Automatic tumor region labeling method for whole slide pathological images of colon cancer

Abstract

This invention discloses an automatic tumor region annotation method for whole-section pathological images of colorectal cancer, belonging to the field of pathological image processing technology. It employs an improved Inception-V3 network, embedding a colorectal cancer feature extraction branch to extract cellular, glandular, and tissue-level morphological features associated with colorectal cancer from stained whole-section pathological images. A fused feature map is obtained by weighting the contribution coefficients of colorectal cancer gene morphology. Based on the colorectal cancer morphological fingerprint vector, an improved Otsu algorithm is used to generate an initial segmentation threshold. The optimal segmentation threshold is obtained through iterative optimization using a colorectal cancer gene morphology matching degree formula, thus initially dividing the tumor candidate region. This invention deeply couples the molecular features of core driver mutation genes in colorectal cancer with the morphological features of pathological images, constructing a dedicated morphological fingerprint database for gene-morphology coupling. This ensures that tumor region annotation aligns with the clinical diagnostic criteria combining molecular and histopathological aspects of colorectal cancer, improving the clinical adaptability of the annotation results.

Description

Technical Field

[0001] This invention relates to the field of pathological image processing technology, and in particular to a method for automatic annotation of tumor regions in whole-section pathological images of colon cancer. Background Technology

[0002] Colorectal cancer is one of the most common malignant tumors worldwide. Pathological diagnosis is the core basis for the diagnosis and staging of colorectal cancer, and accurate annotation of tumor regions in whole-section pathological images is a crucial step in pathological diagnosis. Currently, tumor region annotation in whole-section pathological images of colorectal cancer mainly relies on manual work by pathologists. This method is not only time-consuming and labor-intensive, but also greatly affected by the physician's experience and work status, making it difficult to guarantee the consistency and objectivity of the annotation results. Furthermore, it suffers from annotation errors caused by blurred boundaries between poorly differentiated tumors and normal tissue, and difficulty in distinguishing between colonic adenomas and cancerous tissue.

[0003] With the application of artificial intelligence technology in the field of digital pathology, deep learning-based methods for pathological image region segmentation and annotation have emerged. These methods mostly use general image analysis algorithms to extract universal morphological features of images without combining the molecular pathological features and specific morphological characteristics of colorectal cancer. The segmentation threshold and boundary optimization rules are all general settings, which cannot adapt to the specific characteristics of glandular structural heterogeneity, abnormal nucleocytoplasmic ratio, and morphological features associated with mutated genes in colorectal cancer pathological images. This results in a high false positive rate for tumor region annotation and insufficient accuracy in identifying the boundaries of poorly differentiated colorectal cancer and adjacent tissues.

[0004] Meanwhile, existing technologies do not deeply couple the molecular characteristics of somatic mutations in colorectal cancer with the morphological characteristics of pathological images. They neglect the regulatory role of core driver mutation genes on tissue and cell morphology during the development of colorectal cancer, resulting in annotation results that only remain at the level of image morphology. This fails to meet the clinical diagnostic criteria that combine molecular pathology and histopathology of colorectal cancer, and makes it difficult to meet the actual needs of accurate diagnosis and early screening of colorectal cancer. Summary of the Invention

[0005] To address the aforementioned technical problems, this invention provides a method for automatic tumor region annotation in whole-section pathological images of colorectal cancer. The technical solution adopted is as follows:

[0006] An automatic tumor region annotation method for whole-section pathological images of colon cancer includes the following steps:

[0007] Step 1: Obtain the whole-section pathological images of colorectal cancer and the somatic mutation data of the corresponding cases. Perform standardized preprocessing on the whole-section pathological images of colorectal cancer, screen the somatic mutation data to obtain the set of core driver mutation genes of colorectal cancer, and determine the morphological contribution coefficient of each gene.

[0008] Step 2: Based on the functional enrichment results of the core driver mutation genes in colorectal cancer, extract the morphological feature set of colorectal cancer and establish a quantitative model to generate a colon morphological fingerprint vector that integrates gene morphological contribution coefficients and construct a gene-morphological coupled colorectal cancer morphological fingerprint database.

[0009] Step 3: Using the improved Inception-V3 network, embedding a colorectal cancer feature extraction branch, extracting cellular, glandular, and tissue-level colorectal cancer-related morphological features from stained whole-section pathological images, and obtaining a fused feature map by weighting the colorectal cancer gene morphological contribution coefficient;

[0010] Step 4: Based on the colorectal cancer morphological fingerprint vector, the Otsu improved algorithm is used to generate the initial segmentation threshold. The optimal segmentation threshold is obtained by iterative optimization through the colorectal cancer gene morphology matching degree formula, and the tumor candidate region is initially divided.

[0011] Step 5: Optimize the candidate region boundary using edge detection and region growing algorithms, eliminate false positive regions using the colon cancer gene and morphology consistency verification formula, and eliminate isolated noise regions using colon cancer pathology standards to obtain the optimized tumor candidate region.

[0012] Optionally, the standardization preprocessing in step 1 includes the following sub-steps:

[0013] Step 11: Use the OpenSlide tool to divide the whole-section pathological image of colon cancer staining into image blocks according to the set pixels, set the pathological diagnosis magnification, and remove noisy image blocks where the white background ratio exceeds the set first ratio threshold and the effective tissue ratio of colon cancer is lower than the set second ratio threshold.

[0014] Step 12: Normalize the pixel values ​​of the image block using the Z-score normalization formula;

[0015] Step 13: Use the Mutect2 tool to screen for somatic mutations in colorectal cancer and remove synonymous mutations. Use the random forest algorithm combined with the TNM staging criteria for colorectal cancer to screen for the set of core driver mutation genes in colorectal cancer. Calculate the morphological contribution coefficient of each gene using the following formula for colorectal cancer gene morphological contribution coefficient. :

[0016] ;in , where n is the product of the mutation frequency and functional enrichment score of the gene in colorectal cancer, and n is the number of core driver mutation genes.

[0017] Optionally, the Otsu improved algorithm adjusts the pixel weight coefficients for calculating the inter-class variance to 1.2 for tumor tissue and 0.8 for normal tissue to address the difference in pixel proportion between tumor tissue and normal tissue in colorectal cancer pathological images.

[0018] Optionally, step 2, the construction of the colon cancer morphological fingerprint database, includes the following sub-steps:

[0019] Step 21: Use the cluster profile tool to perform GO functional enrichment analysis on the core driver mutation gene set of colorectal cancer, screen biological processes with p<0.05 that are related to the morphology of colorectal cancer tissue or cells, and extract the colorectal cancer morphological feature set.

