An artificial intelligence-based tumor pathological image classification method and system

Abstract

This invention provides an artificial intelligence-based tumor pathology image classification method, belonging to the field of image analysis technology. The method includes the following steps: acquiring tumor pathology image data and performing staining and normalization processing on the tumor pathology image data; extracting multi-scale features from the processed image data using a multi-branch parallel feature extraction network and fusing the features to obtain a multi-scale feature map; identifying and locating the tumor lesion region based on the multi-scale feature map and segmenting the lesion region from the image; classifying the segmented lesion region according to the lesion type; analyzing the lesion grade of the category after classification and outputting a classification grade diagnostic report; achieving more refined classification, further providing doctors with more accurate diagnostic basis, and improving the accuracy and reliability of pathological diagnosis.

Description

Technical Field

[0001] This invention relates to the field of image analysis technology, and in particular to an artificial intelligence-based method and system for classifying tumor pathological images. Background Technology

[0002] In recent years, with the continuous development of computer science and information technology, medical imaging technology has also developed rapidly, and various new imaging devices have emerged, such as computed tomography (CT), digital subtraction angiography (DSA), single-photon emission computed tomography (SPECT), and magnetic resonance imaging (MRI). In prostate diagnosis and treatment, with the continuous development of medical technology, tumor pathology diagnostic images are increasingly widely used in the medical field. Doctors often need to obtain similar prostate cancer pathology diagnostic images based on the patient's images for further analysis. Reference diagnosis helps doctors understand the patient's condition.

[0003] Patent publication number CN118154975B discloses a tumor pathology diagnostic image classification method based on big data, involving the field of image classification technology. It solves the technical problem that classifying solely based on lesion type and severity fails to consider the crucial importance of prostate cancer lesion location and classification for treatment selection and prognostic assessment, hindering doctors from quickly obtaining similar tumor pathology diagnostic images for reference based on lesion location. The method obtains first, second, and third classifications of pathology images based on lesion type, primary lesion location, and lesion index. It further classifies prostate cancer pathology diagnostic images according to the location of the lesion area within various parts of the prostate. Combining features such as pathology type, lesion location, and lesion size, the method provides doctors with more accurate diagnostic information.

[0004] However, the existing technology only segments lesion areas based on pixel binarization, which is susceptible to staining differences and artifacts, resulting in poor robustness. Moreover, it only focuses on the number of pixels, ignoring microscopic features such as cell nuclear morphology and glandular structure, making it difficult to support fine grading. Summary of the Invention

[0005] In view of the above-mentioned prior art, the present invention provides a tumor pathology image classification method and system based on artificial intelligence, which mainly solves the technical problems existing in the background art.

[0006] To achieve the above objectives, the technical solution of this invention is implemented as follows: In a first aspect, this application provides an artificial intelligence-based tumor pathology image classification method, the method comprising the following steps: Tumor pathological image data is acquired and stained and normalized. The processed image data is then subjected to multi-scale feature extraction using a multi-branch parallel feature extraction network, and feature fusion is performed to obtain multi-scale feature maps. The tumor lesion region is identified and located based on the multi-scale feature map, and the lesion region is segmented from the image; The segmented lesion areas are classified according to the lesion type. After classification, the lesion grade of the category is analyzed, and a classification grade diagnosis report is output.

[0007] As a preferred embodiment of this application, the staining normalization processing of tumor pathology image data specifically includes: K-means clustering analysis was performed on tumor pathology image data in color space to obtain several staining clusters; based on the color consistency index within the clusters and the diagnostic readability score given by the pathologist, the optimal cluster image was selected as the standard staining candidate. Color statistics are performed on the optimal cluster image to obtain the mean vector and covariance matrix of the reference color distribution, and a standard coloring space is constructed. Pattern transfer is used to map tumor pathology image data from source staining to standard staining space; and a pattern separation loss function is used to ensure that diagnostic biological tissue information is preserved while normalizing.

[0008] As a preferred embodiment of this application, the processed image data is subjected to multi-scale feature extraction using a multi-branch parallel feature extraction network, and feature fusion is performed to obtain multi-scale image data, specifically including: The feature extraction network includes four parallel branches, which process image regions at different resolutions respectively; Each branch uses an improved ResNet-50 architecture as the backbone network and embeds a compressed excitation attention module in each residual block, enabling the network to adaptively recalibrate the channel feature response. The branch also employs an expanded convolution strategy to gradually expand the receptive field, ensuring that a wider range of contextual information is captured while maintaining resolution. Then, through the cross-scale feature interaction module, a bidirectional feature pyramid network is used to fuse features of different scales to obtain a multi-scale feature map.

