and even among the same pathologist at different time points, severely impacting the consistency and objectivity of clinical diagnosis. Furthermore, manual interpretation suffers from high time consumption and low throughput. An experienced pathologist typically needs 10 to 20 minutes to complete a fibrosis assessment of a full-field digital slide, which is insufficient to meet the needs of large-scale clinical screening.

[0003] To address the aforementioned clinical challenges, existing technologies have attempted to introduce computer-aided analysis methods to improve the objectivity of fibrosis assessment. For example, Chinese invention patent CN116245881A discloses a method and system for assessing renal interstitial fibrosis based on full-field recognition. This approach acquires full-field images of renal tissue sections and performs staining, uses a feature extraction model to identify fibrotic regions, then constructs polygonal regions using edge feature points to estimate the fibrotic area, and combines grayscale trend analysis to determine the direction of fibrosis spread. Finally, an assessment report is generated using a dual-scoring weighted method. This approach achieves, to a certain extent, automatic identification and area estimation of fibrotic regions.

[0004] However, the above scheme still has the following shortcomings: First, the scheme uses a general feature extraction model for fibrosis region identification without performing specific color normalization processing for the color characteristics of Masson stained sections. This results in color differences between different staining batches directly affecting the identification accuracy, especially in multi-center studies and mixed evaluation scenarios of sections from different laboratories, where the identification accuracy caused by staining color differences can decrease by 10% to 15%. Second, the scheme only identifies fibrosis regions and fails to distinguish multiple tissue components such as blue collagen fibers, normal renal parenchyma, glomeruli, renal tubules, and blood vessels at the pixel level. This leads to a lack of precise histological basis for calculating the fibrosis area and makes it impossible to accurately define the renal cortex region as the denominator of the area ratio. Third, the scheme estimates the fibrosis area through polygon approximation. For irregularly shaped diffuse fibrosis regions and linear fibrosis along the renal tubular basement membrane, this method is not suitable. Fourth, the scheme lacks alignment with the internationally accepted Banff pathological grading standard, and cannot directly output the CI score widely used by clinicians. Clinicians need to manually convert the area percentage, increasing the workload and the possibility of human error. Fifth, although the scheme performs diffusion trend analysis, this trend analysis infers the direction of fibrosis diffusion based on the direction of gray value gradient, and does not visualize the spatial distribution pattern and aggregation area of ​​fibrosis in the form of a heat map. It also lacks a systematic classification of fibrosis distribution patterns (diffuse, focal, or mixed). Sixth, the scheme does not have longitudinal follow-up comparison function, and cannot automatically track the fibrosis progression of the same patient at different time points and calculate the annual progression rate, which is difficult to meet the clinical needs for dynamic monitoring of fibrosis progression in the long-term management of chronic kidney disease. Summary of the Invention

[0005] The purpose of this invention is to provide a method and system for quantitative assessment of the degree of fibrosis in renal pathological sections, in order to solve the technical problems in the prior art, such as insufficient objectivity in assessing renal interstitial fibrosis, low accuracy in distinguishing multiple tissue components, lack of standardized grading output, and inability to conduct longitudinal follow-up comparisons.

[0006] To achieve the above objectives, this invention provides a method for quantitatively evaluating the degree of fibrosis in renal pathological sections, comprising the following steps: Step S1, performing color normalization preprocessing on the acquired Masson-stained renal pathological section image, converting the image from RGB color space to optical density space, and extracting the staining basis vector matrix based on singular value decomposition for color distribution mapping, and then performing adaptive limited contrast histogram equalization processing to enhance the contrast between the fibrotic collagen region and the normal renal parenchyma region, obtaining a preprocessed image; Step S2, inputting the preprocessed image into a pre-trained renal multi-component semantic segmentation network for pixel-level classification, the network including an encoder, a decoder, and a channel-space dual-path attention module, the encoder extracting multi-scale feature maps and passing them to the decoder after weighted fusion via the channel-space dual-path attention module, the output including blue collagen fiber deposition area, red normal renal parenchyma area, glomerular area, renal tubular area, and Step S3: Based on the multi-component segmentation mask, count the number of pixels in the blue collagen fiber deposition area and the total number of pixels in the renal cortex area, calculate the fibrosis area ratio, and convert the fibrosis area ratio into a CI score and fibrosis degree determination result according to the preset Banff grading threshold range; Step S4: Perform sliding window local density statistics on the blue collagen fiber deposition area in the multi-component segmentation mask to generate a fibrosis density matrix, perform Gaussian smoothing and pseudo-color mapping on the fibrosis density matrix to obtain a fibrosis spatial distribution heatmap; Step S5: Obtain multiple renal biopsy pathological images of the same patient at different time points and perform steps S1 to S4 respectively, construct a time series based on the fibrosis area ratio at each time point, calculate the annual fibrosis progression rate through linear regression fitting, and compare the annual fibrosis progression rate with the preset progression risk threshold to output the fibrosis progression risk level.

[0007] This invention also provides a quantitative assessment system for the degree of fibrosis in renal pathological sections, comprising: a color normalization preprocessing module for performing color normalization and adaptive contrast enhancement processing of Masson stained section images in step S1; a multi-component semantic segmentation module for performing pixel-level multi-tissue component classification of the renal multi-component semantic segmentation network including a channel-space dual-path attention module in step S2; a fibrosis quantitative scoring module for performing accurate calculation of the fibrosis area ratio and outputting a CI score based on the Banff grading standard in step S3; a spatial distribution analysis module for performing the generation of the fibrosis density matrix and visualization output of the fibrosis spatial distribution heatmap in step S4; and a longitudinal follow-up comparison module for performing the calculation of the annual fibrosis progression rate and determination of progression risk level based on multiple renal biopsies in step S5. The above modules form a data flow-driven closed-loop collaborative architecture, wherein the analysis results of the longitudinal follow-up comparison module can be fed back to the color normalization preprocessing module to optimize the color consistency parameters between images at different time points, ensuring that the longitudinal comparison results are not interfered with by non-biological color variations.

[0008] The beneficial effects of this invention are as follows: color normalization eliminates color difference interference from different staining batches, significantly improving the cross-center generalization ability of the segmentation model; a multi-component semantic segmentation network achieves pixel-level accurate differentiation of multiple tissue components, overcoming the deficiency of insufficient accuracy in polygon approximation area estimation in traditional methods; the introduction of a channel-space dual-path attention module enhances the segmentation accuracy of fibrosis boundary regions; the Banff standard grading achieves direct alignment with international clinical standards, and the output results can be directly applied to clinical decision-making; the heatmap visualization intuitively presents the spatial distribution characteristics of fibrosis, helping clinicians identify fibrosis distribution patterns; and the longitudinal follow-up comparison function enables objective quantitative tracking of disease progression, providing a comprehensive, objective, and standardized fibrosis assessment tool for the long-term management of chronic kidney disease. Attached Figure Description

[0009] Figure 1 This is a flowchart of a method for quantitatively evaluating the degree of fibrosis in kidney pathological sections provided in an embodiment of the present invention.