[0020] Step 22: Perform quantitative pathological modeling of colon cancer based on the specific morphological features of each colon cancer:

[0021] glandular fragmentation , The total area of ​​the broken glands in the image patch. This represents the total area of ​​normal colonic glands of the same area;

[0022] nucleo-cylindrical ratio anomaly , This represents the mean nucleus-to-cytoplasm ratio of cells in the image patch. =0.3 is the critical value for the nucleus-to-cytoplasm ratio of normal colonic epithelial cells;

[0023] Cell arrangement disorder , Correlation coefficient of cell arrangement in image patches The critical value of the correlation coefficient for the normal arrangement of colonic tissue cells;

[0024] Step 23: Generate colonic morphological fingerprint vector m represents the number of morphological features of colon cancer. To determine the contribution coefficient of corresponding gene morphology, a gene-morphology coupled colorectal cancer morphological fingerprint database was constructed.

[0025] Optionally, the method for extracting colon cancer-related morphological features from the stained whole-section pathological images in step 3 is as follows:

[0026] A colorectal cancer feature extraction branch was embedded into the Inception-V3 network, containing three sets of convolutional kernels, corresponding to cellular, glandular, and tissue-level colorectal cancer-specific features, respectively. The fused feature map was obtained using a weighted fusion formula based on the colorectal cancer gene morphology contribution coefficient.

[0027] ,in , For genes The feature contribution weights at the k-th scale This is a feature map of colon cancer at the k-th scale;

[0028] PCA is used to reduce the dimensionality of the fused feature map, retaining principal components with a cumulative contribution rate of ≥90%, and CBAM is used to enhance the feature regions that match the morphological fingerprint of colorectal cancer by embedding a colorectal cancer feature attention module.

[0029] Optionally, step 4, which involves the preliminary delineation of the tumor candidate region, includes the following sub-steps:

[0030] Step 41: Generate the initial threshold using the improved Otsu algorithm. The pixel class variance weights of normal and tumor tissues in colon cancer were adjusted to 1.2:1 to adapt to the pixel distribution characteristics of colon cancer pathological images.

[0031] Step 42: Calculate candidate regions using the colon cancer gene morphology matching formula. :

[0032] ;

[0033] in Let i be the i-th eigenvalue of the morphological fingerprint vector of colon cancer. Let i be the i-th eigenvalue of the candidate region. The coefficient represents the morphological contribution of the corresponding gene, and m represents the number of colorectal cancer-specific morphological features. The closer the value is to 1, the higher the match between the candidate region and the morphological features of colon cancer.

[0034] Step 43, if Then according to Adjust the threshold, where =0.05 is the iteration step size, which decreases by 10% every 5 iterations. The iteration terminates under the condition that... Alternatively, the optimal segmentation threshold can be obtained by iterating 50 times. ;

[0035] Step 44, based on Segmenting the fusion feature map of colon cancer yields preliminary tumor candidate regions. .

[0036] Optionally, in step 5, the improved Canny operator is used to extract the edge contour of the candidate region and calculate the gradient value of the colon cancer edge;

[0037] The formula for the improved Canny operator is: ,in , These are the gradients in the x and y directions, respectively. , The gradient method weights for colorectal cancer case images are used, with edge contours as seed points. A colorectal cancer region growth criterion is applied: the similarity of colorectal cancer features between adjacent pixels is greater than or equal to 0.85, and... Less than or equal to 0.25, the optimized region boundary is obtained ;

[0038] extract The morphological feature vectors are used to calculate the consistency score using a consistency verification formula. ,like The regions identified as false positives were removed; based on the clinical pathological criteria for colon cancer, isolated noise regions with an area less than 800 pixels were removed to obtain the optimized tumor candidate regions.

[0039] Optionally, step 6 is also included, which compares the optimized tumor candidate region with the manually annotated results based on the clinical diagnostic criteria for colorectal cancer, calculates the colorectal cancer validation index, adaptively adjusts the model parameters, and outputs the final annotation results of the colorectal cancer tumor region, including region coordinates, boundary contours, and clinical confidence.

[0040] Optionally, in step 6, based on the TNM staging and pathological diagnostic criteria for colorectal cancer, validation indicators for colorectal cancer are determined, including the crossover ratio of tumor regions. Sensitivity for colon cancer tumor detection Specificity of normal tissue in colon cancer Calculate the clinical confidence score for colon cancer. :

[0041] .

[0042] Optionally, the quantified values ​​of colorectal cancer morphological features are all normalized to the [0,1] interval.

[0043] In summary, the present invention has at least one of the following beneficial technical effects:

[0044] This invention provides an automatic tumor region annotation method for whole-section pathological images of colorectal cancer. It deeply couples the molecular features of the core driver mutation genes of colorectal cancer with the morphological features of the pathological images, and constructs a dedicated morphological fingerprint library of gene-morphology coupling. This makes the tumor region annotation conform to the clinical diagnostic criteria of colorectal cancer that combine molecular pathology and histopathology. It breaks through the limitation of existing technologies that rely solely on general image features and improves the clinical adaptability of the annotation results.

[0045] We designed core algorithms such as the improved Otsu algorithm, gene morphology matching formula, and improved Canny operator specifically for the unique features of colorectal cancer pathological images. These algorithms are adapted to the heterogeneous structure and blurred boundaries of colorectal cancer glands, effectively reducing the identification error of poorly differentiated colorectal cancer and adjacent tissue boundaries, and significantly improving the accuracy of tumor region annotation.

[0046] This technology enables automated annotation of tumor regions in whole-section pathological images of colorectal cancer, replacing traditional manual annotation methods. It significantly improves annotation efficiency, reduces annotation bias caused by human factors, and ensures the consistency and objectivity of annotation results, providing efficient and accurate technical support for the pathological diagnosis and early screening of colorectal cancer. Attached Figure Description

[0047] Figure 1 This is a flowchart illustrating the automatic tumor region annotation method for whole-section pathological images of colon cancer according to the present invention.

[0048] Figure 2 This is a bar chart showing the distribution of morphological contribution coefficients of core driver mutation genes in colorectal cancer according to a specific embodiment of the present invention.

[0049] Figure 3 This is a comparison diagram of tumor candidate region division and boundary optimization in a specific embodiment of the present invention;

[0050] Figure 4 This is a histogram showing the distribution of colorectal cancer validation indicators according to a specific embodiment of the present invention;

[0051] Figure 5 This is a histogram of clinical confidence distribution according to a specific embodiment of the present invention. Detailed Implementation

[0052] The present invention will be further described in detail below with reference to the accompanying drawings.