[0009] As a preferred embodiment of this application, identifying tumor lesion regions based on multi-scale feature maps specifically includes: The visual index of each image region in the multi-scale feature map is calculated to quantify the degree of deviation between the target region and the surrounding normal tissue in the feature space. Then, by setting an adaptive threshold, candidate regions with significant abnormal features are initially screened out. Based on pixel-level and region-level features, a probabilistic graphical model is used to identify lesion regions in candidate regions. Specifically, pixels that are spatially adjacent and have similar features are assigned the same label. For each candidate region, a series of quantitative attributes are calculated to verify and filter out false positives. The quantitative attributes are compared with the standard parameter range in the predefined tumor lesion region knowledge base to eliminate candidate regions whose qualitative attribute features do not clearly conform to the typical manifestations of malignant tumors. Finally, through iterative reasoning, an optimal binary segmentation mask is generated, in which high-probability regions are identified as tumor lesion regions.

[0010] As a preferred embodiment of this application, the location of the tumor lesion area includes: A subpixel-level edge detection algorithm is used to extract the boundary contour of the tumor lesion area, and the minimum bounding rectangle or fitted ellipse of the contour is calculated. The lesion area is then spatially located in the coordinate system of the original pathological image in the form of bounding box coordinates or contour point set coordinate sequence.

[0011] As a preferred embodiment of this application, the segmented lesion areas are classified according to the lesion type, specifically including: Within the located lesion area, firstly, the global morphological features of the area are calculated, including the area, perimeter, compactness, and invasiveness index at the junction with the surrounding normal tissue. Secondly, cellular-scale microscopic features are extracted, including the nucleocytoplasmic ratio of the cell nucleus, the irregularity of the nuclear membrane contour, the chromatin texture entropy value, as well as the cell-level glandular structure integrity score and the density of mitotic figures per unit area. Finally, all morphological features are generated into a quantitative feature vector to characterize the comprehensive properties of the lesion area. The quantitative feature vector is input into a pre-trained multi-class classification model to establish a non-linear mapping relationship between the quantitative feature space and the tumor type label, outputting the probability value of each preset tumor lesion type, and determining the type with the highest probability value as the classification result of the lesion region.

[0012] As a preferred embodiment of this application, after determining the lesion type, the lesion grade of that category is analyzed, specifically including: To assess the characteristics of adenomatous tumors, the degree of glandular formation was first evaluated. A deep learning segmentation algorithm based on U-Net was used to identify glandular structures, calculate the proportion of glandular area to the total area, and classify them into three levels according to clinical standards: greater than 10%, 10-75%, and less than 75%, and assign corresponding scores. Next, nuclear polymorphism analysis was performed. Individual cell nuclei were separated using a nuclear segmentation algorithm, and nuclear size, shape, and chromatin characteristics were measured. The nuclear size variation coefficient and shape irregularity index were calculated, and the degree of variation was divided into three levels: mild, moderate, and severe, and assigned corresponding scores. Finally, the number of mitotic figures was counted. The YOLO object detection algorithm was used to identify and calculate the number of mitotic figures, and the results were divided into three levels: 1-3. A grading model was established, which integrates three indicators: glandular formation degree, nuclear polymorphism, and nuclear mitotic count. Each indicator was assigned a score of 1-3, and the total score of 3-9 points corresponds to grades I-III.

[0013] Secondly, this application also provides an artificial intelligence-based tumor pathology image classification system, the system comprising: Feature acquisition module: Acquires tumor pathology image data, performs staining and normalization processing on the tumor pathology image data, uses a multi-branch parallel feature extraction network to extract multi-scale features from the processed image data, and performs feature fusion to obtain multi-scale feature maps. Identification and localization module: Identifies and localizes the tumor lesion area based on multi-scale feature maps, and segments the lesion area from the image; Classification and grading module: Classifies the segmented lesion areas according to lesion type, and analyzes the lesion grade of the category after classification; Report output module: Generates and outputs a diagnostic report by taking the lesion type and lesion grade output from the classification and grading module.

