A pathological image-based tumor stroma proportion quantification evaluation system

Abstract

The present application relates to the technical field of digital image processing, in particular to a tumor interstitial proportion quantitative evaluation system based on pathological images. The present application automatically obtains tumor epithelial region, tumor interstitial region, tumor region and tumor infiltration front region based on image segmentation, without pre-selecting histological evaluation position for visual evaluation by artificial selection, avoiding the subjective experience dependence of traditional methods, and improving the efficiency of tumor interstitial proportion evaluation. In addition, the present application comprehensively quantitatively evaluates the first tumor interstitial proportion under the overall scale of the to-be-tested pathological image, the second tumor interstitial proportion under the scale of the tumor infiltration front region and the third tumor interstitial proportion under the scale of the tumor region, and finally obtains the tumor interstitial proportion score value which can reflect the tumor interstitial proportion characteristics of different anatomical positions, further improving the accuracy of the evaluation result. The preset rule is to select the window tumor interstitial proportion with the highest tumor interstitial proportion value.

Description

Technical Field

[0001] This invention relates to the field of digital image processing technology, and in particular to a tumor stroma ratio quantitative assessment system based on pathological images. Background Technology

[0002] Tumor tissue is a heterogeneous complex composed of tumor cells and various tumor stroma components. The tumor stroma is a crucial part of tumor tissue, significantly influencing tumor biological behavior through its dynamic interactions with tumor cells. Studies have shown that tumor cells in solid tumors can actively remodel the surrounding tumor stroma microenvironment, forming niches that support their proliferation and survival. Simultaneously, tumor cells can also transform into the tumor stroma through specific signal transduction pathways, thereby driving tumor angiogenesis and leading to tumor development. Notably, the observation of the tumor stroma plays a key role in critical pathological regulation processes such as tumor occurrence, development, and the formation of treatment resistance, directly affecting the malignant phenotypic characteristics of cancer.

[0003] In the field of colorectal cancer, the incidence and mortality rates continue to rise, necessitating precise treatment strategies to improve patient prognosis. Recent evidence-based medicine has shown a significant correlation between tumor stromal ratio (TSR) and colorectal cancer patient prognosis, serving as a microscopic reflection of the complex interactions between tumor cells and the tumor microenvironment. However, current clinical practice suffers from inaccuracies in TSR calculations. First, TSR assessment requires pre-selection of histological assessment sites, with pathologists visually evaluating HE-stained sections under a microscope, introducing a reliance on subjective experience. Second, TSR values ​​at different anatomical locations exhibit significant prognostic differences, leading to insufficient accuracy and severely limiting the widespread application of TSR in clinical treatment decisions. Therefore, there is an urgent need to improve the accuracy of TSR calculations to enhance its clinical value in colorectal cancer prognosis and individualized treatment. Summary of the Invention

[0004] The present invention aims to provide a tumor stroma ratio quantitative assessment system based on pathological images, so as to improve the accuracy of tumor stroma ratio assessment and solve the technical problems of subjective experience dependence and insufficient accuracy of assessment results in the current calculation of tumor stroma ratio.

[0005] To achieve the above objectives, this invention provides a tumor-stromal ratio quantitative assessment system based on pathological images, applicable to the pathological images to be tested; the system includes an image segmentation module, a tumor invasion front detection module, a first tumor-stromal ratio analysis module, a second tumor-stromal ratio analysis module, a third tumor-stromal ratio analysis module, and a tumor-stromal ratio quantitative assessment module; wherein: The image segmentation module is used to segment the pathological image to be tested to obtain the tumor epithelial region and the tumor stroma region, and to segment the pathological image to be tested to obtain the normal region and the tumor region. The tumor invasion front detection module is used to obtain the tumor invasion front region based on the normal region and the tumor region; The first tumor stroma ratio analysis module is used to obtain the first tumor stroma ratio of the pathological image to be tested based on the tumor epithelial region and the tumor stroma region. The second tumor stroma ratio analysis module is used to obtain the second tumor stroma ratio of the tumor invasion front region based on the tumor epithelial region and the tumor stroma region; The third tumor stroma ratio analysis module is used to divide the tumor region into several window regions based on a preset sliding window, thereby obtaining the window tumor stroma ratio of each window region based on the tumor epithelial region and the tumor stroma region, and then taking the window tumor stroma ratio that meets the preset rules as the third tumor stroma ratio. The tumor stroma ratio quantitative assessment module is used to quantitatively assess the pathological image to be tested based on the first tumor stroma ratio, the second tumor stroma ratio, and the third tumor stroma ratio, and obtain a tumor stroma ratio score.

[0006] This invention automatically acquires the tumor epithelial region, tumor stroma region, tumor region, and tumor invasion front region based on image segmentation. It eliminates the need for pre-selection of histological assessment locations for visual evaluation, avoiding the subjective experience reliance of traditional methods and improving the efficiency of tumor stroma ratio assessment. Furthermore, this invention quantitatively assesses the tumor stroma ratio by comprehensively considering the first tumor stroma ratio at the overall scale of the pathological image, the second tumor stroma ratio at the scale of the tumor invasion front region, and the third tumor stroma ratio at the scale of the tumor region. The resulting tumor stroma ratio score reflects the characteristics of tumor stroma ratio at different anatomical locations, further improving the accuracy of the assessment results. The preset rule selects the window of tumor stroma ratio with the highest value.

[0007] To improve the accuracy of quantitative assessment of pathological images and tumor-stromal ratio scoring, in a preferred embodiment, the aforementioned tumor-stromal ratio quantitative assessment system based on pathological images further includes a fourth tumor-stromal ratio analysis module. This fourth tumor-stromal ratio analysis module is used to divide the tumor invasion front region into several window regions based on a preset sliding window, thereby obtaining the window tumor-stromal ratio for each window region based on the tumor epithelial region and the tumor stromal region, and then using the window tumor-stromal ratio that conforms to preset rules as the fourth tumor-stromal ratio.

[0008] In this embodiment, the tumor stroma ratio quantitative assessment module performs quantitative assessment on the pathological image to be tested based on the first tumor stroma ratio, the second tumor stroma ratio, the third tumor stroma ratio, and the fourth tumor stroma ratio to obtain a tumor stroma ratio score. The final tumor stroma ratio score can more accurately reflect the tumor stroma ratio characteristics of the tumor invasion front region, thereby further improving the accuracy of the assessment results.