[0053] This invention discloses a method for automatically annotating tumor regions in whole-section pathological images of colon cancer.

[0054] Reference Figures 1-5 Example 1: An automatic tumor region annotation method for whole-section pathological images of colon cancer, comprising the following steps:

[0055] Step 1: Obtain the whole-section pathological images of colorectal cancer and the somatic mutation data of the corresponding cases. Perform standardized preprocessing on the whole-section pathological images of colorectal cancer, screen the somatic mutation data to obtain the set of core driver mutation genes of colorectal cancer, and determine the morphological contribution coefficient of each gene.

[0056] Step 2: Based on the functional enrichment results of the core driver mutation genes in colorectal cancer, extract the morphological feature set of colorectal cancer and establish a quantitative model to generate a colon morphological fingerprint vector that integrates gene morphological contribution coefficients and construct a gene-morphological coupled colorectal cancer morphological fingerprint database.

[0057] Step 3: Using the improved Inception-V3 network, embedding a colorectal cancer feature extraction branch, extracting cellular, glandular, and tissue-level colorectal cancer-related morphological features from stained whole-section pathological images, and obtaining a fused feature map by weighting the colorectal cancer gene morphological contribution coefficient;

[0058] Step 4: Based on the colorectal cancer morphological fingerprint vector, the Otsu improved algorithm is used to generate the initial segmentation threshold. The optimal segmentation threshold is obtained by iterative optimization through the colorectal cancer gene morphology matching degree formula, and the tumor candidate region is initially divided.

[0059] Step 5: Optimize the candidate region boundary using edge detection and region growing algorithms, eliminate false positive regions using the colon cancer gene and morphology consistency verification formula, and eliminate isolated noise regions using colon cancer pathology standards to obtain the optimized tumor candidate region.

[0060] Example 2, the standardization preprocessing in step 1 includes the following sub-steps:

[0061] Step 11: Use the OpenSlide tool to divide the whole-section pathological image of colon cancer staining into image blocks according to the set pixels, set the pathological diagnosis magnification, and remove noisy image blocks where the white background ratio exceeds the set first ratio threshold and the effective tissue ratio of colon cancer is lower than the set second ratio threshold.

[0062] Step 12: Normalize the pixel values ​​of the image block using the Z-score normalization formula;

[0063] Step 13: Use the Mutect2 tool to screen for somatic mutations in colorectal cancer and remove synonymous mutations. Use the random forest algorithm combined with the TNM staging criteria for colorectal cancer to screen for the set of core driver mutation genes in colorectal cancer. Calculate the morphological contribution coefficient of each gene using the following formula for colorectal cancer gene morphological contribution coefficient. :

[0064] ;in , where n is the product of the mutation frequency and functional enrichment score of the gene in colorectal cancer, and n is the number of core driver mutation genes.

[0065] By employing the above-mentioned technical solutions, the occurrence and development of colorectal cancer are closely related to the abnormal expression of core driver mutant genes, and different mutant genes have specific regulatory effects on the morphology of colon tissue and cells. This step first standardizes the whole-section pathological images, removing invalid noise areas and unifying the image pixel distribution characteristics, laying the foundation for subsequent feature extraction. Simultaneously, it screens the set of core driver mutant genes for colorectal cancer, and by calculating the morphological contribution coefficient of each gene, quantifies the degree of influence of different mutant genes on colon tissue morphology, achieving a preliminary correlation between molecular-level features and morphological-level features, providing a molecular-level weighting basis for the subsequent construction of a morphological fingerprint database.

[0066] Functional enrichment of core driver mutations in colorectal cancer reflects the biological processes of gene regulation, which directly correspond to morphological changes in colon tissue and cells. Step 2, based on the gene functional enrichment analysis results, extracts morphological features related to the occurrence and development of colorectal cancer and establishes a quantitative model. Each morphological feature is then fused with the morphological contribution coefficient of its corresponding gene to generate a targeted morphological fingerprint vector. The core of this process is establishing a quantitative correlation between molecular and morphological features. The constructed morphological fingerprint database accurately reflects the unique morphological characteristics of colorectal cancer, distinguishing it from the morphological characteristics of normal colon tissue and other tumors, providing a specific feature reference standard for subsequent tumor region identification.

[0067] The pathological features of colorectal cancer are manifested in multiple dimensions, including cells, glands, and tissues, with different dimensions contributing differently to tumor region identification. Step 3 embeds a colorectal cancer feature extraction branch into the Inception-V3 network to selectively extract colorectal cancer-related morphological features at different dimensions, avoiding the omission of key features caused by general feature extraction. Simultaneously, the gene morphology contribution coefficient is used to weight and fuse features at different scales, strengthening morphological features highly correlated with colorectal cancer molecular features and weakening irrelevant features. This allows the fused feature map to accurately represent the multi-dimensional morphological characteristics of colorectal cancer tumor regions, improving the specificity of feature representation.

[0068] General threshold segmentation algorithms do not consider the pixel distribution characteristics of colorectal cancer pathological images, making it difficult to accurately delineate the boundary between tumors and normal tissues. Step 4 uses an improved Otsu algorithm to generate an initial threshold, adapting to the pixel ratio difference between colorectal cancer tumors and normal tissues. The matching degree between candidate regions and the morphological fingerprint database is calculated using the colorectal cancer gene-morphology matching formula, and the segmentation threshold is iteratively optimized. This ensures that the threshold adjustment process is always based on the gene-morphological features specific to colorectal cancer, ensuring that the segmented candidate regions can accurately match the molecular and morphological features of colorectal cancer tumors and reducing missegmentation of non-tumor regions.

[0069] Colorectal cancer tumor regions often exhibit blurred and irregular boundaries, making it difficult for general edge detection and region growing algorithms to accurately characterize these boundaries. Step 5 employs an improved Canny operator to extract edge contours, adjusting gradient direction weights to match the edge features of the colorectal cancer pathological image. The boundary is optimized using colorectal cancer feature similarity and edge gradient values ​​as growth criteria, ensuring the boundary closely matches the pathological morphology of the colorectal cancer tumor. False positive regions that do not match the colorectal cancer morphological fingerprint are eliminated using a gene-morphology consistency verification formula, and isolated noise regions are removed using colorectal cancer clinical pathological standards, further improving the accuracy of tumor region annotation and reducing the false positive rate.

