Pathological section image quality evaluation method and computer program product

Abstract

According to the pathological section image quality evaluation method and the computer program product provided by the invention, automatic and objective evaluation of pathological section quality is realized through the deep learning model, and the problems of low efficiency and low recognition rate of an existing manual evaluation system are effectively solved. The dynamic classification strategy can flexibly adapt to different clinical requirements, technical guarantee is provided for accuracy of pathological diagnosis, and the method has remarkable clinical application value.

Description

Technical Field

[0001] This application relates to the field of computer technology. Specifically, it relates to a method for evaluating the quality of pathological section images and a computer program product. Background Art

[0002] The production of pathological sections requires more than ten precise processes, such as specimen collection, fixation, dehydration and clearing, infiltration with paraffin wax, embedding, sectioning, mounting, staining, and coverslipping. Each link is restricted by multiple variable factors: it involves both the operation standardization of technicians and is comprehensively affected by variables such as equipment operation parameters, effective concentration of reagents, and environmental temperature and humidity. In clinical practice, minor deviations in any process may lead to quality defects in the sections, such as tissue shrinkage, abnormal staining, or section fragmentation. It should be noted that poor-quality sections may cause serious consequences such as blurred cell structures and missing key diagnostic areas, directly affecting the accuracy of pathological diagnosis and even inducing the risk of missed diagnosis or misdiagnosis.

[0003] The current quality assessment system mainly relies on the dual-track artificial assessment: the technical department focuses on the evaluation of the section preparation process, while the diagnostic side provides synchronous feedback by pathologists during slide review. The currently widely used sampling in-house quality control (IQC) mode in hospitals can only cover about 5 - 10% of the section production volume, and the recognition rate for minor but clinically significant defects (such as focal detachment, micron-scale knife marks, etc.) is insufficient. Facing a daily production volume of more than a thousand sections, the full-sample quality evaluation requires a complete workload of 3 - 5 person-days, and there is a subjective judgment difference of up to 20% among different evaluators. Therefore, this way of artificially evaluating the quality of pathological sections has problems of low efficiency and low recognition rate. Summary of the Invention

[0004] The purpose of the embodiments of this application is to provide a method for evaluating the quality of pathological section images and a computer program product to solve the problems of low efficiency and low recognition rate in the way of artificially evaluating the quality of pathological sections.

[0005] A method for evaluating the quality of pathological section images provided by the embodiments of this application includes:

[0006] Obtain multiple pathological section images of the same biological tissue, input each pathological section image into a pre-trained quality evaluation model, and obtain the quality evaluation result of each pathological section image;

[0007] Classify the multiple pathological section images according to the quality evaluation results. The pathological section images with a quality evaluation result greater than the first set value are the first type of pathological section images, the pathological section images with a quality evaluation result less than or equal to the first set value and greater than the second set value are the second type of pathological section images, and the pathological section images with a quality evaluation result less than the second set value are the third type of pathological section images; where the first set value is greater than the second set value;

[0008] Use the first type of pathological section image and the second type of pathological section image as available pathological section images.

[0009] In the above technical solution, the automatic and objective evaluation of the quality of pathological sections is realized through a deep learning model, effectively solving the problems of low efficiency and low recognition rate in the current manual evaluation system. Its dynamic classification strategy can flexibly adapt to different clinical needs, providing technical guarantee for the accuracy of pathological diagnosis, and having significant clinical application value.

[0010] Specifically, collect multiple pathological section images (such as HE staining sections) of the same biological tissue. Input each image into a pre-trained quality evaluation model, and output a quality score (such as 0 - 100 points).

[0011] Classification criteria: The first type of pathological section image: score > the first set value (such as 80 points), with excellent quality. The second type of pathological section image: the second set value < score ≤ the first set value (such as 60 - 80 points), with qualified quality. The third type of pathological section image: score ≤ the second set value (such as <60 points), with unqualified quality. The threshold can be dynamically adjusted according to clinical needs, for example, improving the tolerance for rare disease sections.

[0012] Combine the first and second type of images into available pathological section images for subsequent diagnosis. Automatically exclude low-quality images, reduce the workload of manual reexamination, and improve the diagnostic efficiency.

[0013] In some optional embodiments, the training of the quality evaluation model includes:

[0014] Obtain a multi-source pathological section image dataset and perform standardized preprocessing to obtain a dataset, where the dataset contains quality annotation information;

[0015] Use the sliding window segmentation algorithm to segment the whole-slide pathological section into local image patches;

[0016] Construct a feature extraction backbone network based on a deep convolutional neural network to extract multi-scale features from the local image patches;

[0017] Perform channel attention weighted fusion on the multi-scale features to generate a quality score feature vector;

[0018] According to the quality score feature vector and the annotation information, optimize the model parameters through a joint loss function to obtain the final quality evaluation model.