[0014] The beneficial effects of this invention are as follows: By performing staining normalization processing on tumor pathological image data, and then using a multi-branch parallel feature extraction network to extract features at multiple scales, the problem of poor robustness due to staining differences and artifacts during lesion area segmentation is solved. At the same time, the classification accuracy is improved. Furthermore, a grading model is established based on the classification to perform quantitative grading assessment. By integrating three indicators—glandular formation degree, nuclear pleomorphism, and mitotic count—more detailed diagnostic information is provided, achieving more refined grading. This further provides doctors with more accurate diagnostic basis and improves the accuracy and reliability of pathological diagnosis. Attached Figure Description

[0015] Figure 1 This is a flowchart illustrating the steps of an artificial intelligence-based tumor pathology image classification method. Figure 2 This is a schematic diagram of the structure of an artificial intelligence-based tumor pathology image classification system. Detailed Implementation

[0016] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to limit the invention. In the following description, the expression "some embodiments" refers to a subset of all possible embodiments; however, it should be understood that "some embodiments" can be the same subset or different subsets of all possible embodiments and can be combined with each other without conflict.

[0017] In the following description, numerous specific details are set forth in order to provide a more thorough understanding of the invention. However, it will be apparent to those skilled in the art that the invention can be practiced without one or more of these details. In other instances, certain technical features well-known in the art have not been described in order to avoid obscuring the invention.

[0018] It should be understood that the present invention can be embodied in various forms and should not be construed as being limited to the embodiments set forth herein. Rather, providing these embodiments will make the disclosure thorough and complete, and will fully convey the scope of the invention to those skilled in the art. Furthermore, the terminology used herein is intended only to describe particular embodiments and is not intended to limit the invention. When used herein, the singular forms “a,” “an,” and “the” are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the terms “compose” and / or “comprising,” when used in this specification, identify the presence of the stated features, integers, steps, operations, elements, and / or components, but do not exclude the presence or addition of one or more other features, integers, steps, operations, elements, components, and / or groups. When used herein, the term “and / or” includes any and all combinations of the associated listed items.

[0019] It should also be noted that when an element is referred to as being "fixed to" another element, it can be directly attached to the other element or there may be an intervening element. When an element is referred to as being "connected to" another element, it can be directly connected to the other element or there may be an intervening element. The terms "vertical," "horizontal," "inner," "outer," "left," "right," and similar expressions used herein are for illustrative purposes only and do not represent the only possible implementation.

[0020] To fully understand this invention, a detailed structure will be presented in the following description to illustrate the technical solution proposed by this invention. Optional embodiments of the invention are described in detail below; however, in addition to these detailed descriptions, the invention may have other embodiments.

[0021] Please refer to the attached document. Figure 1 In a first aspect, this application provides an artificial intelligence-based tumor pathology image classification method, the method comprising the following steps: Step S1: Obtain tumor pathological image data, perform staining and normalization processing on the tumor pathological image data, extract multi-scale features from the processed image data using a multi-branch parallel feature extraction network, and perform feature fusion to obtain multi-scale feature maps. In some embodiments, staining normalization processing of tumor pathology image data specifically includes: K-means clustering analysis was performed on tumor pathology image data in color space to obtain several staining clusters; based on the color consistency index within the clusters and the diagnostic readability score given by the pathologist, the optimal cluster image was selected as the standard staining candidate. Specifically, by calculating the color feature values ​​of all pixels in the image, it is divided into K staining clusters, usually K=3, corresponding to the main staining components such as the cell nucleus, cytoplasm, and background. Each staining cluster represents a specific distribution pattern of staining components.

[0022] Optionally, a dual standard is used to select the optimal staining cluster: firstly, the intra-cluster color consistency index of each cluster is calculated, and the uniformity of color distribution is assessed by calculating the variance of pixel colors within the cluster; secondly, a senior pathologist scores the diagnostic readability of representative images of each cluster to assess whether it meets clinical diagnostic requirements. The cluster with the highest overall score is ultimately selected as the standard staining candidate.

[0023] Color statistics are performed on the optimal cluster image to obtain the mean vector and covariance matrix of the reference color distribution, and a standard coloring space is constructed. Color statistical analysis is performed on the selected optimal cluster image to calculate the mean vector μ and covariance matrix Σ of all pixels in the cluster in the RGB color space. A standard color space is constructed based on these statistics, its mathematical representation being:

[0024] in, This represents a normal distribution, which defines the range of color variation for standard staining.