[0070] Example 3: The Otsu improved algorithm addresses the difference in pixel proportion between tumor tissue and normal tissue in colorectal cancer pathological images by adjusting the pixel weight coefficients calculated from the inter-class variance to 1.2 for tumor tissue and 0.8 for normal tissue.

[0071] By employing the aforementioned technical solutions, the core logic of the Otsu algorithm is to determine the optimal segmentation threshold by calculating the inter-class variance of the image's gray levels, thereby maximizing the inter-class variance between the segmented foreground and background and effectively distinguishing different regions. However, the general Otsu algorithm assumes that pixel weights are equal in different regions of the image, failing to consider the pixel distribution characteristics of specific pathological images and thus unable to adapt to the actual characteristics of colorectal cancer pathological images.

[0072] In colorectal cancer pathological images, there is a significant difference in the pixel proportion between tumor tissue and normal tissue, and tumor regions often exhibit focal distribution characteristics. If equal weights are used to calculate the inter-class variance, the higher pixel proportion of normal tissue can easily lead to a segmentation threshold biased towards normal tissue, resulting in incomplete segmentation or boundary offset of the tumor region. To address this issue, this improved algorithm adjusts the pixel weight coefficients in the inter-class variance calculation based on the pixel distribution patterns of colorectal cancer pathological images, setting the weight coefficient for tumor tissue to 1.2 and for normal tissue to 0.8. This strengthens the contribution of tumor tissue pixels to the inter-class variance calculation while weakening the influence of normal tissue pixels, making the calculated inter-class variance results more closely reflect the pixel distribution characteristics of the colorectal cancer tumor region. The adjusted algorithm can accurately capture the pixel grayscale features of tumor tissue, and the generated initial segmentation threshold can effectively distinguish between colorectal cancer tumor tissue and normal tissue, avoiding threshold deviation caused by differences in pixel proportion and improving the accuracy of subsequent tumor candidate region segmentation.

[0073] Example 4, the construction of the colon cancer morphological fingerprint database in step 2 includes the following sub-steps:

[0074] Step 21: Use the cluster profile tool to perform GO functional enrichment analysis on the core driver mutation gene set of colorectal cancer, screen biological processes with p<0.05 that are related to the morphology of colorectal cancer tissue or cells, and extract the colorectal cancer morphological feature set.

[0075] Step 22: Perform quantitative pathological modeling of colon cancer based on the specific morphological features of each colon cancer:

[0076] glandular fragmentation , The total area of ​​the broken glands in the image patch. This represents the total area of ​​normal colonic glands of the same area;

[0077] nucleo-cylindrical ratio anomaly , This represents the mean nucleus-to-cytoplasm ratio of cells in the image patch. =0.3 is the critical value for the nucleus-to-cytoplasm ratio of normal colonic epithelial cells;

[0078] Cell arrangement disorder , Correlation coefficient of cell arrangement in image patches The critical value of the correlation coefficient for the normal arrangement of colonic tissue cells;

[0079] Step 23: Generate colonic morphological fingerprint vector m represents the number of morphological features of colon cancer. To determine the contribution coefficient of corresponding gene morphology, a gene-morphology coupled colorectal cancer morphological fingerprint database was constructed.

[0080] By employing the above technical solution and utilizing the Cluster Profile tool to conduct GO functional enrichment analysis, based on the functional expression patterns of core driver mutation genes in colorectal cancer, biological processes directly related to morphological changes in colorectal cancer tissue cells are screened out. Based on this, a set of colorectal cancer morphological features is extracted, ensuring that the extracted morphological features are not general image morphological features, but rather specific morphological features regulated by colorectal cancer mutation genes and directly related to the occurrence and development of colorectal cancer. This fundamentally guarantees the tumor specificity and clinical relevance of the morphological features, excluding interference from morphological features unrelated to colorectal cancer.

[0081] Pathological quantitative modeling of colorectal cancer-specific morphological features transforms the unique histiocellular morphological changes of colorectal cancer into calculable quantitative indicators. Using the morphological characteristics of normal colon tissue as a reference, the degree of morphological abnormality of tumor tissue relative to normal tissue is quantified through ratio calculation. Glandular fragmentation, abnormal nuclear-cytoplasmic ratio, and disordered cell arrangement are all typical morphological features that distinguish colorectal cancer tumor tissue from normal tissue. Through corresponding quantitative formulas, intuitive morphological differences are transformed into standardized values, achieving quantitative characterization of morphological features and providing a calculable numerical basis for subsequent feature matching and threshold segmentation.

[0082] A colonic morphological fingerprint vector was generated and a morphological fingerprint database was constructed. The quantified morphological features were then fused with the morphological contribution coefficients of corresponding regulatory genes, ensuring that the feature values ​​of each dimension of the morphological fingerprint vector reflected both the degree of morphological abnormality in colon cancer tissue cells and the weight of the molecular regulatory mechanisms behind those morphological features. Different core driver mutations contribute differently to morphological changes in colon cancer. By weighting the gene morphological contribution coefficients, gene-related morphological features that significantly influence colon cancer morphological changes were strengthened, while features with less influence were weakened. This ensured that the final morphological fingerprint database accurately and comprehensively reflected the overall morphological features under gene regulation in colon cancer, allowing subsequent tumor region identification to be both consistent with morphological features and molecular pathological mechanisms.

[0083] Example 5, the method for extracting colon cancer-related morphological features from the whole-section pathological images of stained colon cancer in step 3 is as follows:

[0084] A colorectal cancer feature extraction branch was embedded into the Inception-V3 network, containing three sets of convolutional kernels, corresponding to cellular, glandular, and tissue-level colorectal cancer-specific features, respectively. The fused feature map was obtained using a weighted fusion formula based on the colorectal cancer gene morphology contribution coefficient.

[0085] ,in , For genes The feature contribution weights at the k-th scale This is a feature map of colon cancer at the k-th scale;

[0086] PCA is used to reduce the dimensionality of the fused feature map, retaining principal components with a cumulative contribution rate of ≥90%, and CBAM is used to enhance the feature regions that match the morphological fingerprint of colorectal cancer by embedding a colorectal cancer feature attention module.

[0087] By adopting the above technical solution, a colorectal cancer feature extraction branch is embedded in the Inception-V3 network and three convolutional kernel groups are set up. The core basis is that the pathological morphological features of colorectal cancer are reflected in different scales of cells and glandular tissues, and the features at different scales have different representational dimensions. The dedicated convolutional kernel groups can specifically capture the unique morphological features of colorectal cancer at each scale, avoid the indiscriminate extraction of features by general networks, effectively avoid the interference of non-tumor-related features, and prevent the omission of key morphological features of colorectal cancer, so as to achieve the targeted and accurate extraction of colorectal cancer-related morphological features at each scale.

