Animal pathology image data enhancement analysis method based on image processing

By constructing a multivariate mapping relationship and a pathological semantic perception enhancement network, the systematic color difference problem in animal pathological image staining technology was solved, achieving efficient image quality improvement and model training efficiency enhancement, while reducing the annotation burden.

CN123071192APending Publication Date: 2026-11-10HUNAN ANSHENMEI PHARM RES INST CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611292846.6
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-08-25
Publication Date
2026-11-10

AI Technical Summary

Technical Problem

Existing animal pathology image staining techniques suffer from systematic color differences across batches and institutions, lacking physical interpretability. Furthermore, traditional enhancement strategies cannot differentiate between normal tissues and lesion areas, leading to the easy destruction of key pathological edges and staining details.

Method used

By constructing a multivariate mapping relationship between experimental conditions, image appearance, and staining depth, source-based adaptive normalization is achieved. Combined with multi-scale structural feature matching, tissues are divided into regions with consistent features, unique key features, and unique non-key features. Layered segmentation and enhancement are performed, and adaptive processing is carried out using a pathological semantic perception enhancement network to ensure targeted enhancement of key regions and conventional amplification of non-key regions.

Benefits of technology

It achieves physically interpretable correction across batches of images, maintains tissue structure stability, improves pathological image quality and model training efficiency, reduces the annotation burden on veterinary pathologists, and provides high-quality labeled datasets.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123071192A_ABST
    Figure CN123071192A_ABST
Patent Text Reader

Abstract

This invention discloses a method for enhancing animal pathological image data based on image processing. The method includes artifact detection and preprocessing, fitting a first mapping relationship and a second mapping relationship, staining and normalizing the effective tissue region image, extracting multi-scale structural features and matching them with a public database to determine consistent and unique feature regions, dividing unique key feature regions and unique non-key feature regions, slicing the feature regions into blocks according to different cutting conditions, performing standard enhancement to obtain a first enhanced block, performing adaptive enhancement to obtain a second enhanced block, and stitching the first standard enhanced block, the second enhanced block, and the normalized image together according to stitching conditions to obtain an enhanced animal pathological image. The image is then input into a few-sample structure annotation network to obtain annotation information. This method can directionally enhance key pathological edges and staining details while maintaining the stability of normal tissue structures, effectively improving the quality of animal pathological images and the efficiency of subsequent model training.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of digital pathology and medical image processing technology, and in particular to a method for enhancing and analyzing animal pathological image data based on image processing. Background Technology

[0002] Digital pathology enables precise assessment of animal tissue lesions through whole-slice scanning and image analysis, serving as a core tool for veterinary clinical diagnosis, drug toxicology evaluation, and laboratory animal quality control. The image quality of H&E-stained tissue sections directly determines the reliability of subsequent quantitative analysis, while data augmentation is a crucial step in overcoming the bottleneck of scarce animal pathology samples and enhancing the generalization ability of deep learning models.

[0003] Existing staining normalization techniques largely rely on pixel statistical characteristics or generative style transfer, failing to trace the root causes of experimental conditions such as staining concentration, temperature, time, and instrument model. This leads to systematic color differences across batches / institutions and a lack of physical interpretability. Traditional enhancement strategies apply uniform geometric or color transformations to the entire image, easily disrupting key pathological semantics such as cell nucleus-cytoplasm ratio and tissue topology, and cannot perform differentiated processing for normal tissues and lesion areas. Therefore, this invention proposes an animal pathology image data enhancement and analysis method based on image processing. By constructing a multivariate mapping relationship between experimental conditions, image appearance, and staining depth, source-based adaptive normalization is achieved. Based on multi-scale structural feature matching, tissues are divided into consistent features, unique key features, and unique non-key features, and layered segmentation and enhancement are performed. While maintaining the stability of normal tissue structure, key pathological edges and staining details are targeted and enhanced, effectively improving the quality of animal pathology images and the efficiency of subsequent model training, and providing standardized data support for high-precision analysis in scenarios with few samples. Summary of the Invention

[0004] The purpose of this invention is to provide a method for enhancing and analyzing animal pathological image data based on image processing.

[0005] To achieve the above objectives, the present invention is implemented according to the following technical solution: This invention includes the following steps: Artifact detection was performed on whole-section images of H&E-stained tissue sections from mammals to generate artifact masks, and then processed to obtain images of the effective tissue regions. Reference slice data from the same organ and source location in the same family are obtained from a public database. The changes in experimental conditions, appearance features, and staining depth are calculated. The first and second mapping relationships are fitted and obtained. The effective tissue area image is normalized by combining the experimental conditions of the current slice to obtain a normalized image. Extract multi-scale structural features from normalized images, match them with public databases to determine consistent feature regions and unique feature regions, calculate the feature quality index of unique feature regions, and divide unique key feature regions and unique non-key feature regions based on the annotations in the public database. According to the differentiated segmentation rules, the consistent feature region, the unique key feature region, and the unique non-key feature region are segmented into blocks respectively. The unique non-key feature blocks are then enhanced according to the standard to obtain the first enhanced block. Consistency feature blocks and unique key feature blocks are input into the pathological semantic perception enhancement network for adaptive enhancement to obtain the second enhanced block. The first standard enhanced block, the second enhanced block and the corresponding positions of the normalized image are replaced and spliced ​​according to the splicing conditions to obtain the enhanced animal pathological image. The enhanced animal pathology images are input into a few-sample structure annotation network to obtain the annotation information of the enhanced animal pathology images; The reference slice data includes experimental conditions, appearance characteristics, and staining depth; The standard enhancements include random horizontal flipping, random vertical flipping, rotation in multiples of 90°, random perturbation of the HSV color space, and addition of Gaussian noise.

[0006] Furthermore, the method for obtaining an effective tissue region image includes: Artifact detection on the whole slice image: An algorithm based on the fusion of gray-level co-occurrence matrix and local binary mode is used to perform pixel-level annotation of fold, tear, bubble and knife mark regions to generate artifact mask; The area covered by the artifact mask is excluded from the full-slice image to obtain an image of the effective tissue region.