[0025] Pattern transfer is used to map tumor pathology image data from source staining to standard staining space; and a pattern separation loss function is used to ensure that diagnostic biological tissue information is preserved while normalizing.

[0026] Optionally, the style transfer can employ a CycleGAN-based style transfer network to map the source image from the original chromatic space to a standard chromatic space. During network training, the generator G learns to convert the source image into the target style, while the discriminator D is responsible for distinguishing the generated image from the real standard chromatic image. Adversarial training enables the generator to produce images that conform to the standard chromatic distribution.

[0027] In this embodiment, the processed image data is subjected to multi-scale feature extraction using a multi-branch parallel feature extraction network, and feature fusion is performed to obtain multi-scale image data. Specifically, this includes: The feature extraction network includes four parallel branches, which process image regions at different resolutions respectively; Specifically, Branch 1: Process the original resolution image (1× magnification) to capture cellular-level microscopic features; Branch 2: Process the 2x downsampled image to extract cellular community-level mesoscopic features; Branch 3: Process the 4x downsampled image to obtain glandular structure-level features; Branch 4: Process the 8x downsampled image to capture tissue architecture-level macroscopic features.

[0028] Each branch uses an improved ResNet-50 architecture as its backbone network and embeds a compressed excitation attention module in each residual block, enabling the network to adaptively recalibrate the channel feature response. The branch also employs an expanded convolution strategy to gradually expand the receptive field, ensuring that a wider range of contextual information is captured while maintaining resolution. Specifically, the improved ResNet-50 architecture includes: A compressed excitation (SE) attention module is embedded in each residual block, which works in the following way: First, global average pooling is performed on the input features to obtain channel-level statistics. Then, the dependencies between channels are learned through two fully connected layers. Finally, the sigmoid activation function is used to generate channel weights and recalibrate the feature response. A dilated convolution (dilated convolution) strategy is adopted to gradually expand the receptive field: an increasing dilation rate (2, 4, 6, 8) is used in the deeper layers of the network to expand the receptive field while maintaining the feature map resolution, ensuring that a wider range of contextual information can be captured.

[0029] Then, through the cross-scale feature interaction module, a bidirectional feature pyramid network is used to fuse features of different scales to obtain a multi-scale feature map.

[0030] Specifically, the bidirectional feature pyramid network includes a bottom-up path: passing low-level features upward to enrich the detailed information of high-level features; a top-down path: propagating high-level semantic features downward to enhance the semantic representation of low-level features; then using 1×1 convolutional layers to achieve information exchange of features at the same scale; and finally assigning learnable weights to features at different scales to achieve adaptive feature fusion.

[0031] Step S2: Identify and locate the tumor lesion area based on the multi-scale feature map, and segment the lesion area from the image; In this embodiment, identifying tumor lesion regions based on multi-scale feature maps specifically includes: The visual index of each image region in the multi-scale feature map is calculated to quantify the degree of deviation between the target region and the surrounding normal tissue in the feature space. Then, by setting an adaptive threshold, candidate regions with significant abnormal features are initially screened out. Specifically, based on multi-scale feature maps, an anomaly score is calculated for each image region, and the Mahalanobis distance is used to measure the deviation between the target region features and the distribution of normal tissue features. The specific calculation formula is as follows: ; Where f(x) is the region feature vector, and μ and Σ are the mean and covariance matrices of normal tissue features, respectively.

[0032] An adaptive thresholding algorithm is used to dynamically determine the threshold for judging abnormal regions based on the overall feature distribution of the image.

[0033] Then, based on the probabilistic graphical model of pixel-level and region-level features, the candidate regions are identified as lesion regions. Specifically, adjacent pixels with similar features in the space are assigned the same label. For each candidate region, a series of quantitative attributes are calculated to verify and filter out false positives.

[0034] Specifically, the verification and filtering of false positives includes establishing a tumor lesion region knowledge base containing standard parameters, including integrating WHO classification standards, clinical practice guidelines, and expert consensus, and containing typical parameter ranges and multi-parameter combination patterns for various malignant tumors; and using a rule-based reasoning engine for attribute comparison. Rule 1: Geometric attribute anomaly score > threshold T1; Rule 2: Statistical attribute deviation > threshold T2; Rule 3: Morphological attribute malignancy index > threshold T3; Through a multi-level filtering mechanism, candidate regions that do not conform to the typical manifestations of malignant tumors are gradually eliminated.