[0019] In the above technical solution, the deep convolutional neural network (CNN) can automatically learn image features. However, the complexity of pathological images (such as multi-scale structures and staining variations) requires the model to have strong feature extraction and fusion capabilities. This method constructs an efficient quality evaluation model through multi-source data normalization, local image patch segmentation, multi-scale feature extraction, and attention fusion.

[0020] In some alternative embodiments, the normalization preprocessing includes:

[0021] Adopt the HE staining normalization algorithm to eliminate color deviations of different scanning devices;

[0022] Strengthen tissue edge and cell nucleus texture features through the Hessian matrix enhancement algorithm;

[0023] Use adaptive histogram equalization to enhance the image contrast.

[0024] In the above technical solution, due to differences in scanning devices (such as Aperio and Hamamatsu), different staining batches, and changes in lighting conditions, there are significant color deviations and insufficient contrast problems in pathological section images. The normalization preprocessing aims to eliminate these interferences, enhance key pathological features (such as cell nuclei and tissue edges), and improve the robustness and accuracy of subsequent analyses (such as quality evaluation and AI diagnosis).

[0025] In some alternative embodiments, the feature extraction backbone network adopts ResNet-50.

[0026] In the above technical solution, in pathological image analysis, ResNet-50 has become a commonly used backbone network due to its strong feature extraction ability and computational efficiency. Its residual connections effectively alleviate the problem of gradient disappearance in deep networks, enabling the model to learn more complex pathological features (such as cell nucleus morphology and tissue texture). By combining multi-scale feature fusion and attention mechanism, the performance of ResNet-50 in pathological tasks can be further improved.

[0027] In some alternative embodiments, channel attention weighted fusion of multi-scale features includes:

[0028] Perform spatial pyramid pooling on the shallow feature map to extract local details;

[0029] Adopt global average pooling on the deep feature map to capture semantic information;

[0030] Dynamically adjust the feature weights through the channel attention module to achieve cross-level feature fusion.

[0031] In the above technical solution, the channel attention module adaptively adjusts the weights according to the input features to avoid the conflict between shallow details and deep semantics. Specifically, in the tumor region, the weight of the deep features increases; in the edge region, the weight of the shallow features increases. The local details extracted by SPP (such as the nuclear morphology) are combined with the global semantics captured by GAP (such as the glandular structure) to improve the lesion recognition ability. The local details extracted by SPP (such as the nuclear morphology) are combined with the global semantics captured by GAP (such as the glandular structure) to improve the lesion recognition ability.

[0032] In some alternative embodiments, the combined loss function consists of the following components:

[0033] The Focal Loss for the classification task, which is used to handle the problem of unbalanced quality levels;

[0034] The Dice Loss for the segmentation task, which optimizes the ability to detect fuzzy regions;

[0035] The contrastive learning loss function, which enhances the feature similarity of similar-quality samples.

[0036] In the above technical solution, the goal of the Focal Loss for the classification task is to handle class imbalance and enhance the attention to difficult-to-classify samples. The goal of the Dice Loss for the segmentation task is to optimize the segmentation of fuzzy regions and improve the overlap between the target and the background. The goal of the contrastive learning loss NT-Xent Loss is to enhance the feature similarity of similar samples and improve the consistency of quality evaluation. The combined loss function realizes the collaborative optimization of classification, segmentation, and feature learning by integrating the Focal Loss, Dice Loss, and contrastive learning loss.

[0037] In some alternative embodiments, after using the first type of pathological section image and the second type of pathological section image as available pathological section images, it further includes:

[0038] For the first type of pathological section image, high-quality slice features of different tissue regions are extracted;

[0039] For the third type of pathological section image, the local regions of the image are matched using the high-quality slice features, and the effective regions are screened out;

[0040] The effective regions of the third type of pathological section image are also used as available pathological section images.

[0041] In the above technical solution, on the basis of directly using the first and second types of slices as available images, the potential diagnostic value of the third type of low-quality slices is further explored. Although the overall quality of the third type of slices (score ≤ the second set value) is poor, they may contain local high-quality regions (such as cell clusters not affected by detachment or knife marks). Through feature migration technology, high-quality features of the first type of slices can be extracted and used for local area screening of the third type of slices, improving resource utilization.

[0042] Specifically, for the first type of pathological slice images (score > 80 points), different tissue regions (such as epithelial tissue, connective tissue, tumor cell clusters) are identified and labeled through a semantic segmentation model (such as U-Net++). Geometric features such as the nuclear morphology and the regularity of tissue arrangement are extracted (such as Hessian matrix texture analysis). The contrast, saturation, and consistency of HE staining are quantified (such as RGB channel histogram statistics). The above features are encoded into a high-dimensional vector (such as 512 dimensions) as a high-quality feature template.