[0007] Furthermore, the method for obtaining the normalized image includes: Reference slide data from the same family and organ source location as the target animal were obtained from a public database. The changes in experimental conditions, appearance characteristics, and staining depth among different reference slides were calculated. The staining depth was calculated using optical density. Least squares fitting is performed on the changes in experimental conditions and changes in apparent characteristics to determine the mapping matrix to be estimated and to establish the first mapping relationship. Ridge regression estimation was performed on the changes in experimental conditions and staining depth to determine the mapping vector to be estimated and to establish a second mapping relationship. Extract the experimental conditions of the current slice, select recommended experimental conditions from the same family, genus, and organ source location from the public database to calculate the experimental condition deviation, and predict the appearance feature deviation and staining depth deviation of the current slice based on the first mapping relationship and the second mapping relationship. An adaptive color transformation is performed on the effective tissue region image based on the apparent feature deviation to obtain a color-normalized image. The overall contrast of the color-normalized image is then compensated with Gamma based on the staining depth deviation to obtain a normalized image.

[0008] Furthermore, the method for dividing the region into unique key feature regions and unique non-key feature regions includes: The normalized image is divided into several structural regions, and multi-scale structural features of each structural region are extracted; the multi-scale structural features include cell-level features, tissue-level features, and relation-level features. Calculate the Mahalanobis distance between the multi-scale structural feature vectors of each structural region and the mean of the standard animal slice template in the public database. When the Mahalanobis distance is less than the first distance threshold, the corresponding structural region is classified as a consistent feature region; otherwise, the corresponding structural region is classified as a unique feature region. Calculate the Euclidean distance between the multi-scale structural feature vector of the unique feature region and the structural feature vector of the annotated region of the same family of mammals in the public database. When the Euclidean distance is less than a second distance threshold, match the unique feature region with the annotation information. Calculate the annotation confidence of the unique feature region based on the matching result. The matching result includes the associated disease nature and annotation frequency. Calculate the feature quality index of the unique feature region. When the product of the annotation confidence and the feature quality index is greater than the criticality threshold, the corresponding unique feature region is classified as a unique critical feature region; otherwise, it is classified as a unique non-critical feature region. The feature quality index is positively correlated with the resolution score and the edge sharpness score, and negatively correlated with the information entropy.

[0009] Furthermore, the method for segmenting the consistent feature region, the unique key feature region, and the unique non-key feature region into blocks includes: Regions exhibiting consistent characteristics are cut according to the natural boundaries of histological structural units, with a fixed cut size. Adjacent slices are set with 50% overlap, and the proportion of consistent feature pixels within a single slice is greater than 80%, thus obtaining a set of consistent feature slices; the fixed value of the slice size is from 128×128 pixels to 1024×1024 pixels; preferably, the fixed value of the slice size is 256×256 pixels or 512×512 pixels. The unique key feature region is divided by expanding the bounding rectangle of the key structure by 10% and then cutting, with the cut block size adaptive value. By fitting the minimum bounding rectangle, the proportion of key feature pixels is greater than 30%, and the block boundaries avoid the artifact mask marking area, thus obtaining a unique set of key feature blocks; The unique, non-critical feature regions are divided into non-overlapping sections according to a fixed grid, with a fixed grid size of [size missing]. After removing the artifact mask marking area, the remaining part is directly output to obtain a unique non-critical feature block set; the fixed grid size range is 256×256 pixels to 2048×2048 pixels; preferably, the fixed grid size range is 256×256 pixels; The adaptive value of the cutting size The method for determining the width of the rectangle is as follows: perform minimum bounding rectangle fitting on the unique key feature region to obtain the rectangle width. With height Expand outwards with a preset margin expansion ratio of 0.1 to obtain the expanded width. and the height after expansion Take the maximum value of the expanded size as the base side length, and round it up to the smallest available size in the preset size range set. .

[0010] Furthermore, the method for obtaining the second enhanced slice includes: Consistency feature blocks and unique key feature blocks are input into the pathological semantic perception enhancement network, and processed sequentially through a shared coding layer, a structure preservation layer, an edge-anchor enhancement layer, a staining enhancement layer, an artifact removal layer, a generative enhancement layer, and a decoding layer to output a second enhancement block; the second enhancement block includes a consistency second enhancement block and a key second enhancement block; The pathological semantic perception enhancement network is constrained by a weighted combination of structure preservation loss, edge preservation loss, pathological semantic loss and adversarial loss.

[0011] Furthermore, the method for obtaining enhanced animal pathological images includes: Enhanced animal pathological images are obtained by replacing and stitching the first standard enhanced block, the second enhanced block, and the corresponding positions of the normalized image according to the pathological topological continuity constraints, key feature embedding constraints, and artifact filling rules. The pathological topological continuity constraint includes: if the pixel difference at the overlapping boundary of adjacent consistent feature blocks is less than a pixel difference threshold, otherwise, linear fusion based on distance weight is performed; The key feature embedding constraint includes: when a unique key feature block is embedded into the background, the boundary gradient difference is less than the gradient difference threshold. The artifact filling rule includes: the artifact mask marked area is filled by radial basis function interpolation of neighboring non-artifact pixels.

[0012] Furthermore, the method for obtaining annotation information for enhanced animal pathological images includes: Enhanced animal pathological images are input into a few-shot structure annotation network. The network outputs pathological structure-level annotations through segmentation, regression, and topology branches to obtain annotation information for the enhanced animal pathological images. The few-shot structure annotation network adopts an episodic training strategy and performs structural annotation in the latent space through prototype similarity. The pathological structure-level annotations include cell nuclear contours, cytoplasmic contours, gland shapes, vascular branch topology, interstitial density distribution, and intercellular connectivity. The few-shot structure annotation network improves the model's annotation accuracy through segmentation loss, regression loss, and topological loss.

