Immunohistochemistry response prediction system based on dynamic analysis

Abstract

The application relates to the technical field of digital pathology and artificial intelligence, in particular to an immunohistochemical efficacy prediction system based on dynamic analysis. The system comprises an image processing module, which is used for acquiring and preprocessing paired immunohistochemical digital pathology sections before and after treatment of the same patient; a dynamic registration and analysis module, which is used for high-precision image registration of the paired sections and extraction of dynamic quantitative parameters of biomarker expression changes with treatment; a spatial heterogeneity quantification module, which is used for dividing tumor functional subareas and quantifying spatial distribution heterogeneity parameters of biomarkers in different functional subareas; and a fusion prediction module, which is used for feature fusion of the dynamic quantitative parameters and the spatial distribution heterogeneity parameters and output of a clinical efficacy prediction result through a survival analysis model. The system realizes accurate prediction of tumor immunotherapy efficacy by performing dynamic and quantitative analysis on paired immunohistochemical sections before and after treatment and integrating multi-modal data.

Description

Technical Field

[0001] This application relates to the fields of digital pathology and artificial intelligence technology, and in particular to an immunohistochemical efficacy prediction system based on dynamic analysis. This system achieves accurate prediction of the efficacy of tumor immunotherapy by dynamically and quantitatively analyzing paired immunohistochemical sections before and after treatment and integrating multimodal data. Background Technology

[0002] In recent years, deep learning-based artificial intelligence technology has been widely applied in digital pathological image analysis to assist in tumor diagnosis and prognosis prediction. For example, the SMMILe framework, jointly proposed by Xi'an Jiaotong University and the University of Cambridge, can locate, delineate boundaries, and analyze subtype distribution of tumor regions in pathological sections using only patient-level labels (weak supervision). Its innovation lies in solving the problem of traditional models that "emphasize classification but neglect localization." However, this framework is mainly for single static sections and cannot analyze the dynamic evolution of biomarkers during treatment.

[0003] The BEPH pathological model developed by Yu Zhangsheng's team at Shanghai Jiao Tong University improves the model's generalization ability across different cancer diagnostic tasks by learning common features from massive amounts of unlabeled data through self-supervised learning. However, such models are insufficient in quantifying the spatial heterogeneity of cell distribution (such as the difference between the tumor core and the invasive border) when analyzing the tumor microenvironment.

[0004] On the other hand, RuiPath, a multimodal interactive large model jointly developed by Ruijin Hospital in Shanghai and Huawei, integrates massive amounts of slide data and clinical knowledge, improving the efficiency of clinical slide reading. However, it still faces challenges in the deep integration of multimodal data (such as imaging, genome, and pathology), and the model's performance is easily affected by differences in staining and scanning standards among different medical centers, and its generalization ability needs to be strengthened.

[0005] In summary, existing technologies primarily focus on the analysis of single, static pathological sections, lacking the ability to capture dynamic changes before and after treatment, and failing to fully quantify the spatial heterogeneity within tumors. Furthermore, they suffer from bottlenecks in multimodal data fusion and cross-center generalization. This limits the accuracy and clinical applicability of AI models in predicting the efficacy of treatments with significant dynamic changes, such as immunotherapy.

[0006] The above background information is provided only to aid in understanding the concept and technical solution of this invention. It does not necessarily belong to the prior art of this patent application. In the absence of clear evidence that the above information was disclosed on the filing date of this patent application, the above background information should not be used to evaluate the novelty and inventiveness of this application. Summary of the Invention

[0007] The purpose of this invention is to provide an immunohistochemical efficacy prediction system based on dynamic analysis. This addresses the following problems in existing technologies: 1) inability to effectively capture and analyze the dynamic changes in biomarker expression before and after treatment; 2) insufficient quantification of the spatial heterogeneity of biomarker distribution within the tumor microenvironment; 3) difficulty in effectively integrating multimodal clinical data such as pathological images and genomics; and 4) model susceptibility to technological variations, leading to decreased generalization ability during cross-center validation. To achieve the above objectives, the embodiments of this application disclose the following technical solutions:

[0008] In a first aspect, embodiments of this application provide an immunohistochemical efficacy prediction system based on dynamic analysis, comprising:

[0009] The image processing module is used to acquire and preprocess paired immunohistochemical digital pathology slides of the same patient before and after treatment.