[0043] The third type of pathological slice images are divided into non-overlapping small blocks (such as 256×256 pixels), and the feature vector of each block is calculated. The cosine similarity or Euclidean distance is used to measure the similarity between the small blocks of the third type of slices and the first type of feature template. A threshold is set (such as similarity > 0.7) to screen out the effective regions. Connected component analysis is performed on adjacent effective small blocks to merge them into a complete effective region.

[0044] A binary mask is generated according to the screening result to mark the effective regions. The mask is multiplied by the original third type of slice image to extract the effective regions. Contrast enhancement and noise suppression are performed on the effective regions (such as non-local means filtering). The processed effective regions are used as supplementary available pathological slice images and output together with the first and second type of images.

[0045] In some alternative embodiments, for the first type of pathological slice images, high-quality slice features of different tissue regions are extracted, including:

[0046] Different tissue regions are separated from the first type of pathological slice images through a segmentation algorithm;

[0047] A pre-trained pathological feature extraction model is used to perform multi-scale feature extraction on the segmented tissue regions;

[0048] The extracted features are standardized and dimension-reduced to generate high-quality slice features in the format of numerical feature vectors.

[0049] In the above technical solution, in pathological slice analysis, different tissue regions (such as tumor cells, stroma, necrosis regions) have unique morphological and biological characteristics. To accurately extract high-quality features of the first type of pathological slice images (score > 80 points), it is necessary to combine tissue segmentation, multi-scale feature learning, and dimensionality reduction techniques to generate quantifiable and highly discriminative feature vectors.

[0050] Specifically, a pre-trained U-Net++ or DeepLabV3+ model is used. The first type of pathological slice image (such as 2048×2048 pixels) is input, and pixel-level tissue class labels (such as epithelium, stroma, blood vessels, necrosis) are output. Conditional random fields (CRFs) or morphological operations (such as opening and closing operations) are applied to eliminate isolated noise points and smooth the region boundaries. A segmentation mask of the same size as the original image is generated, and each pixel corresponds to a tissue class.

[0051] ResNet-50 or EfficientNet-B3 is used as the backbone, the classification head is removed, and the feature pyramid structure is retained. Each segmented tissue region is cropped into multiple scales (such as 256×256, 512×512, 1024×1024 pixels) to simulate pathological fields at different magnifications. Multilevel pooling (such as 1×1, 2×2, 4×4 grids) is performed on the feature maps of each scale to fuse global and local information. An SE module or a CBAM module is introduced to adaptively weight important feature channels. A multi-scale feature tensor (such as C×H×W, C = 2048 channels) is generated for each tissue region.

[0052] The feature tensor is normalized in the channel dimension to eliminate the dimension difference. The feature vector is standardized with a mean of 0 and a variance of 1 to improve the model convergence. The principal components that retain 95% of the variance energy are retained, and the feature dimension is reduced from 2048 to 512 to generate a fixed-length numerical feature vector for support vector retrieval and similarity calculation.

[0053] In some alternative embodiments, for the third type of pathological slice images, high-quality slice features are used to match local regions of the images, and effective regions are screened out, including:

[0054] Different tissue regions are separated from the third type of pathological slice images through a segmentation algorithm;

[0055] A pre-trained pathological feature extraction model is used to perform multi-scale feature extraction on the segmented tissue regions;

[0056] The extracted features are standardized and dimensionally reduced to generate slice features to be matched in the format of numerical feature vectors;

[0057] For each tissue region, calculate the similarity between the corresponding high-quality slice features and the corresponding slice features to be matched. If the similarity is higher than the threshold, the tissue region corresponding to the slice features to be matched is regarded as a valid region.

[0058] The third type of pathological slice (score ≤ the second set value) is often discarded due to poor overall quality (such as uneven staining, knife marks or detachment), but its local region may contain key diagnostic information. By using the high-quality features of the first type of slice (high score quality) to perform feature matching on the local regions of the third type of slice, valid regions can be accurately screened out, improving resource utilization and diagnostic accuracy.

[0059] Specifically, use a pre-trained U-Net++ or DeepLabV3+ model to input the third type of pathological slice image (such as 2048×2048 pixels) and output pixel-level tissue category labels (such as epithelium, stroma, blood vessels, necrosis). Apply conditional random fields (CRF) or morphological operations (such as opening and closing operations) to eliminate isolated noise points and smooth the region boundaries. Generate a segmentation mask of the same size as the original image, with each pixel corresponding to a tissue category.