[0013] The beneficial effects of this invention are: This invention is a method for enhancing and analyzing animal pathological image data based on image processing. Compared with existing technologies, this invention has the following technical advantages: This invention constructs a first mapping relationship and a second mapping relationship, and predicts the deviation of appearance features and staining depth based on the deviation between the current slice and the recommended experimental conditions. This enables traceable adaptive staining normalization, allowing systematic color differences across batches to be corrected in a physically explainable way, and improving the stability and reproducibility of normalization. This invention divides a normalized image into regions of consistent features, regions of unique key features, and regions of unique non-key features, enabling the enhancement strategy to have the ability to perceive pathological semantics and avoiding the application of indiscriminate transformations to the entire image. This invention ensures that the slices meet their respective enhancement requirements in terms of size, overlap rate, and feature ratio by setting differentiated cutting constraints, providing structured input units for layered processing, and eliminating slice boundary effects by setting splicing conditions, thus ensuring the continuity of the organizational structure and visual naturalness of the enhanced full-slice image. This invention performs standard enhancement on unique non-critical features and adaptive enhancement on the pathological semantic perception enhancement network for consistent features and unique critical features. It achieves synergistic optimization of conventional amplification of non-critical regions, targeted enhancement of critical regions, and preservation of consistent region structure, thereby expanding sample diversity while preventing distortion of key pathological structures. This invention significantly reduces the annotation burden on veterinary pathologists by inputting enhanced images into a few-sample structure annotation network, outputting structural annotation information such as cells, glands, and blood vessels without requiring extensive manual annotation, and providing high-quality labeled datasets for downstream quantitative analysis. Attached Figure Description

[0014] Figure 1 This is a flowchart illustrating the steps of the animal pathological image data enhancement and analysis method based on image processing according to the present invention. Detailed Implementation

[0015] The present invention will be further described below through specific embodiments. The illustrative embodiments and descriptions herein are used to explain the present invention, but are not intended to limit the present invention.

[0016] The present invention provides an animal pathological image data enhancement and analysis method based on image processing, comprising the following steps: like Figure 1 As shown, this embodiment includes the following steps: Artifact detection was performed on whole-section images of H&E-stained tissue sections from mammals to generate artifact masks, and then processed to obtain images of the effective tissue regions. Reference slice data from the same organ and source location in the same family are obtained from a public database. The changes in experimental conditions, appearance features, and staining depth are calculated. The first and second mapping relationships are fitted and obtained. The effective tissue area image is normalized by combining the experimental conditions of the current slice to obtain a normalized image. Extract multi-scale structural features from normalized images, match them with public databases to determine consistent feature regions and unique feature regions, calculate the feature quality index of unique feature regions, and divide unique key feature regions and unique non-key feature regions based on the annotations in the public database. According to the differentiated segmentation rules, the consistent feature region, the unique key feature region, and the unique non-key feature region are segmented into blocks respectively. The unique non-key feature blocks are then enhanced according to the standard to obtain the first enhanced block. Consistency feature blocks and unique key feature blocks are input into the pathological semantic perception enhancement network for adaptive enhancement to obtain the second enhanced block. The first standard enhanced block, the second enhanced block and the corresponding positions of the normalized image are replaced and spliced ​​according to the splicing conditions to obtain the enhanced animal pathological image. The enhanced animal pathology images are input into a few-sample structure annotation network to obtain the annotation information of the enhanced animal pathology images; The reference slice data includes experimental conditions, appearance characteristics, and staining depth; The standard enhancements include random horizontal flipping, random vertical flipping, rotation in multiples of 90°, random perturbation of the HSV color space, and addition of Gaussian noise.

[0017] In this embodiment, the method for obtaining an effective tissue region image includes: Artifact detection on the whole slice image: An algorithm based on the fusion of gray-level co-occurrence matrix and local binary mode is used to perform pixel-level annotation of fold, tear, bubble and knife mark regions to generate artifact mask; The area covered by the artifact mask is excluded from the full-slice image to obtain an image of the effective tissue region.

[0018] In this embodiment, the method for obtaining a normalized image includes: Reference slide data from the same family and organ source location as the target animal were obtained from a public database. The changes in experimental conditions, appearance characteristics, and staining depth among different reference slides were calculated. The staining depth was calculated using optical density. Least squares fitting is performed on the changes in experimental conditions and apparent characteristics to determine the mapping matrix to be estimated, and the first mapping relationship is established, expressed as: ; ; in For the vector of apparent feature changes, The mapping matrix to be estimated is... Let the vector represent the changes in experimental conditions. The residual of the first mapping, The optimal mapping matrix to be estimated is... The total number of samples, , for The vector of changes in the apparent characteristics of the sample and the vector of changes in experimental conditions. It is the L2 norm. Let be the regularization coefficient of the first mapping. It is the Frobenius norm; Ridge regression estimation was performed on the changes in experimental conditions and staining depth to determine the mapping vector to be estimated, and a second mapping relationship was established, expressed as: ; ; in This represents the change in staining depth. Let be the mapping vector to be estimated. The residual of the second mapping, The optimal mapping vector to be estimated. for The amount of change in staining depth of the sample. is the regularization coefficient for the second mapping; Extract the experimental conditions of the current slice, select recommended experimental conditions from the same family, genus, and organ source location from the public database to calculate the experimental condition deviation, and predict the appearance feature deviation and staining depth deviation of the current slice based on the first mapping relationship and the second mapping relationship. Adaptive color transformation is performed on the effective tissue region image based on apparent feature deviation: the original RGB of the effective tissue region image is converted to the optical density space, an adaptive correction matrix based on apparent feature deviation is applied to the optical density space of the effective tissue region image, and finally the image is inversely transformed back to the RGB space to obtain the color-transformed image. The normalized image is obtained by performing Gamma compensation on the overall contrast of the color normalized image based on the staining depth deviation. In actual evaluation, the experimental conditions include hematoxylin concentration, eosin concentration, staining temperature, staining time, scanning instrument model, section thickness, section size, fixation time, and fixative type. The apparent features are macroscopic statistics that do not affect cell structure features, including mean and standard deviation of brightness, mean and standard deviation of hematoxylin optical density, mean and standard deviation of eosin optical density, hue concentration, gray-level co-occurrence matrix energy, and sharpness index. The formula for calculating staining depth is: ; in For the first The staining depth of the slide, In order to effectively organize the region, A set of color channels, including red. ,green and blue , This represents the maximum transmission intensity captured by the scanner. For pixels exist The transmission intensity value of the color channel; the L2 norm of the optical density vector (i.e., the total light absorption) is calculated to reflect the color point situation. The larger the value, the stronger the overall absorption of light by the tissue, resulting in deeper staining. A value close to 0 indicates a near-blank background (very light staining or no tissue). The optical density spatial representation of the effective tissue region image is as follows: ; in For pixels The optical density at that location is represented by, For pixels The original RGB at the location; Optical density correction is expressed as: ; ; in For color transformation images, To effectively organize the optical density representation of a region image, This is an adaptive correction matrix based on apparent feature bias. Let be the vector of apparent characteristic changes in the hematoxylin channel. Let be the vector of apparent feature changes in the eosinophilic channel. Based on hue deviation The rotation correction matrix; The formula for calculating Gamma compensation is: ; in For pixels RGB after Gamma compensation For normalized images, For pixels After optical density correction, the RGB values ​​are corrected. This is the Gamma compensation coefficient. This is the depth compensation coefficient.