[0088] The fusion feature map obtained by weighting the contribution coefficients of colorectal cancer gene morphology is based on the molecular pathological mechanism of colorectal cancer morphological features being regulated by core driver mutation genes. It combines the gene regulation weights at the molecular level with multi-scale morphological features at the image level. By calculating the fusion weight coefficients at each scale, the fusion weights of feature maps at different scales are matched with the contribution of gene regulation. This strengthens morphological features regulated by core driver mutation genes and highly correlated with colorectal cancer tumors, while weakening the influence of irrelevant features. This allows the fusion feature map to comprehensively reflect the multi-scale morphological features of colorectal cancer while conforming to its molecular pathological regulatory patterns, thus improving the specificity and targeting of feature representation.

[0089] The core idea of ​​performing PCA dimensionality reduction on the fused feature map and retaining principal components with a cumulative contribution rate of ≥90% is to remove redundant information and noise in the fused features. While preserving the key morphological features of colorectal cancer to the greatest extent, the feature dimension is reduced, the complexity of subsequent calculations is reduced, the efficiency of feature processing is improved, and the overfitting problem caused by high-dimensional features is avoided, thus ensuring the effectiveness and robustness of the features.

[0090] Embedding a colorectal cancer feature attention module in CBAM and strengthening the feature regions that match the colorectal cancer morphological fingerprint utilizes the feature focusing characteristics of the attention mechanism. Using the colorectal cancer morphological fingerprint database as a reference standard, the network automatically identifies and focuses on regions in the fused feature map that match the morphological features of colorectal cancer tumors. This further enhances the representation intensity of these key region features, suppresses feature signals from background and non-tumor regions, and enables the extracted features to more accurately point to the colorectal cancer tumor region, laying a solid feature foundation for the accurate segmentation of subsequent tumor candidate regions.

[0091] Example 6, step 4, the preliminary segmentation of the tumor candidate region includes the following sub-steps:

[0092] Step 41: Generate the initial threshold using the improved Otsu algorithm. The pixel class variance weights of normal and tumor tissues in colon cancer were adjusted to 1.2:1 to adapt to the pixel distribution characteristics of colon cancer pathological images.

[0093] Step 42: Calculate candidate regions using the colon cancer gene morphology matching formula. :

[0094] ;

[0095] in Let i be the i-th eigenvalue of the morphological fingerprint vector of colon cancer. Let i be the i-th eigenvalue of the candidate region. The coefficient represents the morphological contribution of the corresponding gene, and m represents the number of colorectal cancer-specific morphological features. The closer the value is to 1, the higher the match between the candidate region and the morphological features of colon cancer.

[0096] Step 43, if Then according to Adjust the threshold, where =0.05 is the iteration step size, which decreases by 10% every 5 iterations. The iteration terminates under the condition that... Alternatively, the optimal segmentation threshold can be obtained by iterating 50 times. ;

[0097] Step 44, based on Segmenting the fusion feature map of colon cancer yields preliminary tumor candidate regions. .

[0098] By adopting the above technical solution and using the improved Otsu algorithm to generate the initial threshold, the core is to address the inherent characteristic of uneven pixel proportions between tumor and normal tissue in colorectal cancer pathological images. By adjusting the pixel weights of the inter-class variance, the contribution of tumor tissue pixels in the threshold calculation is enhanced. This allows the generation of the initial threshold to no longer follow the general rule of equal weights, but rather adapt to the pixel distribution pattern of colorectal cancer pathological images. This reduces the threshold offset caused by the excessive proportion of normal tissue pixels from the source, enabling the initial threshold to initially distinguish between colorectal cancer tumor and normal tissue regions, and providing a reasonable basic threshold for subsequent iterative optimization.

[0099] The formula for calculating the matching degree of candidate regions using the gene-morphology matching degree of colorectal cancer is the core of using the gene-morphology coupling features specific to colorectal cancer as the core basis for candidate region determination. The gene-morphology contribution coefficient, the feature value of the colorectal cancer morphological fingerprint vector, and the actual morphological feature value of the candidate region are deeply integrated into the matching degree calculation. This allows the matching degree result to not only reflect the similarity between the candidate region and the morphological features of colorectal cancer, but also to reflect the underlying molecular gene regulatory weights. This enables the determination of whether the candidate region conforms to the tumor characteristics of colorectal cancer from both morphological and molecular dimensions, avoiding misjudgments caused by relying solely on image grayscale features, and improving the specificity and accuracy of candidate region determination.

[0100] The core of dynamic iterative optimization of the threshold based on matching results is to establish a threshold adjustment mechanism guided by the gene-morphology matching degree of colorectal cancer. Using a preset matching degree standard as the target, the threshold is gradually adjusted through a combination of fixed step size and step size decay. This ensures the stability of the threshold adjustment while allowing for fine-tuning in the later stages of iteration, continuously bringing the threshold closer to the characteristics of colorectal cancer tumors. Simultaneously, a clear iteration termination condition is set to ensure that the candidate region corresponding to the final threshold meets the preset colorectal cancer morphology matching standard, while avoiding excessive computational resource consumption due to over-iteration, thus achieving a balance between efficiency and accuracy.

[0101] In Example 7, step 5, an improved Canny operator is used to extract the edge contours of the candidate region and calculate the gradient value of the colon cancer edge.

[0102] The formula for the improved Canny operator is: ,in , These are the gradients in the x and y directions, respectively. , The gradient method weights for colorectal cancer case images are used, with edge contours as seed points. A colorectal cancer region growth criterion is applied: the similarity of colorectal cancer features between adjacent pixels is greater than or equal to 0.85, and... Less than or equal to 0.25, the optimized region boundary is obtained ;

[0103] extract The morphological feature vectors are used to calculate the consistency score using a consistency verification formula. ,like The regions identified as false positives were removed; based on the clinical pathological criteria for colon cancer, isolated noise regions with an area less than 800 pixels were removed to obtain the optimized tumor candidate regions.

[0104] By adopting the above technical solution, the improved Canny operator is used to calculate the gradient value of the colorectal cancer edge and extract the edge contour. The core is to assign differentiated weights to the gradient direction based on the edge distribution characteristics of the colorectal cancer pathological image, adapt to the characteristic performance of the edge of the colorectal cancer tumor region in different directions, and make the gradient value calculation more in line with the actual gray-scale change law of the colorectal cancer edge. This avoids the problems of inaccurate edge extraction and blurred edge omission caused by the equal weight calculation of the general operator, thereby accurately capturing the true edge contour of the colorectal cancer tumor region and providing an accurate seed point basis for subsequent boundary optimization.