[0009] It should be noted that the above-mentioned preset rules are used to determine the window of the largest tumor stroma ratio, and then to determine the window region of the largest tumor stroma ratio within the tumor region, as well as the third tumor media ratio corresponding to the window region.

[0010] Similarly, the aforementioned preset rules are also used to determine the window of the largest tumor stroma ratio, thereby determining the window region of the largest tumor stroma ratio within the tumor infiltration front region, and the fourth tumor mediator ratio corresponding to the window region.

[0011] Further, the segmentation of the pathological image to be tested to obtain the tumor epithelial region and the tumor stroma region includes: Acquire several original pathological images; The tumor epithelial region and tumor stroma region of each original pathological image are identified based on a preset first magnification, and then a first pathological image training set is obtained based on all the original pathological images. The preset semantic segmentation model is trained based on the first pathological image training set to obtain the first pathological region segmentation model; The pathological image to be tested is segmented based on the first pathological region segmentation model to obtain the tumor epithelial region and the tumor stroma region.

[0012] It should be noted that the first magnification is 10×, which is greater than the second magnification, and is used to identify and segment the tumor epithelial region and the tumor stroma region at a more microscopic scale.

[0013] In this implementation, a standardized labeling method is used to uniformly identify the tumor epithelial and stromal regions of the original pathological image at a magnification of 10×, constructing a first pathological image training set with scale consistency. This effectively solves the problem of region identification bias caused by differences in magnification. The pathological region segmentation model trained based on this training set can significantly improve the image segmentation accuracy of tumor epithelial and stromal regions, especially suitable for distinguishing complex tissue structures. Segmenting the pathological image under test based on the first pathological region segmentation model to obtain the tumor epithelial and stromal regions not only reduces the subjective error of visual assessment but also enhances the model's generalization ability through standardized processing, providing a highly reliable tissue region segmentation foundation for subsequent quantitative assessment of tumor-stromal ratio.

[0014] Further, the segmentation of the pathological image to be tested to obtain normal areas and tumor areas includes: Acquire several original pathological images; The normal and tumor regions of each of the original pathological images are identified based on a preset second magnification, and then a second pathological image training set is obtained based on all the original pathological images. The second pathological region segmentation model is obtained by training the preset semantic segmentation model based on the second pathological image training set. The pathological image to be tested is segmented based on the second pathological region segmentation model to obtain normal regions and tumor regions.

[0015] It should be noted that the second magnification is 5×, which is less than the first magnification. This is used to identify and segment normal and tumor regions on a more macroscopic scale, which is beneficial for subsequent accurate localization of tumor invasion boundaries based on normal and tumor regions.

[0016] In this implementation, a second magnification factor of 5× is used to accurately identify normal and tumor regions in the original pathological image, constructing a highly discriminative second pathological image training set. This effectively solves the problem of blurred tumor boundary recognition caused by differences in tissue contrast. The pathological region segmentation model trained based on this training set can significantly improve the segmentation accuracy of normal and tumor tissues, which is beneficial to improving the efficiency of tumor invasion boundary localization. This method reduces the misjudgment rate of tissue morphology by normalizing the magnification factor, while enhancing the model's sensitivity to heterogeneous tumor regions, providing an effective automated tissue partitioning basis for tumor invasion front region identification and segmentation and tumor stroma ratio quantitative assessment.

[0017] Furthermore, obtaining the tumor invasion front region based on the normal region and the tumor region includes: The normal region is subjected to morphological expansion to obtain a normally expanded region, and the tumor region is subjected to morphological expansion to obtain a tumor expanded region. Obtain the overlapping area between the normal expansion region and the tumor expansion region; Obtain the boundary line between the overlapping region and the normal region; Using the extension direction of the boundary line as the expansion direction, morphological expansion operations are performed on the overlapping region to obtain the tumor invasion front region.

[0018] In this implementation, a multi-level morphological dilation operation is used to accurately extract the tumor invasion front region. First, morphological dilation is performed on both the normal region image and the tumor region image to generate an overlapping area. Then, directional dilation is performed on the overlapping area based on the extension direction of the boundary line between the overlapping area and the normal region. This method effectively overcomes the shortcomings of traditional algorithms in accurately identifying regions with blurred tumor boundaries. By constraining the dilation process by the boundary line direction, the extraction results of the front region better reflect the actual biological invasion characteristics of the tumor, significantly improving the accuracy of segmentation and localization of the invasion front region, and providing a precise analytical scale for calculating the tumor stroma ratio at the tumor invasion front.

[0019] Furthermore, the specific process of obtaining the first tumor-stromal ratio of the pathological image to be tested based on the tumor epithelial region and the tumor stroma region is shown in the following expression: The first tumor stroma ratio is: ; in, This refers to the tumor epithelial region within the pathological image. This refers to the tumor stroma region within the pathological image.

[0020] It should be noted that the specific process of obtaining the second tumor stroma ratio of the tumor invasion front region based on the tumor epithelial region and the tumor stroma region is shown in the following expression: The second tumor stroma ratio is: ; in, This refers to the tumor epithelial region within the tumor invasion front region. This refers to the tumor stroma region within the tumor invasion front region.

[0021] Similarly, the specific process of dividing the tumor region into several window regions based on a preset sliding window, and thereby obtaining the window-to-tumor-stromal ratio of each window region based on the tumor epithelial region and the tumor stroma region, is shown in the following expression: The ratio of tumor stroma in the window is: ; in, This refers to the tumor epithelial region within the window area. This refers to the tumor stroma region within the window area.

[0022] In a preferred embodiment, the fourth tumor-stromal ratio analysis module further divides the tumor invasion front region into several window regions based on a preset sliding window. This allows for the acquisition of the window tumor-stromal ratio for each window region based on the tumor epithelial region and the tumor stromal region. The window tumor-stromal ratio that conforms to a preset rule is then used as the fourth tumor-stromal ratio. The specific process for determining the window tumor-stromal ratio is shown in the following expression: The ratio of tumor stroma in the window is: ; in, This refers to the tumor epithelial region within the window area. This refers to the tumor stroma region within the window area.