[0019] In this embodiment, the method for dividing the region into unique key feature regions and unique non-key feature regions includes: The normalized image is divided into several structural regions, and multi-scale structural features of each structural region are extracted; the multi-scale structural features include cell-level features, tissue-level features, and relation-level features. Calculate the Mahalanobis distance between the multi-scale structural feature vectors of each structural region and the mean of the standard animal slice template in the public database. When the Mahalanobis distance is less than the first distance threshold, the corresponding structural region is classified as a consistent feature region; otherwise, the corresponding structural region is classified as a unique feature region. Calculate the Euclidean distance between the multi-scale structural feature vector of the unique feature region and the structural feature vector of the annotated region of the same family of mammals in the public database. When the Euclidean distance is less than a second distance threshold, match the unique feature region with the annotation information. Calculate the annotation confidence of the unique feature region based on the matching result. The matching result includes the associated disease nature and annotation frequency. Calculate the feature quality index of the unique feature region. When the product of the annotation confidence and the feature quality index is greater than the criticality threshold, the corresponding unique feature region is classified as a unique critical feature region; otherwise, it is classified as a unique non-critical feature region. The feature quality index is positively correlated with the resolution score and the edge sharpness score, and negatively correlated with the information entropy. In practical evaluation, existing image segmentation methods (thresholding segmentation, edge detection, connected component analysis, morphological operations, watershed algorithm, or any combination of methods) are used to quickly divide the normalized image into several structural regions. Existing feature extraction techniques (gray-level co-occurrence matrix, local binary pattern, histogram of oriented gradients, scale-invariant feature transformation, hierarchical feature activation map of pre-trained convolutional neural network, or any combination of methods) are used to extract multi-scale structural features of the structural regions. Artifact masked regions are excluded during feature extraction. Multiscale structural features include cellular features (nuclear area, nuclear perimeter, nuclear roundness, nuclear-cytoplasmic ratio), tissue-level features (glandular / ductal shape factor, stroma density, number of vascular branches), and relational features (nearest neighbor cell distance, cell orientation entropy). The standard animal slide template is determined by statistically analyzing the common structural features of multiple normal slides from the same organ source location of mammals of the same family. The formulas for calculating annotation confidence and feature quality index are as follows: ; ; in For annotation confidence level, For unique feature regions The amount of annotation information in multi-scale structural feature matching. To associate annotation information with the nature of the disease, This is a disease characteristic function, taking the value 1 for malignant lesions, 0.6 for inflammatory lesions, and 0.3 for benign lesions. The normalization constant is For characteristic quality index, , , For feature weights, The resolution score for unique feature regions is determined by the ratio of the resolution to the baseline resolution. The edge sharpness score for unique feature regions is determined by the ratio of the edge sharpness to the baseline edge sharpness. This is the normalized value of the information entropy of the unique feature region.

[0020] In this embodiment, the method for dividing the consistent feature region, the unique key feature region, and the unique non-key feature region into blocks includes: Regions exhibiting consistent characteristics are cut according to the natural boundaries of histological structural units, with a fixed cut size. Adjacent slices are set with 50% overlap, and the proportion of consistent feature pixels within a single slice is greater than 80%, thus obtaining a set of consistent feature slices; the fixed value range of the slice size is 128×128 pixels to 1024×1024 pixels. The unique key feature region is divided by expanding the bounding rectangle of the key structure by 10% and then cutting, with the cut block size adaptive value. By fitting the minimum bounding rectangle, the proportion of key feature pixels is greater than 30%, and the block boundaries avoid the artifact mask marking area, thus obtaining a unique set of key feature blocks; The unique, non-critical feature regions are divided into non-overlapping sections according to a fixed grid, with a fixed grid size of [size missing]. After removing the artifact mask marking area, the remaining part is directly output to obtain a unique set of non-critical feature blocks; the fixed grid size ranges from 256×256 pixels to 2048×2048 pixels. The adaptive value of the cutting size The method for determining the width of the rectangle is as follows: perform minimum bounding rectangle fitting on the unique key feature region to obtain the rectangle width. With height Expand outwards with a preset margin expansion ratio of 0.1 to obtain the expanded width. and the height after expansion Take the maximum value of the expanded size as the base side length, and round it up to the smallest available size in the preset size range set. ; In practical assessments, when cutting areas with consistent characteristics, histological structural units refer to the original functional units in mammalian tissue sections that are surrounded by the basement membrane, stroma, or intrinsic connective tissue and have independent morphological boundaries, including but not limited to: glands (acini and ductal structures), liver lobules, glomeruli, vascular bundles, nerve bundles, muscle fiber bundles, lymphoid follicles, and cartilaginous lacunae. The pixel size corresponds to the physical size after scanner calibration. At 20× objective magnification, 256×256 pixels correspond to a physical tissue area of ​​approximately 128μm×128μm. When cutting out regions with unique key features, the preset size range includes 64 pixels, 128 pixels, 256 pixels, 512 pixels, and 1024 pixels. 1024 pixels is taken as the upper limit constraint. If the expanded size exceeds the upper limit constraint, the center is cropped with a side length of 1024 pixels to ensure that the proportion of key feature pixels is still greater than 30%. When cutting non-critical feature regions, the grid is arranged in rows and columns at equal intervals on the entire normalized image without overlap. For any grid block, if the proportion of artifact mask marked pixels exceeds 30%, the corresponding block is directly discarded; otherwise, after the artifact mask marked region is set to zero, the remaining valid pixels are directly output and do not participate in subsequent enhancement processing.