[0105] The core of optimizing the boundary by using the edge contour as the seed point and combining it with the growth criteria of the colorectal cancer region is to establish a region growth rule that fits the pathological features of colorectal cancer. The similarity of colorectal cancer features and the edge gradient value are used as dual judgment criteria. This ensures that the morphological features of the grown region are highly matched with the features of the colorectal cancer tumor, while limiting the growth range to within the actual tumor edge. This prevents the boundary from growing too much or too little, and allows the optimized boundary to accurately fit the irregular shape of the colorectal cancer tumor region, solving the boundary offset and contour distortion problems that are easily caused by general growth algorithms.

[0106] The core of calculating scores and eliminating false positive regions through the consistency verification formula is to perform secondary matching and verification between the actual morphological features of the candidate region and the morphological fingerprint of colorectal cancer gene-morphology coupling. This determines whether the region is a real colorectal cancer tumor region from both molecular and morphological dimensions, effectively eliminating false positive regions that are only similar in edge features but do not match the core morphological features of colorectal cancer fingerprint. This ensures the tumor specificity of the candidate region at the feature level and reduces the probability of misjudgment.

[0107] The core of eliminating isolated noisy regions by combining image analysis results with the actual standards for clinical diagnosis of colorectal cancer is to integrate the results with the actual standards for clinical diagnosis of colorectal cancer. Using the area of ​​the smallest tumor lesion in clinical colorectal cancer as the criterion, isolated small regions caused by image noise and tissue debris are eliminated. These regions have no clinical or pathological significance and are not real tumor lesions. This operation further purifies the candidate regions, ensuring that the optimized tumor regions fully conform to the clinical and pathological criteria for colorectal cancer, thus guaranteeing the clinical applicability of the annotation results.

[0108] Example 8 also includes step 6, which compares the optimized tumor candidate region with the manually annotated results based on the clinical diagnostic criteria for colorectal cancer, calculates the colorectal cancer validation index, adaptively adjusts the model parameters, and outputs the final annotation results of the colorectal cancer tumor region, including region coordinates, boundary contours, and clinical confidence.

[0109] In Example 9, step 6, based on the TNM staging and pathological diagnostic criteria for colorectal cancer, validation indicators for colorectal cancer are determined, including the crossover ratio of tumor regions. Sensitivity for colon cancer tumor detection Specificity of normal tissue in colon cancer Calculate the clinical confidence score for colon cancer. :

[0110] .

[0111] By employing the aforementioned technical solution, the optimized tumor candidate regions are compared with manually annotated results based on clinical diagnostic standards for colorectal cancer. The manually annotated regions completed by pathologists according to clinical diagnostic guidelines for colorectal cancer are used as the gold standard. This gold standard integrates clinical diagnostic requirements such as TNM staging and histopathological feature determination for colorectal cancer, serving as an authoritative reference for colorectal cancer tumor region annotation. By comparing the algorithm-annotated regions with this gold standard, the degree of conformity between the automated annotation results and clinical diagnostic requirements in tumor region localization and extent division can be intuitively and accurately measured, providing a real and effective reference for subsequent quantitative evaluation and parameter adjustment.

[0112] Based on the TNM staging and pathological diagnostic criteria for colorectal cancer, three validation indicators were determined: tumor region crossover ratio (CGL), tumor detection sensitivity, and normal tissue specificity. Quantitative evaluation dimensions highly aligned with the actual needs of colorectal cancer pathological diagnosis were selected to comprehensively assess the annotation effectiveness. The CGL reflects the degree of overlap between the algorithm-annotated region and the clinical gold standard region, directly reflecting the accuracy of tumor region localization. Tumor detection sensitivity measures the algorithm's ability to identify and cover colorectal cancer tumor regions, ensuring that real tumor regions are not missed in clinical practice. Normal tissue specificity assesses the algorithm's ability to distinguish between tumor tissue and normal colon tissue, preventing normal tissue from being misclassified as tumor regions. These three indicators constitute a complete evaluation system for colorectal cancer tumor annotation effectiveness from different dimensions, comprehensively reflecting the algorithm's performance in accuracy, detection rate, and specificity, aligning with the clinical evaluation logic of colorectal cancer pathological diagnosis.

[0113] The clinical confidence score for colorectal cancer integrates and quantifies three independent validation indicators, transforming them into an intuitive percentage score that comprehensively reflects the overall clinical compliance of the automated annotation results. By averaging the three indicators and converting them to a percentage, the confidence score takes into account the accuracy of tumor localization, the completeness of tumor detection, and the accuracy of distinguishing normal tissues. This achieves both a quantitative representation of the annotation effect and makes the results easier for pathologists to understand and refer to, providing a clear quantitative basis for the acceptance of automated annotation results in clinical diagnosis.

[0114] Example 10: The quantified values ​​of the morphological characteristics of colon cancer were all normalized to the [0,1] interval.

[0115] The following specific embodiments illustrate the implementation principle of the present invention:

[0116] We selected 100 H&E-stained whole-section pathological images of colorectal cancer from the TCGA database and the corresponding somatic mutation data. All images were standard sections fixed in formalin and embedded in paraffin. The pathological diagnosis magnification was uniformly set to 20x. Three physicians with more than 10 years of experience in colorectal cancer pathological diagnosis were selected to manually annotate the images according to the clinical diagnostic criteria for colorectal cancer, which served as the gold standard for verification.

[0117] Step 1: Multimodal data standardization preprocessing:

[0118] Using the OpenSlide tool, 100 whole-section pathological images of colorectal cancer were divided into image blocks of 512×512 pixels. The first proportion threshold was set to 30% and the second proportion threshold was set to 10%. Noisy image blocks with a white background proportion of more than 30% and an effective tissue proportion of less than 10% were removed, resulting in a total of 86,420 effective image blocks.

[0119] The Z-score normalization formula is used to normalize the pixel values ​​of all valid image blocks, thereby unifying the pixel distribution characteristics of the images.

[0120] Somatic mutations in each case were screened using the Mutect2 tool, and synonymous mutations were removed. A set of core driver mutation genes, including APC, KRAS, RELN, ZNF469, and HMCN1, was obtained using a random forest algorithm combined with the TNM staging criteria for colorectal cancer. The morphological contribution coefficients of each gene were calculated: APC = 0.28, KRAS = 0.25, RELN = 0.21, ZNF469 = 0.15, and HMCN1 = 0.11.