[0023] Furthermore, the step of dividing the tumor region into several window regions based on a preset sliding window includes: Identify tumor cells in the tumor region; For any given window region, all boundaries of the window region overlap with the tumor cells.

[0024] In this implementation, a preset sliding window of a fixed size is used to traverse the tumor image and extract local features of the tumor region. To prevent the sliding window boundary from touching non-tumor areas, such as necrotic tissue, normal tissue, or blank edges, which could distort the calculated tumor-stromal ratio within the sliding window, this implementation limits the presence of tumor cells at the top, bottom, left, and right boundaries of the sliding window. This ensures that the sliding window is completely within the tumor region, eliminating interference from non-tumor areas and thus improving the accuracy of subsequent calculations of the tumor-stromal ratio.

[0025] Furthermore, it also includes a tumor-stromal ratio prognostic analysis module, which is used to perform the following steps: acquiring a number of original pathological images, and then acquiring a tumor-stromal ratio score for each of the original pathological images; Based on the tumor-stromal ratio score of each original pathological image, the original pathological images are divided into a first group and a second group according to a preset score classification rule. Based on the first group and the second group, a pre-set proportional hazards regression model was trained to obtain the prognostic analysis results of tumor stroma proportion.

[0026] This implementation introduces a tumor-stromal ratio prognostic analysis module, enabling automated quantitative assessment based on the tumor-stromal ratio and automated prognostic analysis of pathological images. This module first divides the images into high and low groups based on the tumor-stromal ratio score of the original pathological images. Then, it uses the grouped data to train a proportional hazards regression model, which outputs the prognostic analysis results for the tumor-stromal ratio. This approach effectively addresses the subjectivity and efficiency bottlenecks of manual assessment in traditional pathological analysis, significantly improving the efficiency and reproducibility of tumor-stromal ratio prognostic analysis.

[0027] Further, the step of training a preset proportional hazards regression model based on the first group and the second group to obtain the tumor stroma proportion prognostic analysis results includes: Obtain the first survival curve for the first group and the second survival curve for the second group; A statistical difference was confirmed between the first survival curve and the second survival curve. Then, a pre-set proportional hazards regression model was trained based on the first group, the first survival curve, the second group, and the second survival curve to obtain the prognostic analysis results of the tumor stroma ratio.

[0028] In this implementation, during the training phase of the proportional hazards regression model, the accuracy of the prognostic analysis results is ensured by acquiring and verifying statistical differences between the survival curves of the first and second groups. This design first generates corresponding survival curves based on tumor-stromal ratio groups. After confirming statistical differences, survival data with significant differences are input into the proportional hazards regression model for training. This technical approach effectively avoids data noise interference caused by ambiguous group boundaries in traditional prognostic models, significantly improving the accuracy of the hazard regression model. Simultaneously, using statistical differences as a pre-validation condition for model training ensures that the output tumor-stromal ratio prognostic analysis results have clear clinical stratification value, ensuring the degree of influence of tumor-stromal ratio and tumor-stromal ratio score on the survival of colorectal cancer patients and their independence compared to other clinical indicators.

[0029] Further, obtaining the first survival curve of the first group and the second survival curve of the second group includes: The first survival curve and the second survival curve were plotted using the Kaplan-Meier method.

[0030] It should be noted that the Kaplan-Meier method is a nonparametric statistical method for survival analysis, which constructs survival curves by calculating the survival probability at time points and its standardized values. This method is based on an ordered sequence of event occurrence times and uses a product-limit estimator to calculate the survival function point-by-point: the survival probability is updated when an event occurs in the sample, and for censored data, its survival information up to the point of censoring is retained. In this invention, this method is applied to the visualization of survival data after tumor stroma proportion grouping, generating clinically interpretable survival curves and providing standardized input for subsequent statistical difference validation and proportional hazards regression model training.

[0031] Furthermore, confirming a statistical difference between the first survival curve and the second survival curve includes: The first and second survival curves were evaluated using the log-rank test to obtain statistical test values. The statistical test value is confirmed to meet the preset statistical threshold, thereby confirming that there is a statistical difference between the first survival curve and the second survival curve.

[0032] It should be noted that the log-rank test is a standard nonparametric statistical test method in survival analysis, used to compare the statistical significance of differences between two or more groups of survival curves. Its core advantage lies in its equal weighting of all event time points and its effective handling of censored data. In this invention, this method is applied to verify the statistical difference between the survival curves of the first and second groups, providing statistical significance for training the proportional hazards regression model and ensuring the accuracy of the prognostic analysis results. Attached Figure Description

[0033] Figure 1 This is a schematic diagram of the structure of a tumor stroma ratio quantitative assessment system based on pathological images provided in an embodiment of the present invention; Figure 2 This is a schematic diagram of another tumor stroma ratio quantitative assessment system based on pathological images provided in an embodiment of the present invention; Figure 3 This is an HE staining image provided in an embodiment of the present invention; Figure 4 This is a schematic diagram illustrating the precise identification of normal and tumor regions in an HE-stained image at 5× magnification, as provided in an embodiment of the present invention. Figure 5 This is a schematic diagram illustrating the identification of the tumor invasion front region based on normal and tumor regions, provided by an embodiment of the present invention. Figure 6This is a schematic diagram of the segmentation of 12 tissue types in a pathological image to be tested under HE staining at 10× magnification, as provided in an embodiment of the present invention. Figure 7 This is a partial schematic diagram of an HE staining image provided in an embodiment of the present invention; Figure 8 This is a partial schematic diagram of the segmentation of 12 tissue types in an HE staining image provided by an embodiment of the present invention. Detailed Implementation

[0034] The present invention will now be described in detail with reference to the accompanying drawings and embodiments. It should be noted that the following detailed descriptions are exemplary and intended to provide further detailed explanation of the invention. 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 application belongs; the terminology used herein in the specification is for the purpose of describing particular embodiments only and is not intended to limit the application; the terms "comprising" and "having," and any variations thereof, in the specification, claims, and foregoing drawings, are intended to cover non-exclusive inclusion. The terms "first," "second," etc., in the specification, claims, or foregoing drawings are used to distinguish different objects, not to describe a particular order.