[0010] Figure 2 This is an architecture diagram of the kidney pathological section fibrosis degree image quantitative assessment system provided in this embodiment of the invention. Detailed Implementation

[0011] To make the above-mentioned objectives, technical solutions, and beneficial effects of the present invention clearer, the present 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 only for explaining the present invention and are not intended to limit the present invention.

[0012] Please see Figure 1 As shown, this embodiment of the invention provides a method for quantitatively assessing the degree of fibrosis in kidney pathological sections. This method is suitable for objectively quantifying the degree of fibrosis in kidney biopsy tissue sections stained with Masson's trichrome stain. In the Masson's trichrome staining system, collagen fibers are stained blue, muscle fibers and erythrocytes are stained red, and cell nuclei are stained black. This staining characteristic lays the optical foundation for subsequent automatic identification of fibrotic tissue. This embodiment uses high-resolution full-field images acquired by a digital whole-section scanner as input. The preferred scanning resolution is 0.25 μm / pixel corresponding to a 40x objective lens, and the image format is the standard SVS or NDPI format. The specific implementation process of this method is described in detail below with reference to each step.

[0013] Step S1: Color normalization preprocessing. In one embodiment of the present invention, step S1 specifically includes two sub-processes: tissue region detection and color normalization. The two work together to ensure that the subsequent segmentation network obtains input data with stable quality.

[0014] Specifically, tissue region detection is first performed on the acquired Masson-stained renal pathological slide images. Preferably, the original slide images are scaled to 1 / 16 of their original resolution to reduce computational load, then converted to the HSV color space, the saturation channel S is extracted, and a threshold of 0.05 to 0.15 is set. Pixels with saturation higher than this threshold are identified as tissue regions, and pixels with saturation lower than this threshold are identified as blank background regions. In one embodiment of the invention, to eliminate the interference of staining artifacts and isolated noise points at tissue edges, morphological opening operations are used to process the binarization results, where the structuring element is a circular kernel with a diameter of 15 pixels. Furthermore, isolated regions with an area smaller than a preset minimum connected component area threshold (preferably 1000 pixels) are removed, ultimately obtaining an effective tissue mask. This effective tissue mask will be used as a region of interest constraint in all subsequent processing steps to avoid background contamination of the analysis results.

[0015] After completing the tissue region detection, color normalization is performed on the image within the effective tissue region. In actual clinical pathology workflows, due to differences in staining reagent concentration, staining time, section thickness, and the color response characteristics of digital scanners across different laboratories, the color rendering of the same grade of fibrotic tissue may vary significantly in different sections. This invention employs a color normalization strategy based on singular value decomposition to eliminate this batch-to-batch color difference effect.

[0016] Specifically, the transformation of Masson-stained kidney pathological section images from the RGB color space to the optical density space is expressed as follows:

[0017] ,

[0018] in, For the first The optical density value of each color channel These represent the red, green, and blue color channels, respectively. For the first The pixel intensity value of each color channel ranges from 1 to 255 (to avoid division by zero errors in logarithmic operations, pixel values ​​of 0 are truncated to 1). The incident light intensity constant is typically set to 255. The physical significance of this conversion lies in transforming the light intensity domain to the chemical absorption characteristic domain, so that the spectral absorption characteristics of different dyes exhibit linearly separable vector directions in the optical density space.

[0019] Preferably, in the optical density space, all pixels constituting the effective tissue region... The optical density matrix undergoes singular value decomposition, retaining the first two principal components as staining basis vectors, corresponding to the spectral absorption directions of blue collagen dye and red myofiber dye in Masson staining, respectively. In one embodiment of the invention, the staining basis vector matrix of the current image is mapped to the staining basis vector matrix of a pre-established standard reference image using the least squares method, and the RGB values ​​of each pixel under the standard color distribution are reconstructed accordingly. The standard reference image is preferably a benchmark image selected by senior pathologists from a library of high-quality stained slides, whose color contrast and saturation between the blue collagen region and the red solid region are within the clinically recognized optimal range.

[0020] After color normalization, adaptive contrast-limited histogram equalization is further performed on the color-normalized image. Specifically, the CLAHE algorithm is used to enhance local contrast of the image, with its cropping and limiting parameters... Set to between 2.0 and 4.0, preferably 3.0, and set the block size to... Pixel. The technical significance of this step lies in the fact that even after color normalization, the contrast between the blue collagen deposits in some mildly fibrotic areas and the surrounding normal tissue may still be low. Local contrast enhancement using the CLAHE algorithm can effectively improve the visual discriminability of these blurred-boundary areas, thus providing clearer boundary information for subsequent semantic segmentation networks. After the synergistic processing of color normalization and contrast enhancement, the final output is a preprocessed image. This preprocessed image exhibits highly consistent color distribution characteristics and clear tissue component boundaries across different staining batches.

[0021] It is worth further explaining that, in a preferred embodiment of the present invention, the quality of color normalization is verified through an automated quality assessment process. Specifically, the statistical moment differences between the color-normalized image and the standard reference image in each channel of the Lab color space are calculated, including the mean difference. , , and standard deviation difference , , When the mean difference of any channel exceeds a preset threshold (preferably 5.0) or the standard deviation difference exceeds a preset threshold (preferably 8.0), the system determines that the color normalization effect is poor and automatically switches to the alternative standard reference image to re-execute the color normalization process. This quality control mechanism effectively ensures that each image entering the subsequent segmentation network has stable and reliable color quality. In addition, in the longitudinal follow-up comparison scenario (step S5), the standard reference selection strategy for color normalization is different from that in the single evaluation scenario. The color distribution of the patient's first puncture slice after normalization is used as the patient's exclusive longitudinal reference benchmark to ensure the consistency of longitudinal comparison.