[0035] In this embodiment, the quantitative attributes include geometric attributes: region area, perimeter, compactness, and aspect ratio; statistical attributes: average size, distribution density, and texture uniformity of cell nuclei within the region; and morphological attributes: irregularity of region boundaries. Specifically, for the quantification of geometric attributes, the area can be calculated by counting pixels to determine the absolute size of the lesion region, and then converted to its actual physical size by combining this with the image resolution; the perimeter can be precisely calculated using a boundary tracing algorithm to determine the region's contour length; and the compactness can be determined using a formula. Calculations are used to assess the compactness of the region, where... Indicates area, The perimeter is represented by the aspect ratio; the aspect ratio is calculated by the ratio of the length and width of the smallest bounding rectangle, reflecting the morphological characteristics of the region.

[0036] For the analysis of statistical attributes, the average size of the cell nucleus can be identified by the nuclear segmentation algorithm to identify individual cell nuclei and calculate the average diameter and area; the distribution density reflects the density of cell arrangement by calculating the number of cell nuclei per unit area; and the texture uniformity can be evaluated by using local binary mode variance to assess the consistency of the texture. For morphological attribute assessment, boundary irregularity can be calculated using fractal dimension to determine the boundary complexity; boundary curvature analysis assesses the degree of boundary infiltration into surrounding tissues by calculating the average curvature and rate of change of curvature, and then calculating the infiltration index.

[0037] The quantitative attributes are compared with the standard parameter range in the predefined tumor lesion region knowledge base to eliminate candidate regions whose qualitative attribute features do not clearly conform to the typical manifestations of malignant tumors. Finally, through iterative reasoning, an optimal binary segmentation mask is generated, in which high-probability regions are identified as tumor lesion regions.

[0038] In this embodiment, locating the tumor lesion area includes: A subpixel-level edge detection algorithm is used to extract the boundary contour of the tumor lesion area, and the minimum bounding rectangle or fitted ellipse of the contour is calculated. The lesion area is then spatially located in the coordinate system of the original pathological image in the form of bounding box coordinates or contour point set coordinate sequence.

[0039] Optionally, a sub-pixel edge detection algorithm based on Zernike moments is specifically adopted, including: The images of the lesion area are preprocessed, including Gaussian filtering for noise reduction and anisotropic diffusion filtering for enhancement. The Zernike moment for each pixel is calculated, and the probability of edge existence is determined by the phase information of the moment. The precise sub-pixel position of the edge is calculated using the amplitude information of the moment. Non-maximum suppression and double threshold hysteresis processing are used to optimize the edge detection results.

[0040] Specifically, the minimum bounding rectangle of the contour is calculated using a rotating caliper algorithm. First, the convex hull of the contour is calculated, and then the minimum area rectangle is found using the rotating caliper method. The output rectangle parameters include: center coordinates (x, y), width w, height h, and rotation angle θ.

[0041] Specifically, the calculation of the fitted ellipse for the contour is based on fitting the ellipse equation using the least squares method: The geometric parameters of the ellipse are calculated through eigenvalue decomposition, including the center coordinates. , major semi-axis a and minor semi-axis b, rotation angle θ.

[0042] Finally, the localization results are mapped to the coordinate system of the original pathological image. To maintain coordinate consistency, affine transformation is used to process the coordinate transformation, and the transformation parameters are recorded to output localization data in a standardized format.

[0043] In this embodiment, the segmented lesion areas are classified according to the lesion type, specifically including: Within the located lesion area, firstly, the global morphological features of the area are calculated, including the area, perimeter, compactness, and invasiveness index at the junction with the surrounding normal tissue. Secondly, cellular-scale microscopic features are extracted, including the nucleocytoplasmic ratio of the cell nucleus, the irregularity of the nuclear membrane contour, the chromatin texture entropy value, as well as the cell-level glandular structure integrity score and the density of mitotic figures per unit area. Finally, all morphological features are generated into a quantitative feature vector to characterize the comprehensive properties of the lesion area. The quantitative feature vector is input into a pre-trained multi-class classification model to establish a non-linear mapping relationship between the quantitative feature space and the tumor type label, outputting the probability value of each preset tumor lesion type, and determining the type with the highest probability value as the classification result of the lesion region.