[0021] In this embodiment, the method for obtaining the second enhanced slice includes: Consistency feature blocks and unique key feature blocks are input into the pathological semantic perception enhancement network, and processed sequentially through a shared coding layer, a structure preservation layer, an edge-anchor enhancement layer, a staining enhancement layer, an artifact removal layer, a generative enhancement layer, and a decoding layer to output a second enhancement block; the second enhancement block includes a consistency second enhancement block and a key second enhancement block; The pathological semantic perception enhancement network is constrained by a weighted combination of structure preservation loss, edge preservation loss, pathological semantic loss, and adversarial loss, expressed as: ; ; ; ; ; in For the joint loss function, , , , For joint weights, To preserve structural losses, To preserve loss at the edge, For pathological semantic loss, To combat the losses, For the second enhanced cutting Slice from the original input The structural similarity includes brightness and contrast. , For consistency feature index, For indexing unique key features, For the set of anchor points, For the key second enhancement cut block Anchor point Spatial gradient at that location The edge enhancement coefficient, For the second enhanced cutting Feature vectors in the encoder's latent space A discriminator to distinguish between real chunks and augmented chunks; In practical evaluation, the shared coding layer employs a frozen feature encoder to map the input consistent feature blocks and unique key feature blocks to the latent space through forward propagation, outputting consistent latent space features. and key latent space features The frozen feature encoder is a pre-trained pathological baseline model. The structure-preserving layer learns and constructs structure-preserving operators through a self-attention mechanism, and uses these operators to modulate the output features of the shared coding layer element-wise to obtain consistent latent space-preserving features. Maintain characteristics of key latent spaces The expression is: ; in Different from latent space features Structure preservation operator, , To maintain structural strength, To be consistent with latent space characteristics Single-channel spatial importance maps with the same spatial resolution are generated by a self-attention module for latent space features. After averaging the channel dimensions, it is generated by 3×3 convolution and Sigmoid activation; The edge-anchor reinforcement layer identifies key latent space retention features. The set of local maxima anchor points of the edge gradient map is used to perform targeted sharpening of the neighborhood features of the anchor points, outputting key latent space edge features. The expression is: ; ; in This is an edge gradient map. , Preserving characteristics for key latent spaces In respectively , Partial derivative in direction, These are the spatial coordinates on the latent space feature map. To enhance the coefficient, Let be the channel modulation coefficients with the same dimension as the feature vector, derived from the anchor point. The edge direction is generated by 1×1 convolution mapping. For the set of anchor points, The width of the Gaussian decay; The staining enhancement layer preserves the consistency latent space feature. and key latent space edge features Projecting the staining subspace, applying a staining perturbation based on experimental condition residuals, and then projecting back to the original latent space yields consistent latent space staining features. and key latent space staining features The expression is: ; ; ; in , For latent space feature projections on the stained subspace. The projection matrix of the colored subspace. Features of the perturbed subspace, including consistent subspace features. and key color subspace features , Let be the perturbation vector of the colored subspace, which follows a Gaussian normal distribution. The baseline variance of the staining perturbation. This represents the deviation from the current experimental conditions; The artifact elimination layer utilizes prior information from the artifact mask to construct artifacts in the latent space that suppress the output features of the staining enhancement layer by artifact suppression gating, thereby obtaining consistent latent space suppression features. and key latent space inhibition features The expression for artifact suppression gating is: ; in The latent space suppression features after artifact suppression include consistent latent space suppression features. and key latent space inhibition features , For artifact masking Latent space representation after encoder downsampling; The generative enhancement layer performs layered, controllable enhancement of the artifact removal layer in the latent space, and suppresses consistent latent space features. Perform controllable transformations to maintain distance between classes (preserving the commonalities of normal tissue structure in animals of the same family); inhibit key latent space features. It performs intra-class diversity extension transformation (simulating reasonable variations of the same pathological structure under different experimental conditions), employs hierarchical enhancement operators to perform hierarchical controllable enhancement, and outputs consistent latent space enhancement features. and key latent space enhancement features The expression is: ; ; in For hierarchical enhancement operators, Latent space inhibition features, including consistent latent space inhibition features. and key latent space inhibition features , These are explicit transformation parameters, including hue shift, morphological corrosion / expansion intensity, and rotation angle. A lightweight Transformer generator. For cross-attention mechanism, Position encoding embedding for explicit transform parameters; The decoding layer decodes the output features of the generative enhancement layer and outputs the second enhancement block according to the cutting region; In the joint loss function, the structure preservation loss ensures that the pathological structure is not distorted after enhancement, while the loss is based on structural similarity. During calculation, the brightness comparison term only calculates the mean value of the cell nucleus region, while the contrast comparison term weights the edge region; the edge preservation loss only applies to key features, and the edge enhancement coefficient... Typically, a value of 0.2 to 0.5 is used to ensure that the edges of key features are enhanced rather than smoothed; the pathological semantic loss constrains semantic consistency by pre-training the pathological basic model, that is, enhancing the cosine similarity of the pre- and post-cut pieces in the latent space of the pathological basic model; the adversarial loss ensures that the distribution of enhanced features is realistic.

[0022] In this embodiment, the method for obtaining enhanced animal pathological images includes: Enhanced animal pathological images are obtained by replacing and stitching the first standard enhanced block, the second enhanced block, and the corresponding positions of the normalized image according to the pathological topological continuity constraints, key feature embedding constraints, and artifact filling rules. The pathological topological continuity constraint includes: if the pixel difference at the overlapping boundary of adjacent consistent feature blocks is less than a pixel difference threshold, otherwise, linear fusion based on distance weight is performed; The key feature embedding constraint includes: when a unique key feature block is embedded into the background, the boundary gradient difference is less than the gradient difference threshold. The artifact filling rule includes: the artifact mask marked area is filled by radial basis function interpolation of neighboring non-artifact pixels.

[0023] In this embodiment, the method for obtaining annotation information of enhanced animal pathological images includes: Enhanced animal pathological images are input into a few-shot structure annotation network. The network outputs pathological structure-level annotations through segmentation, regression, and topology branches to obtain annotation information for the enhanced animal pathological images. The few-shot structure annotation network adopts an episodic training strategy and performs structural annotation in the latent space through prototype similarity. The pathological structure-level annotations include cell nuclear contours, cytoplasmic contours, gland shapes, vascular branch topology, interstitial density distribution, and intercellular connectivity. The few-shot structure annotation network improves the model's annotation accuracy through segmentation loss, regression loss, and topological loss; In actual evaluation, the few-sample structure labeling network does not make a judgment on the disease results. In the few-sample structure labeling network, the segmentation branch outputs the structural masks of cells, glands and blood vessels, the regression branch outputs the continuous quantity of cell nucleus size and density, and the topology branch outputs the cell adjacency graph and the blood vessel branch connection matrix. The segmentation loss is a weighted sum of the Dice loss and the cross-entropy loss, the regression loss is a smoothed L1 loss for continuous quantities such as cell size and density, and the topological loss is a binary cross-entropy loss for edge prediction of the cell adjacency graph and the blood vessel connection matrix. Few-shot training strategy: Training is performed using enhanced chunks and their corresponding standard template annotations (sparse annotations transferred from public databases). An episode-based training strategy is adopted, with N episodes randomly selected in each round. cThere are 1 structural category, with N for each category. s Support sample and N q Each query sample is labeled using its similarity to its nearest prototype. During training, the pathological semantic perception enhancement network and the few-shot structure annotation network are optimized alternately. When the pathological semantic perception enhancement network is fixed, the annotation ability of the few-shot structure annotation network is trained. When the few-shot structure annotation network is fixed, the enhancement strategy of the pathological semantic perception enhancement network is fine-tuned through annotation feedback (so that the enhanced features are more conducive to structural parsing). The enhanced animal pathology images are input into a few-sample structure annotation network to obtain the annotation information of the enhanced animal pathology images.