[0022] Step S2: Multi-component semantic segmentation. In one embodiment of the present invention, step S2 inputs the preprocessed image output from step S1 into a pre-trained kidney multi-component semantic segmentation network to achieve pixel-level accurate classification of various functional components of kidney tissue. Unlike existing technologies that only perform foreground-background binary classification of fibrotic areas, the segmentation network of the present invention can simultaneously distinguish five key tissue components, including blue collagen fiber deposition areas (category label C1), red normal renal parenchyma areas (category label C2), glomerular areas (category label C3), renal tubular areas (category label C4), and vascular areas (category label C5), as well as background areas (category label C0), for a total of six semantic categories. The core advantage of this multi-component segmentation strategy is that the accurate calculation of the fibrosis area ratio requires accurate definition of the denominator, namely the total area of ​​the renal cortex, and the determination of the total area of ​​the renal cortex depends on the accurate location of cortical landmark structures such as glomeruli and renal tubules.

[0023] Specifically, the kidney multi-component semantic segmentation network employs an encoder-decoder architecture, with the encoder using ResNet-50 as the backbone network. This network was pre-trained on the ImageNet dataset to achieve good general feature representation capabilities. The encoder extracts feature maps at four scales sequentially through four residual module groups. The spatial resolutions of the feature maps at each scale are 1 / 4, 1 / 8, 1 / 16, and 1 / 32 of the input image, corresponding to 256, 512, 1024, and 2048 feature channels, respectively. This multi-scale feature extraction enables the network to capture both the global morphological features of large-scale structures such as the glomeruli and the fine details of small-scale structures such as fibrosis boundaries.

[0024] One of the key innovations of this invention lies in the introduction of a channel-space dual-path attention module between the encoder and decoder to enhance the segmentation accuracy of fibrotic boundary regions. In Masson stained sections, the boundary between fibrotic collagen regions and normal renal parenchyma often exhibits a gradual transition, and traditional convolutional operations are prone to classification confusion in these boundary regions. The channel-space dual-path attention module adaptively weights the feature maps from both the channel and spatial dimensions through two parallel attention paths.

[0025] Specifically, the channel attention branch first performs global average pooling and global max pooling on the input feature map along the spatial dimension, obtaining two channel description vectors respectively. Then, the two description vectors are concatenated and passed through a bottleneck structure with a compression ratio of 16 (i.e., two fully connected layers, with the number of channels in the middle layer being 1 / 16 of the number of input channels), and then the channel attention weight vector is generated by the Sigmoid activation function. The calculation process for channel attention is expressed as follows:

[0026] ,

[0027] in, The input feature map has a dimension of . , For the number of channels, and These represent the height and width of the feature map, respectively. This indicates a global average pooling operation. This represents the global max pooling operation. This represents a multilayer perceptron containing one hidden layer. The Sigmoid activation function has an output range of 1. , For dimension The channel attention weight vector.

[0028] The spatial attention branch performs average and maximum operations on the channel-weighted feature maps along the channel dimension, concatenating the two single-channel feature maps and then passing them through a convolutional kernel of size [size missing]. Convolutional layers and a sigmoid activation function are used to generate a spatial attention weight map. The computational process for spatial attention is expressed as follows:

[0029] ,

[0030] in, This is the feature map after channel attention weighting. This indicates element-wise multiplication and broadcasting along the channel dimension. and These represent the average pooling and max pooling operations along the channel dimension, respectively. This indicates a splicing operation along the channel dimension. Indicates the kernel size as Two-dimensional convolution operation, For dimension The spatial attention weight map. Finally, the feature map enhanced by the dual-path attention module is... .

[0031] The decoder employs an architecture combining step-by-step upsampling and skip connections. Starting with the lowest resolution feature map, it sequentially upsamples the feature maps to twice the resolution using bilinear interpolation, and then concatenates and fuses them with the feature maps of the corresponding layers in the encoder. Convolutional layers, BatchNorm layers, and ReLU activation function are used for feature refinement. The final output layer uses... Convolution and Softmax activation functions are used to output a 6-channel probability map. The sum of the probability values ​​of each pixel in the 6 categories is 1. The category with the highest probability is taken as the segmentation label of the pixel, and finally a multi-component segmentation mask map is obtained.

[0032] Preferably, in one embodiment of the present invention, the training process of the kidney multi-component semantic segmentation network is as follows: First, a training dataset is constructed, collecting at least 500 full-view images of Masson-stained kidney slices from different patients and different staining batches. These images are then pixel-level annotated by at least three senior pathologists with more than five years of experience in kidney pathology diagnosis, with the consistency of the annotations ensured through a majority voting mechanism. During training, a combined loss function is used to optimize the network parameters. The combined loss function is defined as:

[0033] ,

[0034] in, For weighted cross-entropy loss, For Dice's loss, and For weighting coefficients, preferably , In the weighted cross-entropy loss, the class weight of the blue collagen fiber deposition area (class C1) is set to 2 to 4 times, preferably 3 times, the class weight of the normal renal parenchyma area (class C2), to compensate for the class imbalance caused by the small proportion of fibrotic areas in the slice. Training uses the Adam optimizer, with an initial learning rate set to... The batch size was set to 8, the training epochs to 100, and a cosine annealing strategy was used to dynamically adjust the learning rate. The input image was cropped to... Image patches of pixels, data augmentation strategies include random horizontal flipping, random rotation ( ), random scaling (0.8x to 1.2x), and color jitter.

[0035] In one embodiment of this invention, an online hard sample mining strategy is also introduced during the training process. Specifically, in each training iteration, the cross-entropy loss is first calculated for all pixels in the current batch, and then the pixels are sorted in descending order of loss value, with only the top 70% of pixels with the highest loss value retained for gradient backpropagation. This strategy allows the network training to focus on regions that are difficult to segment, such as the transition boundary between fibrosis and normal tissue, blurred regions with uneven staining, and pathological structures with large morphological variations, thereby effectively improving the discrimination accuracy of the segmentation model in these hard sample regions. During the inference phase, for a full-view digital slice image with an input resolution of 40000×30000 pixels, the system uses an overlapping block strategy to divide it into multiple The image is divided into pixel blocks, with each adjacent block maintaining a 64-pixel overlap to eliminate boundary effects. After block inference, the overlapping areas are probabilistically fused using a soft voting strategy, ultimately outputting a complete multi-component segmentation mask for the entire slice. In a hardware environment equipped with a single NVIDIA RTX 4090 GPU, the average inference time for processing a typical 40x objective full-view digital slice is approximately 3 to 5 minutes, far faster than the 10 to 20 minutes required for manual interpretation by pathologists.

[0036] Step S3: Calculation of Fibrosis Area Ratio and Banff Grading. In one embodiment of the present invention, step S3, based on the multi-component segmentation mask output in step S2, achieves accurate calculation of the fibrosis area ratio and outputs a Banffci score conforming to international standards. The core technical idea of ​​this step is that the accurate calculation of the fibrosis area ratio depends not only on the accurate segmentation of the fibrosis region (numerator) but also on the accurate definition of the renal cortex region (denominator). Traditional methods often simply use the area of ​​the entire slice as the denominator, which artificially dilutes the fibrosis area ratio due to the inclusion of the renal medullary region.