[0035] It should be understood that although the steps in the flowcharts of the accompanying figures are shown sequentially as indicated by the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the accompanying figures may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but can be executed at different times, and their execution order is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the sub-steps or stages of other steps.

[0036] The incidence and mortality rates of colorectal cancer continue to rise annually, and further research and validation are needed to advance prognostic studies in colorectal cancer precision diagnosis and treatment. Studies have demonstrated that the tumor-to-stromal ratio is associated with the prognosis of colorectal cancer patients. The tumor-to-stromal ratio may be a microscopic reflection of the complex biological processes involving the interaction between tumor cells and the tumor microenvironment. Given the current variability in the location of tumor-to-stromal ratio assessments, further validation and standardization are crucial for improving the clinical significance of the tumor-to-stromal ratio in predicting prognosis and treatment decisions for colorectal cancer. This invention utilizes deep learning technology to automatically identify and calculate the tumor-to-stromal ratio at different locations, obtaining a comprehensive tumor-to-stromal score for risk stratification of patients, assisting physicians in developing more appropriate treatment plans, and enabling precise diagnosis and treatment of colorectal cancer patients.

[0037] The main objective of this invention is to overcome the shortcomings and deficiencies of the prior art and provide a method for automatically calculating the tumor stroma ratio and tumor stroma score based on digital pathological slides. This method can automatically calculate the tumor stroma ratio at different locations, which can greatly save the time and manpower costs of quantitative tumor stroma scoring for colorectal cancer and reduce inter-observer variability.

[0038] To achieve the above objectives, refer to Figure 1 and Figure 4 This invention provides a tumor stroma ratio quantitative assessment system based on pathological images, applicable to pathological images under test; the system includes an image segmentation module 100, a tumor invasion front detection module 200, a first tumor stroma ratio analysis module 310, a second tumor stroma ratio analysis module 320, a third tumor stroma ratio analysis module 330, and a tumor stroma ratio quantitative assessment module 400; wherein: The image segmentation module 100 is used to segment the pathological image to be tested to obtain the tumor epithelial region and the tumor stroma region, and to segment the pathological image to be tested to obtain the normal region and the tumor region. The tumor invasion front detection module 200 is used to obtain the tumor invasion front region based on the normal region and the tumor region; The first tumor stroma ratio analysis module 310 is used to obtain the first tumor stroma ratio of the pathological image to be tested based on the tumor epithelial region and the tumor stroma region. The second tumor stroma ratio analysis module 320 is used to obtain the second tumor stroma ratio of the tumor invasion front region based on the tumor epithelial region and the tumor stroma region; The third tumor stroma ratio analysis module 330 is used to divide the tumor region into several window regions based on a preset sliding window, thereby obtaining the window tumor stroma ratio of each window region based on the tumor epithelial region and the tumor stroma region, and then taking the window tumor stroma ratio that meets the preset rules as the third tumor stroma ratio. The tumor stroma ratio quantitative assessment module 400 is used to quantitatively assess the pathological image to be tested based on the first tumor stroma ratio, the second tumor stroma ratio, and the third tumor stroma ratio to obtain a tumor stroma ratio score.

[0039] It should be noted that, in this embodiment, the pathological image to be tested is an HE-stained slide image (Hematoxylin and eosin (H&E) stains, H&E stains).

[0040] This invention automatically acquires the tumor epithelial region, tumor stroma region, tumor region, and tumor invasion front region based on image segmentation. It eliminates the need for pre-selection of histological assessment locations for visual evaluation, avoiding the subjective experience reliance of traditional methods and improving the efficiency of tumor stroma ratio assessment. Furthermore, this invention quantitatively assesses the tumor stroma ratio by comprehensively considering the first tumor stroma ratio at the overall scale of the pathological image, the second tumor stroma ratio at the scale of the tumor invasion front region, and the third tumor stroma ratio at the scale of the tumor region. The resulting tumor stroma ratio score reflects the characteristics of tumor stroma ratio at different anatomical locations, further improving the accuracy of the assessment results. The preset rule selects the window of tumor stroma ratio with the highest value.

[0041] Specifically, the tumor stroma ratio score can be obtained by comprehensively considering the stroma content of the entire pathological image slice under test, the tumor invasion front area, and the area with the highest tumor stroma ratio within the tumor region, thus obtaining a more accurate quantitative assessment result of the tumor stroma ratio.

[0042] Among them, the preset rules are used to determine the window of the largest tumor stroma ratio, and then determine the window region of the largest tumor stroma ratio within the tumor region, as well as the third tumor media ratio corresponding to the window region.

[0043] In a preferred embodiment, after defining the tumor region obtained by segmenting the tumor lesion area under 5× magnification of the full-view pathological section, the tissue composition within each window is calculated by operating a sliding window of a fixed size, and the area with the highest proportion of tumor stroma within the tumor region is automatically identified.

[0044] refer to Figure 2To improve the accuracy of quantitative assessment of pathological images and tumor-stromal ratio scoring, in a preferred embodiment, the aforementioned tumor-stromal ratio quantitative assessment system based on pathological images further includes a fourth tumor-stromal ratio analysis module 340. The fourth tumor-stromal ratio analysis module 340 is used to divide the tumor invasion front region into several window regions based on a preset sliding window, thereby obtaining the window tumor-stromal ratio for each window region based on the tumor epithelial region and the tumor stromal region, and then using the window tumor-stromal ratio that conforms to preset rules as the fourth tumor-stromal ratio.

[0045] Among them, the preset rules are used to determine the window of the largest tumor stroma ratio, and then determine the window region of the largest tumor stroma ratio in the tumor infiltration front region, as well as the fourth tumor media ratio corresponding to the window region.

[0046] In this embodiment, the tumor stroma ratio quantitative assessment module 400 performs quantitative assessment on the pathological image to be tested based on the first tumor stroma ratio, the second tumor stroma ratio, the third tumor stroma ratio, and the fourth tumor stroma ratio to obtain a tumor stroma ratio score. The final tumor stroma ratio score can more accurately reflect the tumor stroma ratio characteristics of the tumor invasion front region, thereby further improving the accuracy of the assessment results.