[0024] Furthermore, taking the whole liver H&E stained image of a certain batch of experimental rats SP2024-LV-089 (Rattus norvegicus, Muridae, Mammalia) as an example, the specific implementation process of the method of the present invention is illustrated. Artifact detection and tissue region detection were performed on the batch of full-section images. An algorithm based on the fusion of gray-level co-occurrence matrix and local binary mode was used to perform pixel-level annotation of folded, torn, bubble, and scalpel marks, generating an artifact mask. A total of 12 folded regions, 5 torn regions, 8 bubble regions, and 3 scalpel marks were detected in this batch of slices. The number of pixels covered by the artifact mask was about 4.7%. The areas covered by the artifact mask were excluded from the full-section images to obtain the effective tissue region images. The total number of effective tissue region pixels was 15,728,640. Reference slides from the Muridae family (the same family as the target animal) and with the liver as the organ source were extracted from public databases. A first and second mapping relationship were established (the changes in experimental conditions, apparent features, and staining depth were calculated as follows: normalized differences were calculated for continuous variables (concentration, temperature, time); for categorical variables (instrument model, fixative type), Hamming distance or cosine distance was calculated after one-hot encoding, and these were concatenated to form a change vector. The changes in apparent features and staining depth were the standardized differences between the corresponding features and the target reference values). The experimental condition deviation for the current slide batch number SP2024-LV-089 was [missing information]. (Normalized difference) Predicts the apparent feature bias of the current slice based on the first mapping relationship. (Normalized difference) The staining depth bias predicted according to the second mapping relationship is -0.32, of which the apparent feature bias of the hematoxylin channel is -0.15, the apparent feature bias of the eosin channel is 0.08, and the hue bias is 0.05 rad; where the apparent feature bias is the scalar component obtained after dimensionality reduction or principal component projection of the channel mean, and the apparent feature biases of the hematoxylin channel and the eosin channel are scalar optical density scaling biases. The effective tissue region image is transformed from RGB space to optical density space, an adaptive correction matrix based on the predicted appearance feature deviation is applied, and then inversely transformed back to RGB space to obtain a color transformation image. Subsequently, Gamma compensation is performed according to the staining depth deviation. When the depth compensation coefficient is 0.15, the Gamma compensation coefficient is 1.048. Darkening compensation is performed, and a normalized image is obtained after Gamma compensation. The normalized image was quickly divided into several structural regions using the watershed algorithm combined with connected component analysis. The hierarchical feature activation map and gray-level co-occurrence matrix of the pre-trained ResNet-50 were combined to extract the multi-scale structural features of each structural region. Three typical structural regions were selected for illustration: region A-001 is the radial hepatocyte cord region around the central vein of the liver lobule, region A-002 is the abnormal proliferative nodule region of the portal area, and region A-003 is the region of mild dilation of the hepatic sinusoidal space. The Mahalanobis distances between the multi-scale structural feature vectors of each structural region and the mean values ​​of the standard animal slice templates of mouse livers in the public database were calculated to be 2.1 / 8.5 / 7.2, respectively. The standard animal slice templates were determined by statistically analyzing the common structural feature distribution of liver slices from 50 normal Wistar rats. Based on the first distance threshold of 3.0, region A-001 was classified as a consistent feature region, and regions A-002 and A-003 were classified as unique feature regions. Similarly, the remaining structural regions were classified. Taking the unique feature regions A-002 and A-003 as examples, the Euclidean distance between their multi-scale structural feature vectors and the structural feature vectors of annotated regions of the same family of mammals in the public database was calculated. Region A-002 matched with 3 annotated regions, with annotation frequencies of n1=45 (associated with malignancy), n2=30 (associated with inflammation), and n3=25 (associated with benignity), and the total annotation frequency was 100. Region A-003 matched with 2 annotated regions, with annotation frequencies of n1=20 (associated with benignity) and n2=15 (associated with inflammation), and the total annotation frequency was 35. The normalization constant was set to 50, and the annotation confidence scores of the two regions were calculated to be 0.610 and 0.216, respectively. With feature weights of 0.4 / 0.35 / 0.25, the feature quality indices of the A-002 unique feature region (resolution score 0.85, edge sharpness score 0.78, normalized information entropy 0.65) and the A-003 unique feature region (resolution score 0.6, edge sharpness score 0.55, normalized information entropy 0.8) are calculated to be 0.668 / 0.618. Based on these, the criticality judgment products are calculated to be 0.428 and 0.133. According to the criticality threshold of 0.15, the A-002 unique feature region is classified as a region with critical features, and the A-003 unique feature region is classified as a region with non-critical features. Similarly, the remaining unique feature regions are classified. The consistency feature region A-001 was cut according to the natural boundaries of histological structural units such as liver lobules and hepatocyte cords. The block size was fixed at 256×256 pixels, and 50% overlap was set between adjacent blocks. The proportion of consistent feature pixels in a single block was greater than 80%, resulting in a set of 36 consistent feature blocks. Minimum bounding rectangle fitting is performed on the unique key feature region A-002 to obtain a rectangle with a width of 356 pixels and a height of 284 pixels. The rectangle is then expanded outward to obtain an expanded width of 427.2 pixels and an expanded height of 340.8 pixels. The maximum value of the expanded size, 427.2, is taken as the baseline side length and rounded up to the smallest usable size in the preset size range set {64,128,256,512,1024}. The adaptive value of the segment size is determined to be 512 pixels, the proportion of key feature pixels is greater than 30%, and the segment boundary avoids the artifact mask marking area. A set of unique key feature segments is obtained, consisting of 3 segments. The unique non-critical feature region A-003 is cut into non-overlapping segments using a fixed grid with a grid size of 512×512 pixels. After removing grids where the proportion of artifact mask marker pixels exceeds 30%, the remaining parts are directly output to obtain a set of unique non-critical feature segments, totaling 16 segments. Sixteen unique non-critical feature blocks are enhanced using standard methods to obtain 16 first enhanced blocks. Thirty-six consistency feature blocks and three unique critical feature blocks are input into a pathological semantic perception enhancement network (with a structure preservation strength coefficient of 0.15, an edge enhancement coefficient of 0.40, a Gaussian decay width of 2.5, a staining perturbation baseline variance of 0.01, and multi-constraint loss function weights of 0.25 / 0.35 / 0.3 / 0.1, satisfying that the edge preservation loss weight in the critical feature blocks is greater than the structure preservation loss weight). After processing through a shared encoding layer, a structure preservation layer, an edge-anchor enhancement layer, a staining enhancement layer, an artifact removal layer, a generative enhancement layer, and a decoding layer, 36 consistency second enhanced blocks and 3 critical second enhanced blocks are output, for a total of 39 second enhanced blocks. Sixteen first-enhanced slices, 36 consistent second-enhanced slices, and three key second-enhanced slices are stitched together with the normalized image according to the stitching conditions. The pixel difference threshold is set to 10 (grayscale). When the pixel difference at the overlapping boundary of adjacent consistent feature slices is less than 10, they are stitched together directly; otherwise, linear fusion based on distance weights is performed. The gradient difference threshold is set to 15. When the unique key feature slices are embedded in the background, the boundary gradient difference must be less than 15. The artifact mask marked area is filled by radial basis function interpolation of the neighboring non-artifact pixels. The image is then stitched together with the corresponding position in the normalized image to obtain the enhanced animal pathology image. Enhanced animal pathological images were input into a few-shot structure annotation network, and an episode-based training strategy was used for training and inference. In each episode, five structural categories (hepatocyte cords, portal areas, hepatic sinusoids, central venous area, and bile duct epithelium) were randomly selected, with three support samples and two query samples for each category. A total of 10,000 episodes were trained. During the training process, the pathological semantic perception enhancement network and the few-shot structure annotation network were optimized alternately, and the optimization object was switched every 500 episodes. The original segmented regions were structurally annotated. Three typical segmentation examples are as follows: Consistency feature segmentation (number C-012), with annotation information including hepatocyte cord structure mask and average nucleus area of ​​138 μm. 2 The interstitial density is 0.21, and the number of nodes in the cell adjacency graph is 128. Its original multi-scale features include nuclear roundness of 0.84, interstitial density of 0.21, and nearest neighbor cell distance of 11.8 μm. The unique key feature block (number K-002) is labeled with information including the outline of the abnormal proliferating nodules in the portal area and the average area of ​​the cell nuclei of 76 μm. 2 The topology of blood vessel branches (9 branches) and the intercellular connectivity matrix; its original multi-scale features include a nucleocytoplasmic ratio of 0.48, a gland shape factor of 0.71, and a blood vessel branch count of 9; a unique non-critical feature block (number N-005) with annotation information including a mask of a slightly expanded area of ​​the hepatic sinusoidal space and a stroma density of 0.24; its original multi-scale features include a stroma density of 0.24 and a cell orientation entropy of 2.3; The few-sample structure annotation network improves annotation accuracy by using segmentation loss (weighted sum of Dice loss and cross-entropy loss), regression loss (smoothing L1 loss), and topological loss (edge ​​prediction binary cross-entropy loss), ultimately obtaining complete structural-level annotation information for enhanced animal pathological images; the structural-level annotation only describes the shape, size, density, and connectivity of structures such as cells, glands, and blood vessels, and does not make any judgment on the disease outcome.