[0037] In a preferred embodiment of the present invention, step S3 first performs automatic identification of the renal cortex region. Specifically, the renal cortex boundary is defined based on the spatial distribution characteristics of the glomerular region (category label C3) in the multi-component segmentation mask image. Glomeruli are distributed only within the renal cortex layer; therefore, the distribution density of glomeruli can serve as a reliable spatial marker of the renal cortex. In one embodiment of the present invention, the DBSCAN density clustering algorithm is used to spatially cluster the centroid coordinates of the glomerular region, wherein the neighborhood radius of the cluster is... Set to 500um to 1500um, preferably 800um, minimum sample size Set to 3. Continuous regions with a glomerular density higher than a preset density threshold (preferably one glomerulus per square millimeter) are identified as candidate regions for the renal cortex. The renal cortex boundary contour is then constructed using a convex hull algorithm or an Alpha-Shape algorithm, ultimately yielding the renal cortex mask. .

[0038] After determining the renal cortical mask, the formula for calculating the fibrosis area ratio is:

[0039] ,

[0040] in, The percentage of fibrous area is expressed as a percentage. For pixel coordinates, The set of pixels covered by the renal cortex mask. For pixels Category labels in a multi-component segmentation mask. Category label for blue collagen fiber deposition areas. Category labels for the background area, This is an indicator function; it takes the value 1 when the condition within the parentheses is true, and 0 otherwise. The numerator represents the total number of pixels in the renal cortex that are segmented into blue collagen fiber deposition areas, while the denominator represents the total number of all non-background pixels (i.e., effective tissue pixels) in the renal cortex. The ratio of the two is the fibrosis area ratio.

[0041] After calculating the fibrosis area ratio, it is converted into a CI score level according to a preset Banff grading threshold range. This invention sets the following threshold range based on the definitions of interstitial fibrosis and tubular atrophy (CI score) in the international Banff classification of renal pathology standards: fibrosis area ratio A CI score of 0 indicates no fibrosis or very slight fibrosis. The CI score was grade 1, corresponding to mild fibrosis; The CI score was grade 2, corresponding to moderate fibrosis. The CI score was 3, corresponding to severe fibrosis. The system also outputs a textual description of the degree of fibrosis, including four levels: none / very mild, mild, moderate, and severe, as well as the specific percentage of fibrosis area (accurate to one decimal place), providing clinicians with an intuitive quantitative reference.

[0042] Preferably, in one embodiment of the present invention, the quantitative fibrosis score is further subdivided into two levels: a global score and a regional score. The global score is the fibrosis area ratio and CI score based on the entire renal cortex, as described above. The regional score divides the renal cortex into three regions according to anatomical location: the superficial cortex, the intermediate cortex, and the deep cortex, and calculates the local fibrosis area ratio and local CI score for each region. The clinical significance of this stratified scoring strategy is that some types of chronic kidney disease (such as hypertensive nephropathy) often exhibit a gradient fibrosis distribution pattern that gradually worsens from the deep cortex to the superficial cortex, while other diseases (such as IgA nephropathy) may present a more diffuse distribution. Stratified scoring helps clinicians differentiate the types of fibrosis caused by different etiologies. The division of each region is based on the depth from the renal capsule to the corticomedullary junction. The superficial cortex is defined as the region within 1 / 3 of the depth from the renal capsule, the intermediate cortex is the region from 1 / 3 to 2 / 3 of the depth, and the deep cortex is the region from below 2 / 3 of the depth to the corticomedullary junction.

[0043] Step S4: Generating a thermal map of the spatial distribution of fibrosis. In one embodiment of the present invention, step S4, based on the quantitative score in step S3, further visualizes and analyzes the spatial distribution pattern of fibrotic tissue. Traditional fibrosis assessment only outputs a global area percentage value, which cannot reflect whether the distribution of fibrosis in kidney tissue is diffuse and uniform or focal and concentrated. This spatial distribution information is of great reference value for clinically determining the etiology and prognosis of fibrosis.

[0044] Specifically, a sliding window local density statistics method is used to perform local density statistics on the blue collagen fiber deposition region in the multi-component segmentation mask image. Preferably, the size of the sliding window... Set as Pixels (approximately at 40x objective scanning resolution) The sliding window size (32 pixels) is set to 1 / 4 of the window size to ensure a 75% overlap between adjacent windows, avoiding discontinuous jumps in density estimation. It should be noted that the sliding window size is not fixed at 128×128 pixels. Depending on the slice scanning resolution and tissue structure characteristics, the sliding window size can be adjusted from 64×64 pixels to 256×256 pixels: when using a 20x objective lens, a smaller window of 64×64 pixels to 128×128 pixels is preferred to match the lower spatial resolution; when using a 60x or higher objective lens, a larger window of 192×192 pixels to 256×256 pixels can be selected to cover a sufficient tissue area. Accordingly, the sliding step size is set to 1 / 4 to 1 / 2 of the window size to strike a balance between spatial continuity of density estimation and computational efficiency. For each sliding window position... The proportion of pixels in the blue collagen fiber deposition area within the window to the effective tissue pixels is calculated to form a fibrosis density matrix. The calculation formula is as follows:

[0045] ,

[0046] in, Position in the fiber density matrix The local fibrosis density value at the location, with a range of values ​​being: , For The set of pixels covered by the central sliding window. It is a very small positive number (preferably) (), used to avoid numerical anomalies where the denominator is zero.

[0047] After obtaining the fiber density matrix, Gaussian smoothing is performed on it to eliminate local statistical fluctuations and window boundary effects. The size of the Gaussian smoothing kernel is set to 2 to 4 times the sliding window size, preferably. The Gaussian kernel corresponding to each pixel, and its standard deviation (i.e., when the sliding window size is 128×128 pixels, the size of the Gaussian smoothing kernel is 256×256 pixels to 512×512 pixels; when the sliding window size is 64×64 pixels, the size of the Gaussian smoothing kernel is 128×128 pixels to 256×256 pixels; when the sliding window size is 256×256 pixels, the size of the Gaussian smoothing kernel is 512×512 pixels to 1024×1024 pixels). The density matrix is ​​set to 1 / 6 of the kernel size. The smoothed density matrix generates a thermal map of the spatial distribution of fibrosis using a pseudo-color mapping scheme. This mapping scheme employs a Jet color map, grading from blue (low fibrosis density) through green and yellow to red (high fibrosis density). Preferably, the transparency of the thermal map is set to 0.4 to 0.6, allowing it to be overlaid on the original slice image, enabling clinicians to simultaneously observe the thermal information of the spatial distribution of fibrosis and the underlying tissue morphological details.