[0047] Further, the segmentation of the pathological image to be tested to obtain the tumor epithelial region and the tumor stroma region includes: Acquire several original pathological images; The tumor epithelial region and tumor stroma region of each original pathological image are identified based on a preset first magnification, and then a first pathological image training set is obtained based on all the original pathological images. The preset semantic segmentation model is trained based on the first pathological image training set to obtain the first pathological region segmentation model; The pathological image to be tested is segmented based on the first pathological region segmentation model to obtain the tumor epithelial region and the tumor stroma region.

[0048] It should be noted that the first magnification is 10×, which is greater than the second magnification, and is used to identify and segment the tumor epithelial region and the tumor stroma region at a more microscopic scale.

[0049] It should be understood that the original pathological images are HE-stained slide images (Hematoxylin and eosin (H&E) stains, H&E stains).

[0050] Please refer to Figure 6In a preferred embodiment, considering that the original pathological image includes not only tumor epithelial tissue and tumor stroma tissue, but also many different cellular tissues and structures, in order to improve the accuracy of the first pathological region segmentation model in segmenting and identifying the tumor epithelial region and the tumor stroma region, the first pathological region segmentation model also needs to be able to identify other cellular tissue structures present in the pathological image to avoid misjudging the tumor epithelium or tumor stroma. Therefore, this embodiment first reads the tissue categories of HE-stained pathological images of colorectal cancer patients at 10× magnification; 12 tissue types are defined in the HE-stained original pathological image WSI (Whole Slide Imaging), including tumor epithelium, tumor stroma, background, necrosis, normal glands, normal stroma, submucosa or serosa, muscle, lymphocyte aggregates, mucus, fat, and blood; then, the semantic segmentation model is trained using the read 10× HE-stained pathological image and the labeled ROI (Region of Interest) of the 12 tissue types, wherein the nn U-net semantic segmentation model is preferably used for training. This yields a trained automatic segmentation model for 10×HE-stained pathological images and enables... Figure 3 The HE-stained images shown are segmented into 12 tissue types, such as... Figure 6 As shown, the segmentation results identified tumor epithelium, tumor stroma, necrosis, normal glands, normal stroma, submucosa, muscle, lymphoid aggregates, mucus, fat, and blood. Ultimately, Figure 6 The tumor epithelial region and tumor stroma region in the image segmentation results shown are the image recognition results required by this invention, which is helpful for identifying the tissue components at various locations in the pathological image to be tested.

[0051] More specifically, Figure 7 Yes Figure 3 The diagram shown is a magnified view of a portion of the HE staining image. Figure 8 yes Figure 7 The diagram shows the results of segmenting a local area into 12 tissue types. The segmentation results in this local image identify tumor epithelium, tumor stroma, and necrotic areas.

[0052] In this embodiment, by uniformly standardizing the tumor epithelial and stromal regions of the original pathological images using a 10× magnification, a first pathological image training set with scale consistency is constructed, effectively solving the problem of region identification bias caused by differences in magnification. The pathological region segmentation model trained based on this training set can significantly improve the image segmentation accuracy of tumor epithelial and stromal regions, especially suitable for distinguishing complex tissue structures. Segmenting the pathological image under test based on the first pathological region segmentation model to obtain the tumor epithelial and stromal regions not only reduces the subjective error of visual assessment but also enhances the model's generalization ability through standardized processing, providing a highly reliable tissue region segmentation foundation for subsequent quantitative assessment of the tumor-stromal ratio.

[0053] Figure 3 This paper demonstrates an automated process for analyzing and segmenting pathological images based on artificial intelligence.

[0054] (a): HE staining image of the original full-field section.

[0055] (b): The tissue was initially segmented under a 5x microscope to distinguish tumor tissue from normal tissue.

[0056] (c): The tumor invasion front (black curve) is automatically identified by the algorithm.

[0057] (d): The tissue is finely segmented under a 10x microscope to distinguish various tissue components such as tumors, stroma, necrosis, normal glands, blood vessels, and fat.

[0058] (e): Magnified view of HE staining area.

[0059] (f): Schematic diagram of the automatic segmentation results of the corresponding region. Tumors, stroma and other tissues are automatically labeled with different gray levels.

[0060] Further, the segmentation of the pathological image to be tested to obtain normal areas and tumor areas includes: Acquire several original pathological images; The normal and tumor regions of each of the original pathological images are identified based on a preset second magnification, and then a second pathological image training set is obtained based on all the original pathological images. The second pathological region segmentation model is obtained by training the preset semantic segmentation model based on the second pathological image training set. The pathological image to be tested is segmented based on the second pathological region segmentation model to obtain normal regions and tumor regions.

[0061] It should be noted that the second magnification is 5×, which is less than the first magnification. This is used to identify and segment normal and tumor regions on a more macroscopic scale, which is beneficial for subsequent accurate localization of tumor invasion boundaries based on normal and tumor regions.

[0062] It should be understood that the original pathological images are HE-stained slide images (Hematoxylin and eosin (H&E) stains, H&E stains).

[0063] Please refer to Figure 3 and Figure 4 In a preferred embodiment, considering that the original pathological image includes not only tumor epithelial tissue and tumor stroma tissue, but also many different cellular tissues and structures, in order to improve the accuracy of the first pathological region segmentation model in segmenting and identifying tumor epithelial and tumor stroma regions, the first pathological region segmentation model also needs to be able to identify other cellular tissue structures present in the pathological image to avoid misjudging tumor epithelium or tumor stroma. Therefore, this embodiment first reads the tissue categories of HE-stained pathological images of colorectal cancer patients at 5× magnification; three tissue types are defined in the HE-stained original pathological image WSI (Whole Slide Imaging), including tumor region, normal region, and background region; then, the semantic segmentation model is trained using the read 5× HE-stained pathological image and the labeled ROI (Region of Interest) of the tumor region, normal region, and background region. Preferably, the nn U-net semantic segmentation model is used for training, thereby obtaining the trained 5× HE-stained pathological image automatic segmentation model and performing segmentation on other parts of the image. Figure 3 The HE-stained image shown is segmented into three tissue types. The normal and tumor regions in the final image segmentation result are the image recognition results required by this invention, i.e., as shown below. Figure 4 The green area representing normal tissue and the red area representing tumor tissue are shown to help identify the boundary region at the tumor infiltration front in the pathological image being tested.