[0044] Step S3: Classify the segmented lesion areas according to the lesion type, analyze the lesion level of the category after classification, and output a classification level diagnosis report.

[0045] In this embodiment, based on the characteristics of adenomatous tumors, the degree of glandular formation is first assessed. A deep learning segmentation algorithm based on U-Net is used to identify the glandular structure, calculate the proportion of glandular area to the total area, and classify it into three levels according to clinical standards: greater than 10%, 10-75%, and less than 75%, and assign corresponding scores. Next, nuclear polymorphism analysis was performed. Individual cell nuclei were separated using a nuclear segmentation algorithm, and nuclear size, shape, and chromatin characteristics were measured. The nuclear size variation coefficient and shape irregularity index were calculated, and the degree of variation was divided into three levels: mild, moderate, and severe, and assigned corresponding scores. Finally, the number of mitotic figures was counted. The YOLO object detection algorithm was used to identify and calculate the number of mitotic figures, and the results were divided into three levels: 1-3. A grading model was established, which integrates three indicators: glandular formation degree, nuclear polymorphism, and nuclear mitotic count. Each indicator was assigned a score of 1-3, and the total score of 3-9 points corresponds to grades I-III.

[0046] For example, referring to Table 1, this application uses a pathological slide confirmed by biopsy to be prostate adenocarcinoma as an example to illustrate the specific implementation process of the present invention. The slide is digitally scanned to generate a whole-slice image (WSI) with a resolution of 0.25 μm / pixel.

[0047] Step 1: Glandular formation assessment and duct segmentation: A pre-trained U-Net model is used to segment the lesion region into glandular ducts. Model input: 512×512 pixel image patch; Model output: Binary mask of the glandular duct region; Segmentation accuracy: Dice coefficient reaches 0.93; Area ratio calculation: Total lesion area: 15.6 mm²; Total area of ​​glandular structures: 7.8 mm²; Glandular formation rate: 7.8 / 15.6 × 100% = 50%; Rating assignment: According to clinical standards, 50% falls within the 10%-75% range; Glandular formation score: 2 points (moderate differentiation).

[0048] Step 2: Nuclear polymorphism analysis, cell nucleus segmentation: A total of 1,256 cell nuclei were segmented, with a segmentation accuracy of 94.2%; their characteristics were measured: Core area: average 45.6 μm², standard deviation 12.3 μm²; Nuclear size variation coefficient: (12.3 / 45.6)×100% = 27%; Shape irregularity index: average 0.32 (range 0.12-0.68); Rating assignment: A coefficient of variation of 27% falls within the range of 20%-40%. Nuclear polymorphism score: 2 points (moderate polymorphism).

[0049] Step 3: Nucleus mitosis counting, detection and counting: The YOLOv5 model was used to detect nucleus mitosis, with a detection confidence threshold of 0.75. A total of 8 nucleus mitosis were detected in 10 high-power fields (HPF), with a detection accuracy of 92.5%. Rating assignment: Count result: 8 per 10 HPF, which falls within the range of 6-10 per 10 HPF; Nuclear fission pattern score: 2 points.

[0050] Step 4: Hierarchical calculation and result output.

[0051] Overall rating: Glandular formation degree: 2 points Nuclear polymorphism: 2 points Nuclear fission phenomenon: 2 points Its total score is 2 + 2 + 2 = 6 points. Classification: Based on a total score of 6, it corresponds to Grade II (moderate malignancy), with an overall confidence level of 89.3%. Finally, output the diagnostic report: Generate a structured diagnostic report containing the following: Pathological type: Prostate adenocarcinoma; Grading results: Total score is 6 points; The severity level is II (moderate malignancy). Overall confidence level: 89.3%.

[0052] Table 1 Grading Indicators

[0053] This case is compared with the independent diagnoses of three senior pathology experts: Expert A: Gleason score 3+3=6, Level II; Expert B: Gleason score 3+3=6, Level II; Expert C: Gleason score 3+4=7, Level II.

[0054] It is evident that the grading results of this application are consistent with those of most experts.