[0048] In one embodiment of the present invention, automatic classification of spatial distribution patterns is also performed based on the fiber density matrix. Specifically, the spatial variation coefficient of the fiber density matrix is ​​calculated. :

[0049] ,

[0050] in, The standard deviation of the density values ​​at all effective locations in the fiber density matrix. This is the mean of all valid location density values. When it is determined to be a diffuse uniform distribution pattern, when When it is determined to be a mixed distribution pattern, when The pattern was determined to be a focal concentrated distribution pattern. Furthermore, for focal concentrated distribution patterns, further segmentation was performed using a threshold (the threshold was set to...). Identify fibrosis hotspots, calculate the number of hotspots, the area of ​​each hotspot, and their relative position in the renal cortex, and output a fibrosis aggregation analysis report.

[0051] Preferably, in one embodiment of the present invention, the spatial distribution analysis further includes fibrosis gradient direction analysis. Specifically, a gradient field is calculated on the Gaussian-smoothed fibrosis density matrix, and the gradient direction indicates the direction in which the fibrosis density increases most rapidly. By statistically summarizing the gradient directions of all effective locations within the entire renal cortex region, it can be determined whether fibrosis exhibits a trend of worsening from one direction to another, such as worsening from the capsule side to the medullary side or a pattern of expansion from a local area to the surrounding area. This gradient direction analysis provides valuable spatial reference information for clinicians to understand the pathological trajectory of fibrosis. In addition, the heatmap also supports joint analysis with the glomerular region in the multi-component segmentation mask map. The system automatically counts the local fibrosis density within a preset radius (preferably 200 μm) around each glomerulus and marks the top 10% of glomeruli with the highest fibrosis density as high-risk glomeruli, assisting clinicians in determining which glomeruli are in the leading edge region of fibrosis progression.

[0052] Step S5: Longitudinal follow-up comparison and annual fibrosis progression rate calculation. In one embodiment of the present invention, step S5 addresses the long-term management needs of chronic kidney disease by constructing a longitudinal follow-up comparison module. This module can automatically analyze the fibrosis progression of multiple renal biopsy images of the same patient at different time points. The clinical value of this function lies in the fact that patients with chronic kidney disease often require multiple renal biopsies to assess disease progression and treatment effectiveness. However, manual comparison of fibrosis changes at different time points is highly susceptible to subjective bias, especially when the rate of fibrosis progression is slow (e.g., an annual growth rate of only 2% to 3%), where manual interpretation can hardly reliably detect such minute changes.

[0053] Specifically, renal biopsy pathological images of the same patient at different time points (at least two time points, preferably three or more time points) are obtained, and all processing steps S1 to S4 are executed respectively to obtain the fibrosis area ratio corresponding to each time point. The system includes CI (Chronic Injection) score levels and a thermal map of the spatial distribution of fibrosis. It is worth noting that since renal biopsies at different time points may be performed in different laboratories, their staining conditions and scanning equipment may differ. Therefore, the color normalization preprocessing in step S1 is crucial for ensuring the consistency and reliability of longitudinal comparisons. In one embodiment of the present invention, in longitudinal follow-up mode, the system uses the color distribution of the patient's first biopsy slice image after color normalization as a reference. Images at subsequent time points are mapped to this reference for color normalization, thereby minimizing non-biological color variations between different time points.

[0054] After obtaining the fibrosis area ratio at each time point, the puncture time was used as the basis for the calculation. (Unit: Year) is the independent variable; fibrosis area ratio Using [variable name] as the dependent variable, a time series was constructed, and the annual rate of fibrosis progression was calculated by least squares linear regression fitting. :

[0055] ,

[0056] in, The annual rate of fibrosis progression is expressed as % / year. This represents the total number of longitudinal follow-up time points. For the first The timing of the second kidney biopsy (relative number of years from the time of the first biopsy). For the first The ratio of fibrosis area corresponding to the second renal biopsy. When A positive value indicates that fibrosis is progressing; when... A negative value indicates that fibrosis is regressing (suggesting that anti-fibrotic treatment may be effective). A value close to 0 indicates that the degree of fibrosis is relatively stable.

[0057] Preferably, the system simultaneously calculates the coefficient of determination for linear regression. To evaluate the goodness of linear fit, when The system indicates that fibrosis progression may exhibit non-linear characteristics, and clinicians are advised to make a comprehensive judgment based on clinical information. In non-linear cases, the system further provides quadratic polynomial fitting as an alternative analysis scheme to capture possible accelerating or decelerating progression trends. Furthermore, when only data from two time points are available, the annual fibrosis progression rate degenerates into a simple calculation of the difference in the ratio of fibrosis area between the two points divided by the time interval.

[0058] This invention compares the annual rate of fibrosis progression with a preset risk threshold and outputs a risk level for fibrosis progression. Specifically, when... At / year, it was determined to be low risk, indicating slow progression of renal function, and it was recommended to maintain the current treatment plan and have routine follow-up; when At / year, it was determined to be of medium risk (Moderate Risk), indicating a relatively rapid progression of fibrosis, and it was recommended to intensify anti-fibrotic treatment and shorten the follow-up interval; when A diagnosis of "High Risk" at a certain time (year) indicates rapid progression of renal fibrosis, and it is recommended to urgently evaluate treatment options and consider intensive intervention measures.

[0059] It is worth noting that the aforementioned risk thresholds for progression are based on the international consensus in renal pathology. Patients with an annual fibrosis progression rate exceeding 2% per year have a significantly increased risk of progressing to end-stage renal failure within 5 years, while those exceeding 5% per year have an extremely poor prognosis. The system also outputs a comprehensive follow-up report including the fibrosis area ratio at each time point, CI score, spatial distribution heatmap, and progression trend curve, providing the clinical team with comprehensive longitudinal assessment information.

[0060] In one embodiment of the present invention, the longitudinal follow-up comparison function also includes differential analysis of thermal maps showing the spatial distribution of fibrosis at different time points. Specifically, the fibrosis density matrices generated at different time points for the same patient are spatially registered and then subtracted pixel by pixel to obtain a fibrosis density difference matrix. This difference matrix is ​​visualized in a bidirectional pseudo-color manner: positive areas (increased fibrosis) are rendered in red, negative areas (reduced fibrosis) are rendered in blue, and areas near zero (stable fibrosis) are rendered in white. This differential thermal map allows clinicians to intuitively observe the specific location and extent of fibrosis progression in a spatial dimension, rather than relying solely on global percentage changes in area, thereby providing more refined imaging evidence for targeted treatment strategy adjustments.