[0064] In this embodiment, by employing a 5× second magnification to accurately identify normal and tumor regions in the original pathological image, a highly discriminative second pathological image training set is constructed, effectively resolving the problem of blurred tumor boundary recognition caused by differences in tissue contrast. The pathological region segmentation model trained based on this training set significantly improves the segmentation accuracy between normal and tumor tissues, thus enhancing the efficiency of tumor invasion boundary localization. This method reduces the misjudgment rate of tissue morphology through normalized magnification, while simultaneously enhancing the model's sensitivity to heterogeneous tumor regions, providing an effective automated tissue partitioning foundation for tumor invasion front region identification and segmentation and quantitative assessment of tumor stroma proportions.

[0065] In one possible embodiment, the training set is formed by a professional pathologist drawing three different types of tissue categories on a colorectal cancer HE-stained digital pathological image used for training: tumor area, normal area, and background area. Then, several sub-image blocks of a set size are extracted from each type of tissue area as training samples. Next, manual features are extracted from the training samples to construct feature vectors, and a support vector machine model is used to construct a classifier, thus completing the training of the semantic segmentation model.

[0066] Please refer to Figure 4 and Figure 5 Furthermore, obtaining the tumor invasion front region based on the normal region and the tumor region includes: The normal region is subjected to morphological expansion to obtain a normally expanded region, and the tumor region is subjected to morphological expansion to obtain a tumor expanded region. Obtain the overlapping area between the normal expansion region and the tumor expansion region; Obtain the boundary line between the overlapping region and the normal region; Using the extension direction of the boundary line as the expansion direction, morphological expansion operations are performed on the overlapping region to obtain the tumor invasion front region.

[0067] This embodiment determines the location of the boundary between the tumor area and the normal area, as well as the location of the tumor invasion front, based on the distribution of tumor areas and normal areas in HE-stained digital pathological images of colorectal cancer.

[0068] In a preferred embodiment, morphological dilation is first performed on the binary images of the tumor region and the normal region of the pathological image to be tested. Then, the intersection of the two processed regions is taken as the overlapping region, which is defined as the tumor boundary line in a physiological sense, with a width of 5 pixels. Next, the boundary line coordinates of the boundary between the tumor region and the normal region are found as the boundary line coordinates between the tumor region and the normal region. Finally, the tumor boundary line is appropriately dilated along the vertical direction by a set distance to complete the localization of the tumor infiltration front region.

[0069] In this embodiment, the tumor invasion front region is precisely extracted through multi-level morphological dilation operations. First, the tumor invasion front region is... Figure 4 Morphological dilation is performed on normal tissue images and tumor tissue images respectively to generate overlapping regions. Then, directional dilation is performed on the overlapping regions based on the extension direction of the boundary line between the overlapping regions and normal regions, resulting in the following image. Figure 5 The tumor invasion front region is shown. This method effectively overcomes the shortcomings of traditional algorithms in accurately identifying regions with blurred tumor boundaries. By constraining the expansion process along the boundary line direction, the extraction results of the front region better reflect the actual biological invasion characteristics of the tumor, significantly improving the accuracy of segmentation and localization of the invasion front region, and providing a precise analytical scale for calculating the tumor stroma ratio at the tumor invasion front.

[0070] Furthermore, the specific process of obtaining the first tumor-stromal ratio of the pathological image to be tested based on the tumor epithelial region and the tumor stroma region is shown in the following expression: The first tumor stroma ratio is: ; in, This refers to the tumor epithelial region within the pathological image. This refers to the tumor stroma region within the pathological image.

[0071] It should be noted that the specific process of obtaining the second tumor stroma ratio of the tumor invasion front region based on the tumor epithelial region and the tumor stroma region is shown in the following expression: The second tumor stroma ratio is: ; in, This refers to the tumor epithelial region within the tumor invasion front region. This refers to the tumor stroma region within the tumor invasion front region.

[0072] Similarly, the specific process of dividing the tumor region into several window regions based on a preset sliding window, and thereby obtaining the window-to-tumor-stromal ratio of each window region based on the tumor epithelial region and the tumor stroma region, is shown in the following expression: The ratio of tumor stroma in the window is: ; in, This refers to the tumor epithelial region within the window area. This refers to the tumor stroma region within the window area.

[0073] Specifically, the tumor stroma ratio is defined as the proportion of the area of ​​the tumor epithelial region to the sum of the areas of the tumor epithelial region and the tumor stroma region.

[0074] In a preferred embodiment, the fourth tumor-stromal ratio analysis module 340 further needs to divide the tumor invasion front region into several window regions based on a preset sliding window, thereby obtaining the window tumor-stromal ratio of each window region based on the tumor epithelial region and the tumor stromal region, and then using the window tumor-stromal ratio that conforms to a preset rule as the fourth tumor-stromal ratio. The specific process of the window tumor-stromal ratio is shown in the following expression: The ratio of tumor stroma in the window is: ; in, This refers to the tumor epithelial region within the window area. This refers to the tumor stroma region within the window area.

[0075] In a preferred embodiment, based on the tissue category identification results of HE-stained pathological images of colorectal cancer patients at 10× magnification, the tumor epithelial region and the tumor stroma region are obtained, and the proportion of the tumor epithelial region to the total tumor epithelial region and tumor stroma region is calculated, which is the proportion of tumor stroma in the entire pathological section to be tested.

[0076] In a preferred embodiment, the tissue category identification results at 10× magnification of HE-stained pathological images of colorectal cancer patients are combined to obtain the tumor epithelial region and tumor stroma region within the tumor invasion front. The proportion of the tumor epithelial region to the tumor epithelial region and tumor stroma region within the tumor invasion front region is calculated, which is the tumor stroma ratio of the tumor invasion front region.

[0077] In a preferred embodiment, after defining the tumor region, a fixed-size sliding window operation is used to ensure that tumor cells must exist at all boundaries of the sliding window. Combined with the tissue category identification results of HE-stained pathological images of colorectal cancer patients at 10× magnification, the proportion of the tumor epithelial region to the tumor epithelial region and the tumor stroma region within the region with the highest proportion of tumor stroma is calculated, which is the tumor stroma proportion of the region with the highest proportion of tumor stroma within the tumor region.