[0025] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for enhancing and analyzing animal pathological image data based on image processing, characterized in that, Includes the following steps: S1. Perform artifact detection on the whole-section image of H&E stained tissue sections of mammals to generate an artifact mask, and process it to obtain an image of the effective tissue region; S2. Obtain reference slice data from the same organ source location in the same family from the public database, calculate the changes in experimental conditions, changes in appearance features and changes in staining depth, fit to obtain the first mapping relationship and the second mapping relationship, and combine the experimental conditions of the current slice to perform staining normalization on the effective tissue area image to obtain a normalized image. S3. Extract multi-scale structural features from the normalized image, match them with the public database to determine consistent feature regions and unique feature regions, calculate the feature quality index of the unique feature regions, and divide the unique key feature regions and unique non-key feature regions based on the annotations in the public database. S4. According to the differentiated cutting rules, the consistent feature region, the unique key feature region, and the unique non-key feature region are cut into blocks respectively. The unique non-key feature blocks are then enhanced according to the standard to obtain the first enhanced block. S5. Input the consistency feature block and the unique key feature block into the pathological semantic perception enhancement network for adaptive enhancement to obtain the second enhanced block. According to the splicing conditions, the first standard enhanced block, the second enhanced block and the corresponding position of the normalized image are replaced and spliced ​​to obtain the enhanced animal pathological image. S6. Input the enhanced animal pathology images into the few-sample structure annotation network to obtain the annotation information of the enhanced animal pathology images; The reference slice data includes experimental conditions, appearance characteristics, and staining depth; The standard enhancements include random horizontal flipping, random vertical flipping, rotation in multiples of 90°, random perturbation of the HSV color space, and addition of Gaussian noise.

2. The method for enhancing and analyzing animal pathological image data based on image processing according to claim 1, characterized in that, The method for obtaining an effective tissue region image includes: Artifact detection on the whole slice image: An algorithm based on the fusion of gray-level co-occurrence matrix and local binary mode is used to perform pixel-level annotation of fold, tear, bubble and knife mark regions to generate artifact mask; The area covered by the artifact mask is excluded from the full-slice image to obtain an image of the effective tissue region.