[0061] Preferably, during spatial registration, since the location and angle of renal biopsy samples may differ at different time points, the system employs a rigid registration method based on the spatial distribution pattern of the glomeruli. The system uses the coordinate sets of the glomerular centroids at two time points as the feature points for registration, and estimates rotation and translation parameters using an iterative nearest-point algorithm to transform the density matrix of subsequent time points to the coordinate system of the first time point. When the registration residual exceeds a preset threshold (preferably 300 μm), the system indicates insufficient reliability of the spatial registration and advises clinicians to carefully interpret the results of the difference heatmap.

[0062] Please see Figure 2As shown, this embodiment of the invention also provides a system for quantitatively evaluating the degree of fibrosis in kidney pathological sections, which corresponds one-to-one with each step in the above method embodiments. This system can be deployed on a workstation or cloud server equipped with a GPU computing card. Preferred hardware configurations include an NVIDIA RTX 3090 or higher performance GPU (with at least 24GB of video memory), 64GB or more of system memory, and 2TB or more of solid-state drive storage space.

[0063] The kidney pathological section fibrosis degree image quantitative assessment system includes a color normalization preprocessing module 1, a multi-component semantic segmentation module 2, a fibrosis quantitative scoring module 3, a spatial distribution analysis module 4, and a longitudinal follow-up comparison module 5.

[0064] The color normalization preprocessing module 1 is configured to execute all the processing steps described in step S1. This module receives the original Masson stained slide image output from a digital whole-slice scanner. First, the tissue region detection subunit separates the tissue region from the background region in the HSV color space based on a saturation threshold, obtaining an effective tissue mask. Then, the color normalization subunit converts the image within the effective tissue region from the RGB color space to the optical density space, extracts the staining basis vector matrix based on singular value decomposition, and maps it to the color distribution of a standard reference image. Finally, the contrast enhancement subunit performs CLAHE algorithm processing to enhance the contrast between the fibrotic collagen region and the normal renal parenchyma region, outputting the preprocessed image to the multi-component semantic segmentation module 2. In one embodiment of the invention, the color normalization preprocessing module 1 has a built-in standard reference image library containing multiple sets of standard reference images optimized for different scanner models and different laboratory staining conditions. The system can automatically select the most matching standard reference image based on the metadata of the input image.

[0065] The multi-component semantic segmentation module 2 is configured to perform the pixel-level multi-tissue component classification described in step S2. This module internally encapsulates a pre-trained kidney multi-component semantic segmentation network, which employs an architecture including a ResNet-50 encoder, a channel-space dual-path attention module, and a progressive upsampling decoder, as described in the method embodiments. This module receives the pre-processed image output by the color normalization preprocessing module 1 and processes it according to… The pixel grid is divided into blocks, and each block is input into the semantic segmentation network for inference. The segmentation results of each block are then stitched and fused according to their original spatial positions to obtain a multi-component segmentation mask for the entire slice. Preferably, during the block stitching process, a 64-pixel overlap region is set between adjacent blocks, and the segmentation probability within the overlap region is averaged before determining the final label to eliminate segmentation discontinuity artifacts at the block boundaries.

[0066] The fibrosis quantitative scoring module 3 is configured to perform the fibrosis area ratio calculation and Banff ci score output described in step S3. This module receives the multi-component segmentation mask image output by the multi-component semantic segmentation module 2. First, the cortical recognition subunit automatically defines the renal cortex region based on the glomerular distribution density. Then, the area ratio calculation subunit calculates the ratio of blue collagen fiber deposition area pixels to effective tissue pixels within the renal cortex region. Finally, the grading scoring subunit outputs the ci score level, fibrosis degree determination result, and fibrosis area ratio value accurate to one decimal place based on the Banff threshold range.

[0067] The spatial distribution analysis module 4 is configured to perform the fibrosis heatmap generation and spatial distribution pattern analysis described in step S4. This module receives the multi-component segmentation mask image output by the multi-component semantic segmentation module 2, calculates the fibrosis density matrix through a sliding window density statistics sub-unit, and then generates a fibrosis spatial distribution heatmap via Gaussian smoothing and pseudo-color mapping sub-units. This module can also automatically calculate the spatial variation coefficient and determine the spatial distribution pattern of fibrosis (diffuse, mixed, or focal). When a focal concentrated distribution is identified, it automatically labels the location and area information of the fibrosis hotspot region.

[0068] The longitudinal follow-up comparison module 5 is configured to perform the annual fibrosis progression rate calculation and risk level output described in step S5. This module associates multiple renal biopsy pathology images of the same patient at different time points using a patient identifier. It calls the color normalization preprocessing module 1, the multi-component semantic segmentation module 2, and the fibrosis quantitative scoring module 3 to process the images at each time point, obtaining the fibrosis area ratio at each time point and constructing a time series for linear regression fitting. In one embodiment of the invention, the analysis results of the longitudinal follow-up comparison module 5 can be fed back to the color normalization preprocessing module 1 to optimize the color consistency parameters between images at different time points. For example, when a significant deviation in the color distribution of an image at a certain time point is detected, the system automatically adjusts the color normalization mapping strategy for that image, thereby forming a closed-loop optimization mechanism between the color normalization preprocessing module 1 and the longitudinal follow-up comparison module 5. The module ultimately outputs a comprehensive follow-up report including a trend chart of the fibrosis area ratio at each time point, a CI score change curve, the annual fibrosis progression rate, and the progression risk level.

[0069] In one embodiment of the present invention, the data flow between the above modules is standardized and encapsulated through a unified intermediate data format. Specifically, the preprocessed image output by the color normalization preprocessing module 1 is stored in lossless PNG format with accompanying color normalization parameter metadata; the multi-component semantic segmentation mask output by the multi-component segmentation module 2 is stored in single-channel label image format, with each pixel value corresponding to a category label (0 to 5); the quantitative scoring results output by the fibrosis quantitative scoring module 3 are stored in structured JSON format, including fields such as fibrosis area ratio, CI score level, and stratified region score; the heatmap output by the spatial distribution analysis module 4 is stored in PNG image format with an alpha channel, along with the original numerical data of the density matrix. This standardized intermediate data format design ensures loose coupling and high maintainability between the modules, facilitating the flexible expansion of new analytical function modules according to clinical needs.