[0078] Furthermore, the step of dividing the tumor region into several window regions based on a preset sliding window includes: Identify tumor cells in the tumor region; For any given window region, all boundaries of the window region overlap with the tumor cells.

[0079] In this embodiment, a preset sliding window traverses the tumor image through a fixed-size sliding window to extract local features of the tumor region. To prevent the sliding window boundary from touching non-tumor areas, such as necrotic tissue, normal tissue, or blank edges, which could distort the calculated tumor-stromal ratio within the sliding window, this embodiment limits the presence of tumor cells at the top, bottom, left, and right boundaries of the sliding window. This ensures that the sliding window is completely within the tumor region, eliminating interference from non-tumor areas and thus improving the accuracy of subsequent calculations of the tumor-stromal ratio within the window.

[0080] Furthermore, it also includes a tumor-stromal ratio prognostic analysis module 500, which is used to perform the following steps: acquiring a number of original pathological images, and then acquiring a tumor-stromal ratio score value for each of the original pathological images; Based on the tumor-stromal ratio score of each original pathological image, the original pathological images are divided into a first group and a second group according to a preset score classification rule. Based on the first group and the second group, a pre-set proportional hazards regression model was trained to obtain the prognostic analysis results of tumor stroma proportion.

[0081] In this embodiment, by introducing a tumor-stromal ratio prognostic analysis module 500, automated quantitative assessment based on the tumor-stromal ratio is achieved, enabling automated prognostic analysis of pathological images. This module first divides the images into two groups based on the tumor-stromal ratio score of the original pathological images: those with high scores and those with low scores. Then, it uses the grouped data to train a proportional hazards regression model, which outputs the prognostic analysis results for the tumor-stromal ratio. This approach effectively addresses the subjective bias and efficiency bottleneck of manual assessment in traditional pathological analysis, significantly improving the efficiency and reproducibility of tumor-stromal ratio prognostic analysis.

[0082] Further, the step of training a preset proportional hazards regression model based on the first group and the second group to obtain the tumor stroma proportion prognostic analysis results includes: Obtain the first survival curve for the first group and the second survival curve for the second group; A statistical difference was confirmed between the first survival curve and the second survival curve. Then, a pre-set proportional hazards regression model was trained based on the first group, the first survival curve, the second group, and the second survival curve to obtain the prognostic analysis results of the tumor stroma ratio.

[0083] In this embodiment, during the training phase of the proportional hazards regression model, the accuracy of the prognostic analysis results is ensured by acquiring and verifying statistical differences between the survival curves of the first and second groups. This design first generates corresponding survival curves based on tumor-stromal ratio groups. Only after confirming statistical differences are the survival data with significant differences input into the proportional hazards regression model for training. This technical approach effectively avoids data noise interference caused by ambiguous group boundaries in traditional prognostic models, significantly improving the accuracy of the hazard regression model. Simultaneously, using statistical differences as a pre-validation condition for model training ensures that the output tumor-stromal ratio prognostic analysis results have clear clinical stratification value, ensuring the degree of influence of tumor-stromal ratio and tumor-stromal ratio score as an influencing factor on the survival of colorectal cancer patients and their independence compared to other clinical indicators.

[0084] Further, obtaining the first survival curve of the first group and the second survival curve of the second group includes: The first survival curve and the second survival curve were plotted using the Kaplan-Meier method.

[0085] It should be noted that the Kaplan-Meier method is a nonparametric statistical method for survival analysis, which constructs survival curves by calculating the survival probability at time points and its standardized values. This method is based on an ordered sequence of event occurrence times and uses a product-limit estimator to calculate the survival function point-by-point: the survival probability is updated when an event (such as patient death) occurs in the sample, and for censored data, its survival information up to the point of censoring is retained. In this invention, this method is applied to the visualization of survival data after tumor stroma proportion grouping, generating clinically interpretable survival curves and providing standardized input for subsequent statistical difference validation and proportional hazards regression model training.

[0086] Furthermore, confirming a statistical difference between the first survival curve and the second survival curve includes: The first and second survival curves were evaluated using the log-rank test to obtain statistical test values. The statistical test value is confirmed to meet the preset statistical threshold, thereby confirming that there is a statistical difference between the first survival curve and the second survival curve.

[0087] It should be noted that the log-rank test is a standard nonparametric statistical test method in survival analysis, used to compare the statistical significance of differences between two or more groups of survival curves. Its core advantage lies in its equal weighting of all event time points and its effective handling of censored data. In this invention, this method is applied to verify the statistical difference between the survival curves of the first and second groups, providing statistical significance for training the proportional hazards regression model and ensuring the accuracy of the prognostic analysis results.

[0088] Furthermore, the preset proportional hazards regression model is a Cox proportional hazards regression model, which can evaluate the degree of influence of tumor stroma ratio and tumor stroma ratio score as influencing factors on the survival of colorectal cancer patients and their independence compared with other clinical indicators.

[0089] The Cox proportional hazards regression model is a semi-parametric survival analysis method used to quantify the impact of multiple factors on patient survival time. This model solves for regression coefficients using partial likelihood estimation, without requiring a pre-defined specific form of the baseline hazard function; its core assumption is the proportional hazards principle, meaning that the hazard ratio of any individual remains constant over time. This model can effectively handle missing data in clinical follow-up and outputs the hazard ratios of covariates, i.e., exponential regression coefficients, used to quantify the multiplicative effect of factors on survival risk. In this invention, the model is trained based on statistically validated survival curves of tumor-stromal proportion groups to generate prognostic risk predictors, which serve as the results of the tumor-stromal proportion prognostic analysis.

[0090] The tumor stroma ratio quantitative assessment system based on pathological images provided by the present invention has at least the following advantages compared with the prior art: This invention statistically correlates the tumor stroma ratio with the stroma location, automatically calculates the tumor stroma ratio at different locations, and integrates them to form a tumor stroma score. This reduces the inter-observer variability in the tumor stroma ratio and tumor stroma score of colorectal cancer, and further obtains the tumor stroma ratio and tumor stroma score as prognostic factors. Based on the Kaplan-Meier method, log-rank test, and Cox proportional hazards regression model, the prognostic analysis results of the tumor stroma ratio are obtained, thus improving the accuracy of prognostic analysis.