3. The animal pathological image data enhancement and analysis method based on image processing according to claim 1, characterized in that, The method for obtaining a normalized image includes: Reference slide data from the same family and organ source location as the target animal were obtained from a public database. The changes in experimental conditions, appearance characteristics, and staining depth among different reference slides were calculated. The staining depth was calculated using optical density. Least squares fitting is performed on the changes in experimental conditions and changes in apparent characteristics to determine the mapping matrix to be estimated and to establish the first mapping relationship. Ridge regression estimation was performed on the changes in experimental conditions and staining depth to determine the mapping vector to be estimated and to establish a second mapping relationship. Extract the experimental conditions of the current slice, select recommended experimental conditions from the same family, genus, and organ source location from the public database to calculate the experimental condition deviation, and predict the appearance feature deviation and staining depth deviation of the current slice based on the first mapping relationship and the second mapping relationship. An adaptive color transformation is performed on the effective tissue region image based on the apparent feature deviation to obtain a color-normalized image. The overall contrast of the color-normalized image is then compensated with Gamma based on the staining depth deviation to obtain a normalized image.

4. The animal pathological image data enhancement and analysis method based on image processing according to claim 1, characterized in that, The method for dividing the region into unique key feature regions and unique non-key feature regions includes: The normalized image is divided into several structural regions, and multi-scale structural features of each structural region are extracted; the multi-scale structural features include cell-level features, tissue-level features, and relation-level features. Calculate the Mahalanobis distance between the multi-scale structural feature vectors of each structural region and the mean of the standard animal slice template in the public database. When the Mahalanobis distance is less than the first distance threshold, the corresponding structural region is classified as a consistent feature region; otherwise, the corresponding structural region is classified as a unique feature region. Calculate the Euclidean distance between the multi-scale structural feature vector of the unique feature region and the structural feature vector of the annotated region of the same family of mammals in the public database. When the Euclidean distance is less than a second distance threshold, match the unique feature region with the annotation information. Calculate the annotation confidence of the unique feature region based on the matching result. The matching result includes the associated disease nature and annotation frequency. Calculate the feature quality index of the unique feature region. When the product of the annotation confidence and the feature quality index is greater than the criticality threshold, the corresponding unique feature region is classified as a unique critical feature region; otherwise, it is classified as a unique non-critical feature region. The feature quality index is positively correlated with the resolution score and the edge sharpness score, and negatively correlated with the information entropy.

5. The animal pathological image data enhancement and analysis method based on image processing according to claim 1, characterized in that, The method for dividing the consistent feature region, the unique key feature region, and the unique non-key feature region into blocks includes: Regions exhibiting consistent characteristics are cut according to the natural boundaries of histological structural units, with a fixed cut size. Adjacent slices are set with 50% overlap, and the proportion of consistent feature pixels within a single slice is greater than 80%, thus obtaining a set of consistent feature slices; the fixed value range of the slice size is 128×128 pixels to 1024×1024 pixels. The unique key feature region is divided by expanding the bounding rectangle of the key structure by 10% and then cutting, with the cut block size adaptive value. By fitting the minimum bounding rectangle, the proportion of key feature pixels is greater than 30%, and the block boundaries avoid the artifact mask marking area, thus obtaining a unique set of key feature blocks; The unique, non-critical feature regions are divided into non-overlapping sections according to a fixed grid, with a fixed grid size of [size missing]. After removing the artifact mask marking area, the remaining part is directly output to obtain a unique set of non-critical feature blocks; the fixed grid size ranges from 256×256 pixels to 2048×2048 pixels. The adaptive value of the cutting size The method for determining the width of the rectangle is as follows: perform minimum bounding rectangle fitting on the unique key feature region to obtain the rectangle width. With height Expand outwards with a preset margin expansion ratio of 0.1 to obtain the expanded width. and the height after expansion Take the maximum value of the expanded size as the base side length, and round it up to the smallest available size in the preset size range set. .

6. The animal pathological image data enhancement and analysis method based on image processing according to claim 1, characterized in that, The method for obtaining the second enhanced slice includes: Consistency feature blocks and unique key feature blocks are input into the pathological semantic perception enhancement network, and processed sequentially through a shared coding layer, a structure preservation layer, an edge-anchor enhancement layer, a staining enhancement layer, an artifact removal layer, a generative enhancement layer, and a decoding layer to output a second enhancement block; the second enhancement block includes a consistency second enhancement block and a key second enhancement block; The pathological semantic perception enhancement network is constrained by a weighted combination of structure preservation loss, edge preservation loss, pathological semantic loss and adversarial loss; The structural preservation loss is the loss of pathological structural similarity between the pre- and post-enhancement sections, expressed as: ; in To preserve structural losses, For the second enhanced cutting Slice from the original input The structural similarity includes brightness and contrast. , For consistency feature index, Indexed for unique key features; The edge preservation loss is: in To preserve structural losses, For the set of anchor points, For the key second enhancement cut block Anchor point Spatial gradient at that location This is the edge enhancement coefficient; The pathological semantic loss is the latent space feature cosine distance loss; The adversarial loss is the discriminator cross-entropy loss.

7. The method for enhancing and analyzing animal pathological image data based on image processing according to claim 1, characterized in that, The method for obtaining enhanced animal pathological images includes: Enhanced animal pathological images are obtained by replacing and stitching the first standard enhanced block, the second enhanced block, and the corresponding positions of the normalized image according to the pathological topological continuity constraints, key feature embedding constraints, and artifact filling rules. The pathological topological continuity constraint includes: if the pixel difference at the overlapping boundary of adjacent consistent feature blocks is less than a pixel difference threshold, otherwise, linear fusion based on distance weight is performed; The key feature embedding constraint includes: when a unique key feature block is embedded into the background, the boundary gradient difference is less than the gradient difference threshold. The artifact filling rule includes: the artifact mask marked area is filled by radial basis function interpolation of neighboring non-artifact pixels.

8. The method for enhancing and analyzing animal pathological image data based on image processing according to claim 1, characterized in that, The method for obtaining annotation information for enhanced animal pathological images includes: Enhanced animal pathological images are input into a few-shot structure annotation network. The network outputs pathological structure-level annotations through segmentation, regression, and topology branches to obtain annotation information for the enhanced animal pathological images. The few-shot structure annotation network adopts an episodic training strategy and performs structural annotation in the latent space through prototype similarity. The pathological structure-level annotations include cell nuclear contours, cytoplasmic contours, gland shapes, vascular branch topology, interstitial density distribution, and intercellular connectivity. The few-shot structure annotation network improves the model's annotation accuracy through segmentation loss, regression loss, and topological loss.