[0070] Preferably, the system provided in this embodiment of the invention also includes a visual interactive interface, which integrates functions such as slide browsing, overlay display of segmentation results, adjustment of heatmap transparency, viewing of CI score reports, and display of longitudinal follow-up trend graphs. Clinicians can intuitively review the system's automatic analysis results through this interface and make local corrections to the segmentation results when necessary. The system supports exporting the evaluation report as a standardized pathology report in PDF format. The report content includes basic patient information, slide thumbnails, multi-component segmentation result graphs, fibrosis area ratio and CI score, fibrosis spatial distribution heatmap, and longitudinal follow-up trend analysis results, meeting the workflow requirements of clinical pathology departments and electronic medical record archiving requirements.

[0071] To verify the technical effectiveness of the method and system for quantitatively evaluating the degree of fibrosis in kidney pathological sections provided by this invention, the applicant conducted the following systematic comparative experiment.

[0072] For the experimental dataset, a total of 680 full-field digital slide images of kidney biopsies stained with Masson's stained dye were collected from the renal pathology centers of three tertiary hospitals. These images covered samples from CI scores of 0 to 3, with the following distribution: CI0: 180 images; CI1: 220 images; CI2: 180 images; CI3: 100 images. Of these, 480 images were used for training and validation (divided in an 8:2 ratio), and 200 images served as the independent test set. All slides were independently annotated pixel-wise and scored by five senior pathologists with over 10 years of experience in renal pathology diagnosis, with the majority vote used as the gold standard. The testing environment consisted of a workstation equipped with an NVIDIA RTX 4090 GPU and 128GB of RAM, running Ubuntu 22.04, and using PyTorch 2.0 as the deep learning framework.

[0073] In terms of semantic segmentation accuracy, the kidney multi-component semantic segmentation network of this invention achieved an average Dice coefficient of 0.862 for five tissue components on an independent test set. Specifically, the Dice coefficient for the blue collagen fiber deposition area was 0.841, for the red normal renal parenchyma area it was 0.912, for the glomerular area it was 0.893, for the renal tubular area it was 0.856, and for the vascular area it was 0.808. Compared to the baseline model without a channel-space dual-path attention module (i.e., standard U-Net with ResNet-50 encoder), the addition of the attention module improved the Dice coefficient of the blue collagen fiber deposition area by 0.047, and the average pixel accuracy increased from 0.891 to 0.927. The segmentation accuracy of the boundary region was particularly significantly improved, indicating that the attention mechanism plays a crucial role in enhancing the feature representation of the fibrotic boundary region. Compared with the model without color normalization preprocessing, the average Dice coefficient on cross-center (different hospitals) data improved from 0.783 to 0.862 after adding color normalization, an improvement of 10.1%, which fully verifies the significant gain of color normalization on the model's cross-domain generalization performance.

[0074] Regarding CI score consistency, the intraclass correlation coefficient (ICC) between the CI score results of this invention and the gold standard reference of 5 pathologists reached 0.92, significantly better than the average ICC value of 0.78 among the 5 pathologists. In the test-retest reliability experiment, the same batch of 50 slides were repeatedly processed after a 2-week interval. The system's test-retest ICC reached 0.99, much higher than the test-retest ICC value of 0.82 for manual interpretation, indicating that the scoring consistency and repeatability stability of this invention system are significantly better than those of manual interpretation. Regarding the quantitative accuracy of the fibrosis area ratio, the Pearson correlation coefficient between the fibrosis area ratio output by the system and the gold standard reference of pathologists was [not specified]. The mean absolute error was 2.3%, and the 95% agreement limit was -4.8% to +5.1%.

[0075] In terms of longitudinal follow-up comparison, paired slides from 50 patients with chronic kidney disease who underwent at least two renal biopsies were selected for automatic calculation of the annual fibrosis progression rate. The results showed that the system-calculated annual fibrosis progression rate was 88% consistent with the progression trend assessment by pathologists (Cohen's Kappa = 0.81). Furthermore, the system could detect minute progression with an annual fibrosis area ratio change as low as 1.5%, while the consistency of pathologists' interpretations significantly decreased when the area ratio change was less than 5%. Regarding processing efficiency, the system completed the entire evaluation process (from color normalization to heatmap generation) of a full-view digital slide in an average of 4.2 minutes, only about 28% of the average 15 minutes required for manual assessment by pathologists, increasing throughput by approximately 3.6 times, demonstrating significant clinical application value.

[0076] The experimental results above demonstrate that this invention, through a deep-coupled closed-loop synergy of color normalization preprocessing, multi-component semantic segmentation, Banff standardized scoring, spatial distribution heatmap visualization, and longitudinal follow-up comparison, achieves an objective, accurate, and standardized quantitative assessment of the degree of fibrosis in renal pathological sections. All technical indicators meet or exceed existing technical levels, demonstrating promising prospects for clinical translational applications. It is particularly noteworthy that a tight data coupling relationship is formed between the steps of this invention: the color normalization output of step S1 directly determines the input quality of the segmentation network in step S2; the multi-component segmentation results of step S2 simultaneously provide basic data for the area ratio calculation in step S3 and the heatmap generation in step S4; the quantitative scoring results of step S3 drive the longitudinal time series construction in step S5; and the longitudinal analysis results of step S5 can in turn optimize the color normalization parameter settings in step S1. Overall, this constitutes a closed-loop synergistic architecture combining forward data-driven and backward parameter feedback. This deep coupling mechanism allows the optimization of each link in the system to mutually promote each other, ultimately achieving a comprehensive evaluation effect superior to using any single technical means alone.

[0077] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for quantitatively assessing the degree of fibrosis in kidney pathological sections, characterized in that, Includes the following steps: Step S1: Perform color normalization preprocessing on the acquired Masson stained kidney pathological slide image to obtain a preprocessed image; Step S2: The preprocessed image is input into a pre-trained kidney multi-component semantic segmentation network for pixel-level classification. The kidney multi-component semantic segmentation network includes an encoder, a decoder, and a channel-space dual-path attention module. The encoder extracts multi-scale feature maps and passes them to the decoder after weighted fusion by the channel-space dual-path attention module. The output is a multi-component segmentation mask map including the blue collagen fiber deposition area, the red normal renal parenchyma area, the glomerular area, the renal tubular area, and the vascular area. Step S3: Based on the multi-component segmentation mask, count the number of pixels in the blue collagen fiber deposition area and the number of pixels in the total renal cortex area, calculate the fibrosis area ratio, and convert the fibrosis area ratio into a CI score level and fibrosis degree determination result according to the preset Banff grading threshold range. Step S4: Perform sliding window local density statistics on the blue collagen fiber deposition area in the multi-component segmentation mask to generate a fibrosis density matrix. Perform Gaussian smoothing and pseudo-color mapping on the fibrosis density matrix to obtain a fibrosis spatial distribution heatmap. Step S5: Obtain multiple renal biopsy pathological images of the same patient at different time points and execute steps S1 to S4 respectively. Construct a time series based on the fibrosis area ratio at each time point, calculate the annual fibrosis progression rate by linear regression fitting, and compare the annual fibrosis progression rate with a preset progression risk threshold to output the fibrosis progression risk level.