[0121] Step 2: Construction of the colon cancer morphological fingerprint database:

[0122] The cluster profile tool was used to perform GO functional enrichment analysis on the core driver mutation gene set. Biological processes with p < 0.05 and related to the morphology of colon cancer tissue cells were screened, and three types of colon cancer-specific morphological features were extracted: glandular fragmentation, abnormal nucleocytoplasmic ratio, and disordered cell arrangement.

[0123] Pathological quantitative modeling was performed on three types of morphological features: the total area of ​​broken glands in the image patch was measured and compared with the total area of ​​normal colonic glands of the same area to calculate the degree of gland fragmentation; the mean nuclear-cytoplasmic ratio of cells in the image patch was measured and combined with the critical value of 0.3 for the nuclear-cytoplasmic ratio of normal colonic epithelial cells to calculate the degree of nuclear-cytoplasmic ratio abnormality; the correlation coefficient of cell arrangement in the image patch was calculated and combined with the critical value of 0.8 for the correlation coefficient of cell arrangement in normal colonic tissue to calculate the degree of cell arrangement disorder.

[0124] The quantified three types of morphological features are fused with the morphological contribution coefficients of the corresponding genes to generate colon morphological fingerprint vectors, and a gene-morphology coupled colon cancer morphological fingerprint database is constructed.

[0125] Step 3: Multiscale morphological feature extraction associated with colorectal cancer

[0126] An improved Inception-V3 network was used, embedding a colorectal cancer feature extraction branch. This branch contains three convolutional kernel groups: a 1×1 convolutional kernel group for extracting cell-level features, a 3×3 convolutional kernel group for extracting gland-level features, and a 5×5 convolutional kernel group for extracting tissue-level features.

[0127] The fusion weight coefficients for each scale are calculated: 0.36 for cell-level features, 0.42 for gland-level features, and 0.22 for tissue-level features. The fused feature map is obtained through the weighted fusion formula.

[0128] PCA was used to reduce the dimensionality of the fused feature map, retaining the principal components with a cumulative contribution rate of 92%. By embedding the colorectal cancer feature attention module in CBAM, the feature regions that match the morphological fingerprint database were enhanced.

[0129] Step 4: Preliminary delineation of candidate tumor regions

[0130] The initial threshold is generated using the improved Otsu algorithm. The algorithm adjusts the inter-class variance weights of normal and tumor tissues in colon cancer to 1.2:1 to adapt to the pixel distribution characteristics of the image.

[0131] The matching degree of each candidate region was calculated using the colon cancer gene morphology matching formula. ,right For candidate regions with a threshold < 0.65, the threshold is adjusted with an iteration step size of 0.05, decreasing by 10% every 5 iterations, until the threshold is reached. The optimal segmentation threshold is obtained when the threshold is ≥0.65 or after 50 iterations. .

[0132] in accordance with The fused feature map was segmented to initially identify tumor candidate regions from 100 samples. .

[0133] Step 5: Candidate Region Boundary Optimization and Noise Filtering

[0134] An improved Canny operator was used to extract the edge contours of each candidate region and calculate the gradient value of the colon cancer edge, where the gradient weight in the x-direction was set to 0.6 and the gradient weight in the y-direction was set to 0.4.

[0135] Using the edge contour as seed points, the boundary is optimized according to the growth criteria for colon cancer regions. The similarity of colon cancer features between adjacent pixels must be ≥0.85 and the edge gradient value ≤0.25 to obtain the region with optimized boundary. .

[0136] extract The morphological feature vectors are used to calculate the consistency score using a consistency verification formula. Remove False positive regions <0.55; combined with the clinicopathological criteria for colorectal cancer, isolated noise regions with an area less than 800 pixels were removed to obtain the optimized tumor candidate region.

[0137] Step 6: Verification, Correction, and Result Output:

[0138] The optimized tumor candidate regions were compared with the manually labeled gold standard to calculate the validation index for colorectal cancer, namely the tumor region crossover ratio. Sensitivity for colon cancer tumor detection Specificity of normal tissue in colon cancer .

[0139] right <0.8 or <0.85 or For samples with a value <0.9, adjust the feature fusion weights and threshold iteration step size, and repeat steps 3-5 until all samples meet the validation metric requirements.

[0140] Calculate the clinical confidence score for each sample. The final annotation results include region coordinates, boundary contours, and clinical confidence scores. The clinical confidence scores of all 100 samples were above 88.

[0141] Step 7: Morphological feature normalization processing:

[0142] All quantified values ​​of colorectal cancer morphological features were uniformly normalized to the [0,1] interval to ensure the consistency and comparability of feature data, providing a standardized data foundation for subsequent model reuse and parameter adjustment.

[0143] The above are all preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Therefore, all equivalent changes made in accordance with the structure, shape and principle of the present invention should be covered within the scope of protection of the present invention.

Claims

1. A method for automatic tumor region annotation in whole-section pathological images of colon cancer, characterized in that, Includes the following steps: Step 1: Obtain the whole-section pathological images of colorectal cancer and the somatic mutation data of the corresponding cases. Perform standardized preprocessing on the whole-section pathological images of colorectal cancer, screen the somatic mutation data to obtain the set of core driver mutation genes of colorectal cancer, and determine the morphological contribution coefficient of each gene. Step 2: Based on the functional enrichment results of the core driver mutation genes in colorectal cancer, extract the morphological feature set of colorectal cancer and establish a quantitative model to generate a colorectal cancer morphological fingerprint vector that integrates gene morphological contribution coefficients, and construct a gene-morphological coupled colorectal cancer morphological fingerprint library. Step 3: Using the improved Inception-V3 network, embedding a colorectal cancer feature extraction branch, extracting cellular, glandular, and tissue-level colorectal cancer-related morphological features from stained whole-section pathological images, and obtaining a fused feature map by weighting the colorectal cancer gene morphological contribution coefficient; Step 4: Based on the colorectal cancer morphological fingerprint vector, the Otsu improved algorithm is used to generate the initial segmentation threshold. The optimal segmentation threshold is obtained by iterative optimization through the colorectal cancer gene morphology matching degree formula, and the tumor candidate region is initially divided. Step 5: Optimize the candidate region boundary using edge detection and region growing algorithms, eliminate false positive regions using the colon cancer gene and morphology consistency verification formula, and eliminate isolated noise regions using colon cancer pathology standards to obtain the optimized tumor candidate region.