[0060] Use ResNet-50 or EfficientNet-B3 as the backbone, remove the classification head, and retain the feature pyramid structure. For each tissue region after segmentation, crop it into multiple scales (such as 256×256, 512×512, 1024×1024 pixels) to simulate the pathological fields at different magnifications. Perform multi-level pooling (such as 1×1, 2×2, 4×4 grids) on the feature maps of each scale to fuse global and local information. Introduce an SE module or a CBAM module to adaptively weight important feature channels. Generate multi-scale feature tensors (such as C×H×W, C = 2048 channels) for each tissue region.

[0061] Normalize the feature tensors in the channel dimension to eliminate the dimension difference. Standardize the feature vectors with a mean of 0 and a variance of 1 to improve the model convergence. Retain the principal components with 95% variance energy, reduce the feature dimension from 2048 to 512, and generate a fixed-length numerical feature vector to support vector retrieval and similarity calculation.

[0062] For each tissue region of the third type of slice, retrieve the high-quality feature vectors of the same tissue type in the first type of slice (such as matching tumor features for the tumor region).

[0063] Similarity calculation: Use cosine similarity or Euclidean distance to measure the similarity between the features to be matched (third type) and the high-quality features (first type).

[0064] Set a dynamic threshold (such as cosine similarity > 0.7). If the similarity is higher than the threshold, mark the tissue region as a valid region.

[0065] An electronic device provided by an embodiment of the present application includes: a processor and a memory. The memory stores machine-readable instructions executable by the processor. When the machine-readable instructions are executed by the processor, the method described in any of the above is executed.

[0066] A computer program product provided by an embodiment of the present application includes a computer program / instructions. When the computer program / instructions are executed by a processor, the steps of the method described in any of the above are implemented. Description of the Drawings

[0067] To more clearly illustrate the technical solutions of the embodiments of the present application, the drawings required to be used in the embodiments of the present application will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present application and should not be regarded as limiting the scope. For those of ordinary skill in the art, without creative efforts, other relevant drawings can also be obtained based on these drawings.

[0068] Figure 1 It is a flowchart of the steps of a method for evaluating the quality of a pathological section image provided by an embodiment of the present application;

[0069] Figure 2 It is a flowchart of the steps of a method for evaluating the quality of a pathological section image provided by another embodiment of the present application;

[0070] Figure 3 It is a possible structural schematic diagram of the electronic device provided by the embodiment of the present application. Detailed Embodiments

[0071] The technical solutions in the embodiments of the present application will be described below in conjunction with the drawings in the embodiments of the present application.

[0072] Please refer to Figure 1 , Figure 1 It is a flowchart of the steps of a method for evaluating the quality of a pathological section image provided by an embodiment of the present application, including:

[0073] Step 100: Obtain multiple pathological section images of the same biological tissue, and input each pathological section image into a pre-trained quality evaluation model to obtain the quality evaluation result of each pathological section image;

[0074] Step 200: Classify the multiple pathological section images according to the quality evaluation results. The quality evaluation result greater than the first set value is the first type of pathological section image, the quality evaluation result less than or equal to the first set value and greater than the second set value is the second type of pathological section image, and the quality evaluation result less than the second set value is the third type of pathological section image; wherein, the first set value is greater than the second set value;

[0075] Step 300: Use the first type of pathological section image and the second type of pathological section image as available pathological section images.

[0076] In the embodiments of the present application, the automatic and objective evaluation of the quality of pathological sections is realized through a deep learning model, effectively solving the problems of low efficiency and low recognition rate of the current manual evaluation system. Its dynamic classification strategy can be flexibly adapted to different clinical needs, providing technical guarantee for the accuracy of pathological diagnosis, and having significant clinical application value.

[0077] Specifically, collect multiple pathological section images (such as HE-stained sections) of the same biological tissue. Input each image into a pre-trained quality evaluation model, and output a quality score (such as 0 - 100 points).

[0078] Classification criteria: The first type of pathological section image: score > the first set value (such as 80 points), with excellent quality. The second type of pathological section image: the second set value < score ≤ the first set value (such as 60 - 80 points), with qualified quality. The third type of pathological section image: score ≤ the second set value (such as <60 points), with unqualified quality. The threshold can be dynamically adjusted according to clinical needs, for example, increasing the tolerance for rare disease sections.

[0079] Combine the first and second type of images into available pathological section images for subsequent diagnosis. Automatically exclude low-quality images, reduce the workload of manual reexamination, and improve the diagnosis efficiency.