2. The method for quantitatively assessing the degree of fibrosis in kidney pathological sections according to claim 1, characterized in that, In step S1, the color normalization preprocessing involves converting the kidney Masson stained pathological slide image from RGB color space to optical density space, extracting the staining basis vector matrix based on singular value decomposition, and using the staining basis vector matrix to map the color distribution of the kidney Masson stained pathological slide image to the color distribution of a preset standard reference image to obtain a color normalized image. Adaptive limited contrast histogram equalization is then performed on the color normalized image to enhance the contrast between the fibrotic collagen region and the normal renal parenchyma region. Specifically, the conversion from RGB color space to optical density space uses the negative logarithm of each channel pixel value. The first two principal components are retained as staining basis vectors in the singular value decomposition. The cropping limit parameter for the adaptive limited contrast histogram equalization is set to 2.0 to 4.0, and the block size is set to 8×8 pixels.

3. The method for quantitatively assessing the degree of fibrosis in kidney pathological sections according to claim 1, characterized in that, In step S2, the encoder uses ResNet-50 as the backbone network and extracts feature maps at four scales. The spatial resolution of each scale feature map is 1 / 4, 1 / 8, 1 / 16 and 1 / 32 of the input image, respectively. The compression ratio of the channel attention branch in the channel-space dual-path attention module is set to 16, and the convolution kernel size of the spatial attention branch is set to 7×7.

4. The method for quantitatively evaluating the degree of fibrosis in kidney pathological sections according to claim 1, characterized in that, In step S3, the Banff grading threshold range is specifically as follows: when the fibrous area ratio is less than or equal to 5%, the ci score is level 0; when the fibrous area ratio is greater than 5% and less than or equal to 25%, the ci score is level 1; when the fibrous area ratio is greater than 25% and less than or equal to 50%, the ci score is level 2; and when the fibrous area ratio is greater than 50%, the ci score is level 3.

5. The method for quantitatively assessing the degree of fibrosis in kidney pathological sections according to claim 1, characterized in that, In step S2, the training process of the kidney multi-component semantic segmentation network includes: constructing a training dataset containing at least 500 Masson stained kidney slice images annotated at the pixel level by pathology experts; optimizing network parameters using a combined loss function, which includes a weighted sum of weighted cross-entropy loss and Dice loss, wherein the class weight of the blue collagen fiber deposition area is set to 2 to 4 times the class weight of the normal renal parenchyma area.

6. The method for quantitatively assessing the degree of fibrosis in kidney pathological sections according to claim 1, characterized in that, In step S4, the size of the sliding window is set to 64×64 pixels to 256×256 pixels, the sliding step size is set to 1 / 4 to 1 / 2 of the window size, and the kernel size of the Gaussian smoothing is set to 2 to 4 times the size of the sliding window.

7. The method for quantitatively evaluating the degree of fibrosis in kidney pathological sections according to claim 1, characterized in that, In step S5, when calculating the annual fibrosis progression rate using linear regression fitting, it is required to include fibrosis area ratio data at at least two time points. The preset progression risk threshold is as follows: an annual fibrosis progression rate of less than 2% / year is considered low risk, an annual fibrosis progression rate of greater than or equal to 2% / year and less than 5% / year is considered medium risk, and an annual fibrosis progression rate of greater than or equal to 5% / year is considered high risk.

8. The method for quantitatively assessing the degree of fibrosis in kidney pathological sections according to claim 1, characterized in that, In step S1, before performing color normalization preprocessing, a step of performing tissue region detection on the kidney Masson stained pathological slide image is also included. Specifically, the kidney Masson stained pathological slide image is converted to the HSV color space, the tissue region and the background region are separated based on the threshold set by the saturation channel, and isolated noise regions with an area smaller than the preset minimum connected region area threshold are removed to obtain an effective tissue mask.

9. The method for quantitatively assessing the degree of fibrosis in kidney pathological sections according to claim 1, characterized in that, In step S3, before calculating the fibrosis area ratio, there is also a step of automatic identification of the renal cortex region. Specifically, based on the distribution density characteristics of the glomerular region in the multi-component segmentation mask image, a density clustering algorithm is used to identify the renal cortex boundary. Continuous regions with glomerular distribution density higher than a preset density threshold are identified as renal cortex regions, thus eliminating the interference of the renal medulla region on the calculation of the fibrosis area ratio.

10. A system for quantitatively evaluating the degree of fibrosis in kidney pathological sections, used to implement the method for quantitatively evaluating the degree of fibrosis in kidney pathological sections according to any one of claims 1-9, characterized in that, include: The color normalization preprocessing module is configured to perform color normalization preprocessing on the acquired kidney Masson staining pathological slide image, convert the image from RGB color space to optical density space, extract the staining basis vector matrix based on singular value decomposition and perform color distribution mapping, and then perform adaptive limited contrast histogram equalization processing to output the preprocessed image. The multi-component semantic segmentation module is configured to input the preprocessed image into a pre-trained kidney multi-component semantic segmentation network for pixel-level classification. The network includes an encoder, a decoder, and a channel-space dual-path attention module, and outputs a multi-component segmentation mask map including a blue collagen fiber deposition area, a red normal renal parenchyma area, a glomerular area, a renal tubular area, and a vascular area. The quantitative scoring module for fibrosis is configured to calculate the fibrosis area ratio based on the multi-component segmentation mask image, and convert the fibrosis area ratio into a CI score level and a fibrosis degree determination result according to the Banff grading threshold range. The spatial distribution analysis module is configured to perform sliding window local density statistics on the blue collagen fiber deposition area in the multi-component segmentation mask image, and generate a fibrous spatial distribution heat map after Gaussian smoothing and pseudo-color mapping. The longitudinal follow-up comparison module is configured to acquire multiple renal biopsy pathological images of the same patient at different time points and process them respectively through the color normalization preprocessing module, the multi-component semantic segmentation module and the fibrosis quantitative scoring module. Based on the fibrosis area ratio at each time point, the annual fibrosis progression rate is calculated by linear regression fitting, and the fibrosis progression risk level is output.

Images