[0055] Secondly, please refer to the attached document. Figure 2 This application also provides an artificial intelligence-based tumor pathology image classification system, the system comprising: Feature acquisition module: Acquires tumor pathology image data, performs staining and normalization processing on the tumor pathology image data, uses a multi-branch parallel feature extraction network to extract multi-scale features from the processed image data, and performs feature fusion to obtain multi-scale feature maps. Identification and localization module: Identifies and localizes the tumor lesion area based on multi-scale feature maps, and segments the lesion area from the image; Classification and grading module: Classifies the segmented lesion areas according to lesion type, and analyzes the lesion grade of the category after classification; Report output module: Generates and outputs a diagnostic report by taking the lesion type and lesion grade output from the classification and grading module.

[0056] Thirdly, embodiments of this application also provide an electronic device, including a processor and a memory, wherein the memory stores multiple instructions; the processor loads instructions from the memory to execute steps in any of the artificial intelligence-based tumor pathology image classification methods provided in embodiments of this application.

[0057] Fourthly, embodiments of this application also provide a computer-readable storage medium storing a plurality of instructions adapted for loading by a processor to execute steps in any of the artificial intelligence-based tumor pathology image classification methods provided in embodiments of this application.

[0058] Fifthly, embodiments of this application also provide a computer program product, including a computer program / instructions, which, when executed by a processor, implement the steps in any of the artificial intelligence-based tumor pathology image classification methods provided in embodiments of this application.

[0059] In this embodiment, those skilled in the art will understand that all or part of the steps in the various methods of the above embodiments can be accomplished by instructions, or by instructions controlling related hardware. These instructions can be stored in a computer-readable storage medium and loaded and executed by a processor.

[0060] Therefore, embodiments of this application provide a computer-readable storage medium storing a plurality of instructions that can be loaded by a processor to execute steps in any of the artificial intelligence-based tumor pathology image classification methods provided in embodiments of this application.

[0061] The storage medium may include: read-only memory (ROM), random access memory (RAM), disk or optical disk, etc.

[0062] According to one aspect of this application, a computer program product or computer program is provided, comprising a computer program / instructions stored in a computer-readable storage medium. A processor of an electronic device reads the computer program / instructions from the computer-readable storage medium and executes the computer program / instructions, causing the electronic device to perform the method provided in various optional implementations of any of the steps of the artificial intelligence-based tumor pathology image classification method provided in the above embodiments.

[0063] Since the instructions stored in the storage medium can execute the steps in any of the artificial intelligence-based tumor pathology image classification methods provided in the embodiments of this application, the beneficial effects that any of the artificial intelligence-based tumor pathology image classification methods provided in the embodiments of this application can achieve can be realized, as detailed in the preceding embodiments, and will not be repeated here.

[0064] The above are merely specific embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. The scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. An artificial intelligence-based tumor pathological image classification method, characterized in that, The method includes the following steps: Tumor pathological image data is acquired and stained and normalized. The processed image data is then subjected to multi-scale feature extraction using a multi-branch parallel feature extraction network, and feature fusion is performed to obtain multi-scale feature maps. The staining normalization process for tumor pathology image data specifically includes: performing K-means clustering analysis on the tumor pathology image data in color space to obtain several staining clusters; and selecting the optimal cluster image as the standard staining candidate based on the intra-cluster color consistency index and the given diagnostic readability score. Color statistics are performed on the optimal cluster image to obtain the mean vector and covariance matrix of the reference color distribution, and a standard coloring space is constructed. Pattern transfer is used to map tumor pathology image data from source staining to standard staining space; and a pattern separation loss function is used to ensure that diagnostic biological tissue information is preserved while normalizing. The processed image data is then subjected to multi-scale feature extraction using a multi-branch parallel feature extraction network, followed by feature fusion to obtain multi-scale image data. Specifically, this includes: The feature extraction network includes four parallel branches, which process image regions at different resolutions respectively. Each of the parallel branches uses an improved ResNet-50 architecture as the backbone network and embeds a compressed excitation attention module in each residual block, enabling the network to adaptively recalibrate the channel feature response. Furthermore, an expanded convolution strategy is employed within each of the parallel branches to progressively expand the receptive field, ensuring that a wider range of contextual information is captured while maintaining resolution. Then, through the cross-scale feature interaction module, a bidirectional feature pyramid network is used to fuse features of different scales to obtain a multi-scale feature map. The tumor lesion region is identified and located based on the multi-scale feature map, and the lesion region is segmented from the image; Identifying tumor lesion regions based on multi-scale feature maps specifically includes: The visual index of each image region in the multi-scale feature map is calculated to quantify the degree of deviation between the target region and the surrounding normal tissue in the feature space. Then, by setting an adaptive threshold, candidate regions with significant abnormal features are initially screened out. Based on pixel-level and region-level features, a probabilistic graphical model is used to identify lesion regions in candidate regions. Specifically, pixels that are spatially adjacent and have similar features are assigned the same label. For each candidate region, a series of quantitative attributes are calculated to verify and filter out false positives. The series of quantitative attributes are compared with the standard parameter range in the predefined tumor lesion region knowledge base to eliminate candidate regions whose qualitative attribute features do not clearly conform to the typical manifestations of malignant tumors; finally, through iterative reasoning, an optimal binary segmentation mask is generated, in which high-probability regions are identified as tumor lesion regions. The segmented lesion areas are classified according to the lesion type. After classification, the lesion level of the classification category is analyzed, and a classification level diagnosis report is output.