[0080] In some alternative embodiments, the training of the quality evaluation model includes:

[0081] Obtain a multi-source pathological section image dataset and perform standardized preprocessing to obtain a dataset, which contains quality annotation information;

[0082] Use the sliding window segmentation algorithm to segment the whole-slide pathological section into local image patches;

[0083] Construct a feature extraction backbone network based on a deep convolutional neural network to extract multi-scale features from the local image patches;

[0084] Perform channel attention weighted fusion on the multi-scale features to generate a quality score feature vector;

[0085] According to the quality score feature vector and the annotation information, optimize the model parameters through a joint loss function to obtain the final quality evaluation model.

[0086] In the embodiments of the present application, the deep convolutional neural network (CNN) can automatically learn image features. However, the complexity of pathological images (such as multi-scale structures and staining variations) requires the model to have strong feature extraction and fusion capabilities. This method constructs an efficient quality evaluation model through multi-source data standardization, local image block segmentation, multi-scale feature extraction, and attention fusion.

[0087] In some alternative embodiments, the standardization preprocessing includes:

[0088] Adopt the HE staining normalization algorithm to eliminate the color deviation of different scanning devices;

[0089] Enhance the tissue edge and nuclear texture features through the Hessian matrix enhancement algorithm;

[0090] Use adaptive histogram equalization to enhance the image contrast.

[0091] In the embodiments of the present application, due to differences in scanning devices (such as Aperio and Hamamatsu), different staining batches, and changes in lighting conditions, there are significant color deviation and insufficient contrast problems in pathological section images. The standardization preprocessing aims to eliminate these interferences, enhance key pathological features (such as cell nuclei and tissue edges), and improve the robustness and accuracy of subsequent analyses (such as quality evaluation and AI diagnosis).

[0092] In some alternative embodiments, the feature extraction backbone network uses ResNet-50.

[0093] In the embodiments of the present application, in pathological image analysis, ResNet-50 has become a commonly used backbone network due to its strong feature extraction ability and computational efficiency. Its residual connections effectively alleviate the problem of gradient disappearance in deep networks, enabling the model to learn more complex pathological features (such as cell nucleus morphology and tissue texture). By combining multi-scale feature fusion and attention mechanism, the performance of ResNet-50 in pathological tasks can be further improved.

[0094] In some alternative embodiments, the channel attention weighted fusion of multi-scale features includes:

[0095] Perform spatial pyramid pooling on the shallow feature maps to extract local details;

[0096] Adopt global average pooling on the deep feature maps to capture semantic information;

[0097] Dynamically adjust the feature weights through the channel attention module to achieve cross-level feature fusion.

[0098] In the embodiments of the present application, the channel attention module adaptively adjusts the weights according to the input features to avoid the conflict between shallow details and deep semantics. Specifically, in the tumor region, the weight of the deep features increases; in the edge region, the weight of the shallow features increases. The local details extracted by SPP (such as the morphology of cell nuclei) are combined with the global semantics captured by GAP (such as glandular structures) to improve the lesion recognition ability. The local details extracted by SPP (such as the morphology of cell nuclei) are combined with the global semantics captured by GAP (such as glandular structures) to improve the lesion recognition ability.

[0099] In some alternative embodiments, the joint loss function is composed of the following components:

[0100] The Focal Loss for the classification task, which is used to handle the problem of unbalanced quality levels;

[0101] The Dice Loss for the segmentation task, which optimizes the ability to detect fuzzy regions;

[0102] The contrastive learning loss function, which enhances the feature similarity of samples of the same quality.

[0103] In the embodiments of the present application, the goal of the Focal Loss for the classification task is to handle class imbalance and enhance the attention to difficult-to-classify samples. The goal of the Dice Loss for the segmentation task is to optimize the segmentation of fuzzy regions and improve the overlap between the target and the background. The goal of the contrastive learning loss NT-Xent Loss is to enhance the feature similarity of samples of the same type and improve the consistency of quality evaluation. The joint loss function realizes the collaborative optimization of classification, segmentation, and feature learning by integrating the Focal Loss, Dice Loss, and contrastive learning loss.

[0104] Please refer to Figure 2 , Figure 2 which is a flowchart of the steps of a method for evaluating the quality of a pathological slice image provided in another embodiment of the present application.

[0105] In some alternative embodiments, after using the first type of pathological slice image and the second type of pathological slice image as available pathological slice images, it further includes:

[0106] Step 400: Extract high-quality slice features of different tissue regions from the first type of pathological slice image;

[0107] Step 500: Use the high-quality slice features to match the local regions of the third type of pathological slice image, and filter out the effective regions;

[0108] Step 600: Also use the effective regions of the third type of pathological slice image as available pathological slice images.

[0109] In the embodiments of the present application, on the basis of directly using the first and second types of slices as available images, the potential diagnostic value of the third type of low-quality slices is further explored. Although the overall quality of the third type of slices (score ≤ the second set value) is poor, they may contain local high-quality regions (such as cell clusters not affected by detachment or knife marks). Through feature migration technology, high-quality features of the first type of slices can be extracted and used for screening local regions of the third type of slices, improving resource utilization.