[0091] The term "embodiment" as used herein means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a mutually exclusive, independent, or alternative embodiment. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described; however, any combination of these technical features that does not contradict each other should be considered within the scope of this specification.

[0092] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this application. It should be noted that those skilled in the art can make various improvements and substitutions without departing from the concept of this application, and these improvements and substitutions should also be considered within the scope of protection of this invention. Therefore, the scope of protection of this application should be determined by the appended claims.

Claims

1. A tumor stroma ratio quantitative assessment system based on pathological images, characterized in that, Applicable to pathological images under test; the system includes an image segmentation module, a tumor invasion front detection module, a first tumor-stromal ratio analysis module, a second tumor-stromal ratio analysis module, a third tumor-stromal ratio analysis module, and a tumor-stromal ratio quantitative assessment module; wherein: The image segmentation module is used to segment the pathological image to be tested to obtain the tumor epithelial region and the tumor stroma region, and to segment the pathological image to be tested to obtain the normal region and the tumor region. The tumor invasion front detection module is used to obtain the tumor invasion front region based on the normal region and the tumor region; The first tumor stroma ratio analysis module is used to obtain the first tumor stroma ratio of the pathological image to be tested based on the tumor epithelial region and the tumor stroma region. The second tumor stroma ratio analysis module is used to obtain the second tumor stroma ratio of the tumor invasion front region based on the tumor epithelial region and the tumor stroma region; The third tumor-stromal ratio analysis module is used to identify tumor cells in the tumor region, divide the tumor region into several window regions based on a preset sliding window, and obtain the tumor-stromal ratio of each window region based on the tumor epithelial region and the tumor stromal region. Then, the window tumor-stromal ratio that meets the preset rules is taken as the third tumor-stromal ratio. Wherein, for any window region, all boundaries of the window region overlap with the tumor cells. The preset rules are used to determine the window tumor-stromal ratio with the largest tumor-stromal ratio, and then determine the window region with the largest tumor-stromal ratio in the tumor region, as well as the third tumor mediator ratio corresponding to the window region. The tumor stroma ratio quantitative assessment module is used to quantitatively assess the pathological image to be tested based on the first tumor stroma ratio, the second tumor stroma ratio, and the third tumor stroma ratio, and obtain a tumor stroma ratio score.

2. The tumor stroma ratio quantitative assessment system based on pathological images according to claim 1, characterized in that, The segmentation of the pathological image to be tested to obtain the tumor epithelial region and the tumor stroma region includes: Acquire several original pathological images; The tumor epithelial region and tumor stroma region of each original pathological image are identified based on a preset first magnification, and then a first pathological image training set is obtained based on all the original pathological images. The preset semantic segmentation model is trained based on the first pathological image training set to obtain the first pathological region segmentation model; The pathological image to be tested is segmented based on the first pathological region segmentation model to obtain the tumor epithelial region and the tumor stroma region.

3. The tumor stroma ratio quantitative assessment system based on pathological images according to claim 1, characterized in that, The segmentation of the pathological image to be tested to obtain normal areas and tumor areas includes: Acquire several original pathological images; The normal and tumor regions of each of the original pathological images are identified based on a preset second magnification, and then a second pathological image training set is obtained based on all the original pathological images. The second pathological region segmentation model is obtained by training the preset semantic segmentation model based on the second pathological image training set. The pathological image to be tested is segmented based on the second pathological region segmentation model to obtain normal regions and tumor regions.

4. The tumor stroma ratio quantitative assessment system based on pathological images according to claim 1, characterized in that, The process of obtaining the tumor invasion front region based on the normal region and the tumor region includes: The normal region is subjected to morphological expansion to obtain a normally expanded region, and the tumor region is subjected to morphological expansion to obtain a tumor expanded region. Obtain the overlapping area between the normal expansion region and the tumor expansion region; Obtain the boundary line between the overlapping region and the normal region; Using the extension direction of the boundary line as the expansion direction, morphological expansion operations are performed on the overlapping region to obtain the tumor invasion front region.

5. The tumor stroma ratio quantitative assessment system based on pathological images according to claim 1, characterized in that, The specific process of obtaining the first tumor-stromal ratio of the pathological image to be tested based on the tumor epithelial region and the tumor stroma region is shown in the following expression: The first tumor stroma ratio is: ; in, This refers to the tumor epithelial region within the pathological image. This refers to the tumor stroma region within the pathological image.

6. The tumor stroma ratio quantitative assessment system based on pathological images according to claim 1, characterized in that, It also includes a tumor-stromal ratio prognostic analysis module, which is used to perform the following steps: acquiring several original pathological images, and then acquiring the tumor-stromal ratio score value of each original pathological image; Based on the tumor-stromal ratio score of each original pathological image, the original pathological images are divided into a first group and a second group according to a preset score classification rule. Based on the first group and the second group, a pre-set proportional hazards regression model was trained to obtain the prognostic analysis results of tumor stroma proportion.

7. The tumor stroma ratio quantitative assessment system based on pathological images according to claim 6, characterized in that, The process of training a pre-defined proportional hazards regression model based on the first group and the second group to obtain the prognostic analysis results of the tumor stroma proportion includes: Obtain the first survival curve for the first group and the second survival curve for the second group; A statistical difference was confirmed between the first survival curve and the second survival curve. Then, a pre-set proportional hazards regression model was trained based on the first group, the first survival curve, the second group, and the second survival curve to obtain the prognostic analysis results of the tumor stroma ratio.

8. The tumor stroma ratio quantitative assessment system based on pathological images according to claim 7, characterized in that, The step of obtaining the first survival curve of the first group and the second survival curve of the second group includes: The first survival curve and the second survival curve were plotted using the Kaplan-Meier method.

9. A tumor stroma ratio quantitative assessment system based on pathological images according to claim 7, characterized in that, The confirmation that there is a statistical difference between the first survival curve and the second survival curve includes: The first and second survival curves were evaluated using the log-rank test to obtain statistical test values. The statistical test value is confirmed to meet the preset statistical threshold, thereby confirming that there is a statistical difference between the first survival curve and the second survival curve.

Images