2. The method for automatic tumor region annotation in whole-section pathological images of colon cancer according to claim 1, characterized in that, The standardization preprocessing in step 1 includes the following sub-steps: Step 11: Use the OpenSlide tool to divide the whole-section pathological image of colon cancer staining into image blocks according to the set pixels, set the pathological diagnosis magnification, and remove noisy image blocks where the white background ratio exceeds the set first ratio threshold and the effective tissue ratio of colon cancer is lower than the set second ratio threshold. Step 12: Normalize the pixel values ​​of the image block using the Z-score normalization formula; Step 13: Use the Mutect2 tool to screen for somatic mutations in colorectal cancer and remove synonymous mutations. Use the random forest algorithm combined with the TNM staging criteria for colorectal cancer to screen for the set of core driver mutation genes in colorectal cancer. Calculate the morphological contribution coefficient of each gene using the following formula for colorectal cancer gene morphological contribution coefficient. : ;in , where n is the product of the mutation frequency and functional enrichment score of the gene in colorectal cancer, and n is the number of core driver mutation genes.

3. The method for automatic tumor region annotation in whole-section pathological images of colon cancer according to claim 2, characterized in that, Otsu's improved algorithm addresses the difference in pixel proportion between tumor and normal tissue in colorectal cancer pathological images by adjusting the pixel weight coefficients for calculating inter-class variance to 1.2 for tumor tissue and 0.8 for normal tissue.

4. The method for automatic tumor region annotation in whole-section pathological images of colon cancer according to claim 3, characterized in that, Step 2, the construction of the colon cancer morphological fingerprint database, includes the following sub-steps: Step 21: Use the cluster profile tool to perform GO functional enrichment analysis on the core driver mutation gene set of colorectal cancer, screen biological processes with p<0.05 that are related to the morphology of colorectal cancer tissue or cells, and extract the colorectal cancer morphological feature set. Step 22: Perform quantitative pathological modeling of colon cancer based on the specific morphological features of each colon cancer: glandular fragmentation , The total area of ​​the broken glands in the image patch. This represents the total area of ​​normal colonic glands of the same area; nucleo-cylindrical ratio anomaly , This represents the mean nucleus-to-cytoplasm ratio of cells in the image patch. =0.3 is the critical value for the nucleus-to-cytoplasm ratio of normal colonic epithelial cells; Cell arrangement disorder , Correlation coefficient of cell arrangement in image patches The critical value of the correlation coefficient for the normal arrangement of colonic tissue cells; Step 23: Generate colonic morphological fingerprint vector m represents the number of morphological features of colon cancer. To determine the contribution coefficient of corresponding gene morphology, a gene-morphology coupled colorectal cancer morphological fingerprint database was constructed.

5. The method for automatic tumor region annotation in whole-section pathological images of colon cancer according to claim 4, characterized in that, The method for extracting colon cancer-related morphological features from stained whole-section pathological images in step 3 is as follows: A colorectal cancer feature extraction branch was embedded into the Inception-V3 network, containing three sets of convolutional kernels, corresponding to cellular, glandular, and tissue-level colorectal cancer-specific features, respectively. The fused feature map was obtained using a weighted fusion formula based on the colorectal cancer gene morphology contribution coefficient. ,in , For genes The feature contribution weights at the k-th scale This is a feature map of colon cancer at the k-th scale; PCA is used to reduce the dimensionality of the fused feature map, retaining principal components with a cumulative contribution rate of ≥90%, and CBAM is used to enhance the feature regions that match the morphological fingerprint of colorectal cancer by embedding a colorectal cancer feature attention module.

6. The method for automatic tumor region annotation in whole-section pathological images of colon cancer according to claim 5, characterized in that, Step 4, the preliminary delineation of tumor candidate regions, includes the following sub-steps: Step 41: Generate the initial threshold using the improved Otsu algorithm. The pixel class variance weights of normal and tumor tissues in colon cancer were adjusted to 1.2:1 to adapt to the pixel distribution characteristics of colon cancer pathological images. Step 42: Calculate candidate regions using the colon cancer gene morphology matching formula. : ； in Let i be the i-th eigenvalue of the morphological fingerprint vector of colon cancer. Let i be the i-th eigenvalue of the candidate region. The coefficient represents the morphological contribution of the corresponding gene, and m represents the number of colorectal cancer-specific morphological features. The closer the value is to 1, the higher the match between the candidate region and the morphological features of colon cancer. Step 43, if Then according to Adjust the threshold, where =0.05 is the iteration step size, which decreases by 10% every 5 iterations. The iteration terminates under the condition that... Alternatively, the optimal segmentation threshold can be obtained by iterating 50 times. ; Step 44, based on Segmenting the fusion feature map of colon cancer yields preliminary tumor candidate regions. .

7. The method for automatic tumor region annotation in whole-section pathological images of colon cancer according to claim 6, characterized in that, In step 5, the improved Canny operator is used to extract the edge contours of the candidate region and calculate the gradient value of the colon cancer edge; The formula for the improved Canny operator is: ,in , These are the gradients in the x and y directions, respectively. , The gradient method weights for colorectal cancer case images are used, with edge contours as seed points. A colorectal cancer region growth criterion is applied: the similarity of colorectal cancer features between adjacent pixels is greater than or equal to 0.85, and... Less than or equal to 0.25, the optimized region boundary is obtained ; extract The morphological feature vectors are used to calculate the consistency score using a consistency verification formula. ,like The regions identified as false positives were removed; based on the clinical pathological criteria for colon cancer, isolated noise regions with an area less than 800 pixels were removed to obtain the optimized tumor candidate regions.

8. The method for automatic tumor region annotation in whole-section pathological images of colon cancer according to claim 7, characterized in that, The process also includes step 6, which compares the optimized tumor candidate region with the manually annotated results based on the clinical diagnostic criteria for colorectal cancer, calculates the colorectal cancer validation index, adaptively adjusts the model parameters, and outputs the final annotation results of the colorectal cancer tumor region, including region coordinates, boundary contours, and clinical confidence.

9. The method for automatic tumor region annotation in whole-section pathological images of colon cancer according to claim 8, characterized in that, In step 6, based on the TNM staging and pathological diagnostic criteria for colorectal cancer, validation indicators for colorectal cancer are determined, including the crossover ratio of tumor regions. Sensitivity for colon cancer tumor detection Specificity of normal tissue in colon cancer Calculate the clinical confidence score for colon cancer. : 。 10. The method for automatic tumor region annotation in whole-section pathological images of colon cancer according to claim 9, characterized in that, The quantified values ​​of the morphological characteristics of colon cancer were all normalized to the [0,1] interval.

Images