[0110] Specifically, for the first type of pathological slice images (score > 80 points), different tissue regions (such as epithelial tissue, connective tissue, tumor cell clusters) are identified and labeled through a semantic segmentation model (such as U-Net++). Geometric features such as nuclear morphology and tissue arrangement regularity are extracted (such as Hessian matrix texture analysis). The contrast, saturation, and consistency of HE staining are quantified (such as RGB channel histogram statistics). The above features are encoded into high-dimensional vectors (such as 512-dimensional) as high-quality feature templates.

[0111] The third type of pathological slice images are divided into non-overlapping small blocks (such as 256×256 pixels), and the feature vector of each block is calculated. The cosine similarity or Euclidean distance is used to measure the similarity between the small blocks of the third type of slices and the first type of feature template. A threshold is set (such as similarity > 0.7) to screen out valid regions. Connected component analysis is performed on adjacent valid small blocks to merge and form complete valid regions.

[0112] A binary mask is generated according to the screening result to mark the valid regions. The mask is multiplied by the original third type of slice image to extract the valid regions. Contrast enhancement and noise suppression are performed on the valid regions (such as non-local means filtering). The processed valid regions are used as supplementary available pathological slice images and output together with the first and second type of images.

[0113] In some alternative embodiments, for the first type of pathological slice images, high-quality slice features of different tissue regions are extracted, including:

[0114] Different tissue regions are separated from the first type of pathological slice images through a segmentation algorithm;

[0115] A pre-trained pathological feature extraction model is used to perform multi-scale feature extraction on the segmented tissue regions;

[0116] The extracted features are standardized and dimension-reduced to generate high-quality slice features in the format of numerical feature vectors.

[0117] In the embodiments of the present application, in pathological slice analysis, different tissue regions (such as tumor cells, stroma, necrosis regions) have unique morphological and biological characteristics. To accurately extract high-quality features of the first type of pathological slice images (score > 80 points), it is necessary to combine tissue segmentation, multi-scale feature learning, and dimensionality reduction techniques to generate quantifiable and highly discriminative feature vectors.

[0118] Specifically, a pre-trained U-Net++ or DeepLabV3+ model is used. The first type of pathological slice images (such as 2048×2048 pixels) are input, and pixel-level tissue category labels (such as epithelium, stroma, blood vessels, necrosis) are output. Conditional random fields (CRFs) or morphological operations (such as opening and closing operations) are applied to eliminate isolated noise points and smooth the region boundaries. A segmentation mask of the same size as the original image is generated, and each pixel corresponds to a tissue category.

[0119] ResNet-50 or EfficientNet-B3 is used as the backbone, the classification head is removed, and the feature pyramid structure is retained. For each segmented tissue region, it is cropped into multiple scales (such as 256×256, 512×512, 1024×1024 pixels) to simulate pathological fields at different magnifications. Multi-level pooling (such as 1×1, 2×2, 4×4 grids) is performed on the feature maps at each scale to fuse global and local information. The SE module or CBAM module is introduced to adaptively weight important feature channels. A multi-scale feature tensor (such as C×H×W, C = 2048 channels) is generated for each tissue region.

[0120] The feature tensor is normalized in the channel dimension to eliminate the dimension difference. The feature vector is standardized with a mean of 0 and a variance of 1 to improve the model convergence. The principal components that retain 95% of the variance energy are retained, and the feature dimension is reduced from 2048 to 512 to generate a fixed-length numerical feature vector to support vector retrieval and similarity calculation.

[0121] In some alternative embodiments, for the third type of pathological slice images, high-quality slice features are used to match local regions of the images to screen out effective regions, including:

[0122] Different tissue regions are separated from the third type of pathological slice images through a segmentation algorithm;

[0123] A pre-trained pathological feature extraction model is used to perform multi-scale feature extraction on the segmented tissue regions;

[0124] The extracted features are standardized and dimensionally reduced to generate slice features to be matched in the format of a numerical feature vector;

[0125] For each tissue region, calculate the similarity between the corresponding high-quality slice features and the corresponding slice features to be matched. If the similarity is higher than the threshold, the tissue region corresponding to the slice features to be matched is taken as the effective region.

[0126] The third type of pathological slice (score ≤ the second set value) is often discarded due to poor overall quality (such as uneven staining, knife marks or detachment), but its local area may contain key diagnostic information. By using the high-quality features of the first type of slice (high score quality) to perform feature matching on the local area of the third type of slice, the effective region can be accurately screened out, improving resource utilization and diagnostic accuracy.