2. The tumor pathological image classification method based on artificial intelligence according to claim 1, characterized in that, Localizing the tumor lesion area includes: A subpixel-level edge detection algorithm is used to extract the boundary contour of the tumor lesion area, and the minimum bounding rectangle or fitted ellipse of the contour is calculated. The lesion area is then spatially located in the coordinate system of the original pathological image in the form of bounding box coordinates or contour point set coordinate sequence. 3.The tumor pathological image classification method based on artificial intelligence according to claim 1, characterized in that, The segmented lesion areas are classified according to lesion type, specifically including: Within the located lesion area, firstly, the global morphological features of the area are calculated, including the area, perimeter, compactness, and invasiveness index at the junction with the surrounding normal tissue. Secondly, cellular-scale microscopic features are extracted, including the nucleocytoplasmic ratio of the cell nucleus, the irregularity of the nuclear membrane contour, the chromatin texture entropy value, as well as the cell-level glandular structure integrity score and the density of mitotic figures per unit area. Finally, all morphological features are generated into a quantitative feature vector to characterize the comprehensive properties of the lesion area. The quantitative feature vector is input into a pre-trained multi-class classification model to establish a non-linear mapping relationship between the quantitative feature space and the tumor type label, outputting the probability value of each preset tumor lesion type, and determining the type with the highest probability value as the classification result of the lesion region. 4.The method of claim 3, wherein, After determining the lesion type, the lesion grade of that classification category is analyzed, specifically including: To assess the characteristics of adenomatous tumors, the degree of glandular formation was first evaluated. A deep learning segmentation algorithm based on U-Net was used to identify glandular structures, calculate the proportion of glandular area to the total area, and classify them into three levels according to clinical standards: greater than 10%, 10-75%, and less than 75%, and assign corresponding scores. Next, nuclear polymorphism analysis was performed. Individual cell nuclei were separated using a nuclear segmentation algorithm, and nuclear size, shape, and chromatin characteristics were measured. The nuclear size variation coefficient and shape irregularity index were calculated, and the degree of variation was divided into three levels: mild, moderate, and severe, and assigned corresponding scores. Finally, the number of mitotic figures was counted. The YOLO object detection algorithm was used to identify and calculate the number of mitotic figures, and the results were divided into three levels: 1-3. A grading model was established, which integrates three indicators: glandular formation degree, nuclear polymorphism, and nuclear mitotic count. Each indicator was assigned a score of 1-3, and the total score of 3-9 points corresponds to grades I-III.

5. An artificial intelligence-based tumor pathology image classification system for performing an artificial intelligence-based tumor pathology image classification method according to claim 1, characterized in that, The system includes: Feature acquisition module: Acquires tumor pathology image data, performs staining and normalization processing on the tumor pathology image data, uses a multi-branch parallel feature extraction network to extract multi-scale features from the processed image data, and performs feature fusion to obtain multi-scale feature maps. Identification and localization module: Identifies and localizes the tumor lesion area based on multi-scale feature maps, and segments the lesion area from the image; Classification and grading module: Classifies the segmented lesion areas according to lesion type, and analyzes the lesion grade of the classification category after classification; Report output module: Generates and outputs a diagnostic report by taking the lesion type and lesion grade output from the classification and grading module.

Images