[0127] Specifically, use a pre-trained U-Net++ or DeepLabV3+ model. Input the third type of pathological slice image (such as 2048×2048 pixels) and output pixel-level tissue category labels (such as epithelium, stroma, blood vessel, necrosis). Apply conditional random field (CRF) or morphological operations (such as opening and closing operations) to eliminate isolated noise points and smooth the region boundaries. Generate a segmentation mask of the same size as the original image, with each pixel corresponding to a tissue category.

[0128] Use ResNet-50 or EfficientNet-B3 as the backbone, remove the classification head, and retain the feature pyramid structure. For each tissue region after segmentation, crop it into multiple scales (such as 256×256, 512×512, 1024×1024 pixels) to simulate the pathological fields at different magnifications. Perform multi-level pooling (such as 1×1, 2×2, 4×4 grids) on the feature maps of each scale to fuse global and local information. Introduce the SE module or CBAM module to adaptively weight important feature channels. Generate multi-scale feature tensors (such as C×H×W, C = 2048 channels) for each tissue region.

[0129] Normalize the feature tensors in the channel dimension to eliminate the dimension difference. Standardize the feature vectors with a mean of 0 and a variance of 1 to improve the model convergence. Retain the principal components with 95% variance energy, reduce the feature dimension from 2048 to 512, and generate a fixed-length numerical feature vector to support vector retrieval and similarity calculation.

[0130] For each tissue region of the third type of slice, retrieve the high-quality feature vectors of the same tissue type in the first type of slice (such as matching the tumor region with tumor features).

[0131] Similarity calculation: Use cosine similarity or Euclidean distance to measure the similarity between the features to be matched (the third type) and the high-quality features (the first type).

[0132] Set a dynamic threshold (such as cosine similarity > 0.7). If the similarity is higher than the threshold, mark this tissue region as the effective region.

[0133] Figure 3 shows a possible structure of an electronic device provided by an embodiment of the present application. Referring to Figure 3 , the electronic device includes: a processor, a memory, and a communication interface, and these components are interconnected and communicate with each other through a communication bus and / or other forms of connection mechanisms (not shown).

[0134] Among them, the memory includes one or more (only one is shown in the figure), which can be, but is not limited to, a random access memory (Random Access Memory, abbreviated as RAM), a read-only memory (Read Only Memory, abbreviated as ROM), a programmable read-only memory (Programmable Read-Only Memory, abbreviated as PROM), an erasable programmable read-only memory (Erasable Programmable Read-Only Memory, abbreviated as EPROM), an electrically erasable programmable read-only memory (Electric Erasable Programmable Read-Only Memory, abbreviated as EEPROM), etc. The processor and other possible components can access the memory, read and / or write the data therein.

[0135] The processor includes one or more (only one is shown in the figure), which can be an integrated circuit chip with signal processing capabilities. The above-mentioned processor can be a general-purpose processor, including a central processing unit (Central Processing Unit, abbreviated as CPU), a microcontroller unit (Micro Controller Unit, abbreviated as MCU), a network processor (Network Processor, abbreviated as NP), or other conventional processors; it can also be a dedicated processor, including a neural network processor (Neural-network Processing Unit, abbreviated as NPU), a graphics processing unit (Graphics Processing Unit, abbreviated as GPU), a digital signal processor (Digital Signal Processor, abbreviated as DSP), an application specific integrated circuit (Application Specific Integrated Circuits, abbreviated as ASIC), a field programmable gate array (Field Programmable Gate Array, abbreviated as FPGA), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components. And when there are multiple processors, a part of them can be general-purpose processors, and another part can be dedicated processors.

[0136] The communication interface includes one or more (only one is shown in the figure), which can be used to communicate directly or indirectly with other devices for data interaction. The communication interface can include interfaces for wired and / or wireless communication.

[0137] One or more computer program instructions can be stored in the memory, and the processor can read and run these computer program instructions to implement the method provided by the embodiments of the present application.

[0138] It can be understood that Figure 3 The structure shown is only schematic, and the electronic device may further include more or fewer components than those shown in Figure 3 or have a different structure from that shown in Figure 3 shown. Figure 3 Each component shown in can be implemented by hardware, software, or a combination thereof. The electronic device may be a physical device, such as a PC, laptop, tablet, mobile phone, server, embedded device, etc., or may be a virtual device, such as a virtual machine, virtualization container, etc. Moreover, the electronic device is not limited to a single device, and may also be a combination of multiple devices or a cluster composed of a large number of devices.

[0139] A computer program product provided by an embodiment of the present application includes a computer program / instructions, and when the computer program / instructions are executed by a processor, the steps of any of the above methods are implemented.

[0140] In the embodiments provided in the present application, it should be understood that the disclosed device and method can be implemented in other ways. The device embodiments described above are only illustrative. For example, the division of the units is only a logical function division, and there may be other division methods in actual implementation. For another example, multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point, the displayed or discussed mutual coupling or direct coupling or communication connection may be through some communication interfaces, and the indirect coupling or communication connection of the device or unit may be in an electrical, mechanical or other form.

[0141] In addition, the units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed to multiple network units. Some or all of the units can be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0142] Furthermore, in each embodiment of the present application, the functional modules can be integrated together to form an independent part, or each module can exist alone, or two or more modules can be integrated to form an independent part.

[0143] In this document, relational terms such as first and second are used solely to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations.

[0144] The above description is only for the embodiments of the present application and is not intended to limit the protection scope of the present application. For those skilled in the art, the present application may have various modifications and changes. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included within the protection scope of the present application.

Claims

1. A method for evaluating the quality of a pathological section image, characterized in that, Including: Obtain multiple pathological section images of the same biological tissue, input each pathological section image into a pre-trained quality evaluation model, and obtain the quality evaluation result of each pathological section image; Classify the multiple pathological section images according to the quality evaluation results. The quality evaluation result greater than the first set value is the first type of pathological section image, the quality evaluation result less than or equal to the first set value and greater than the second set value is the second type of pathological section image, and the quality evaluation result less than the second set value is the third type of pathological section image; wherein, the first set value is greater than the second set value; Use the first type of pathological section image and the second type of pathological section image as available pathological section images.

2. The method according to claim 1, wherein The training of the quality evaluation model includes: Obtain a multi-source pathological section image dataset and perform standardized preprocessing to obtain a dataset, and the dataset contains quality annotation information; Segment the whole-field pathological section into local image patches through a sliding window segmentation algorithm; Construct a feature extraction backbone network based on a deep convolutional neural network, and extract multi-scale features from the local image patches; Perform channel attention weighted fusion on the multi-scale features to generate a quality score feature vector; According to the quality score feature vector and the annotation information, optimize the model parameters through a joint loss function to obtain the final quality evaluation model.

3. The method according to claim 2, characterized in that The standardized preprocessing includes: Adopt the HE staining normalization algorithm to eliminate the color deviation of different scanning devices; Enhance the tissue edge and cell nucleus texture features through the Hessian matrix enhancement algorithm; Use adaptive histogram equalization to improve the image contrast.

4. The method according to claim 2, characterized in that The feature extraction backbone network adopts ResNet-50.

5. The method according to claim 2, characterized in that, The channel attention weighted fusion of the multi-scale features includes: Perform spatial pyramid pooling on the shallow feature maps to extract local details; Adopt global average pooling on the deep feature maps to capture semantic information; Dynamically adjust the feature weights through a channel attention module to achieve cross-level feature fusion.

6. The method according to claim 2, wherein The joint loss function is composed of the following components: The Focal Loss of the classification task, which is used to handle the problem of unbalanced quality levels; The Dice Loss of the segmentation task, which optimizes the fuzzy region detection ability; The contrastive learning loss function, which enhances the feature similarity of the same-class quality samples.

7. The method according to claim 1, characterized in that, After using the first type of pathological section image and the second type of pathological section image as available pathological section images, it further includes: For the first type of pathological section image, extract the high-quality slice features of different tissue regions; For the third type of pathological section image, use the high-quality slice features to match the local regions of the image, and screen out the effective regions; Also use the effective regions of the third type of pathological section image as available pathological section images.

8. The method according to claim 7, wherein For the first type of pathological section image, extracting the high-quality slice features of different tissue regions includes: Separate different tissue regions from the first type of pathological section image through a segmentation algorithm; Use a pre-trained pathological feature extraction model to perform multi-scale feature extraction on the segmented tissue regions; Standardize and reduce the dimension of the extracted features to generate high-quality slice features in the format of a numerical feature vector.

9. The method according to claim 8, wherein For the third type of pathological section images, the local regions of the images are matched using high-quality section features, and the effective regions are screened out, including: Separate different tissue regions from the third type of pathological section images through a segmentation algorithm; Use a pre-trained pathological feature extraction model to perform multi-scale feature extraction on the segmented tissue regions; Normalize and reduce the dimensions of the extracted features to generate the section features to be matched in the format of numerical feature vectors; For each tissue region, calculate the similarity between the corresponding high-quality section features and the corresponding section features to be matched. If the similarity is higher than the threshold, the tissue region corresponding to the section features to be matched is used as the effective region.

10. A computer program product, comprising a computer program / instructions, characterized in that, When the computer program / instructions are executed by a processor, the steps of the method according to any one of claims 1-9 are implemented.

Images