[0010] The dynamic registration and analysis module is connected to the image processing module and is used to perform high-precision image registration on the paired immunohistochemical digital pathological slides and extract dynamic quantitative parameters of biomarker expression changes with treatment.

[0011] A spatial heterogeneity quantification module, connected to the image processing module, is used to divide tumor functional sub-regions within a single slice and quantify the spatial distribution heterogeneity parameters of biomarkers in different functional sub-regions.

[0012] The fusion prediction module, connected to the dynamic registration and analysis module and the spatial heterogeneity quantification module, is used to fuse the dynamic quantification parameters and the spatial distribution heterogeneity parameters, and output clinical efficacy prediction results based on the fused features through a survival analysis model.

[0013] Optionally, the high-precision image registration performed by the dynamic registration and analysis module includes, in turn, global rigid registration based on feature points and local non-rigid registration based on a B-spline free deformation model; the dynamic quantification parameters include at least a change index calculated based on the density of positive cells in the same anatomical region before and after treatment.

[0014] Optionally, the dynamic quantification parameter is the immune cell infiltration dynamic index (DII), which is calculated using the following formula:

[0015]

[0016] in, and The positive cell densities were calculated in the tumor invasion border area before and after treatment, respectively.

[0017] Optionally, the tumor functional sub-regions divided by the spatial heterogeneity quantification module include at least the tumor core region and the tumor invasion edge region; the spatial distribution heterogeneity parameters include at least the edge-core polarization index ECPI.

[0018] Optionally, the formula for calculating the edge-core polarization index (ECPI) is as follows:

[0019]

[0020] or

[0021]

[0022] in, and These represent the positive cell densities in the tumor invasion margin region and the tumor core region, respectively, with ε being a minimal constant to prevent division by zero; the tumor invasion margin region is the area extending inward and outward from the tumor-stromal boundary, defined by a predetermined distance based on the histological contour.

[0023] Optionally, the survival analysis model used by the fusion prediction module is a random survival forest model or a Cox proportional hazards model, which is used to output the patient's risk score or the probability of progression-free survival at a specific time point.

[0024] Optionally, the image processing module further includes a staining normalization unit for performing color deconvolution or deep learning-driven color normalization on immunohistochemical slices to eliminate differences between different staining batches.

[0025] Optionally, the system also includes an interpretability report generation module for generating a structured report that integrates the efficacy prediction results, key quantitative parameters, and their visual mapping on pathological sections.

[0026] Optionally, the system supports biomarkers including PD-L1, Ki-67, or HER2.

[0027] Secondly, embodiments of this application provide a computer-readable storage medium having a computer program stored thereon, wherein when the computer program is executed by a processor, it implements the functions of any of the systems described above.

[0028] The beneficial effects of this application are as follows:

[0029] (1) By integrating dynamic change information and spatial heterogeneity features, the system can more comprehensively characterize tumor biological behavior and significantly improve prediction accuracy.

[0030] (2) The dynamic parameters extracted by the system can indicate the treatment response earlier, which helps clinicians to identify potential ineffective treatments and adjust the plan in the early stage, thus realizing early warning.

[0031] (3) Through the built-in staining standardization algorithm and domain adaptive training strategy, the system is more adaptable to different scanner models and staining protocols, and its generalization ability has been optimized.

[0032] (4) The system can automatically generate structured reports that integrate quantitative parameters and visualization results, which greatly improves the efficiency of decision support. Attached Figure Description

[0033] To more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings in the following description are merely exemplary, and those skilled in the art can derive other embodiments based on the provided drawings without creative effort.

[0034] Figure 1 A block diagram of an immunohistochemical efficacy prediction system based on dynamic analysis is provided as an embodiment of this application;

[0035] Figure 2 A schematic diagram of an image dynamic registration process provided in an embodiment of this application;

[0036] Figure 3 A schematic diagram illustrating the extraction of spatial heterogeneity quantification features according to an embodiment of this application;

[0037] Figure 4 This is a schematic diagram of the structure of a multimodal feature dynamic attention fusion mechanism provided in an embodiment of this application;

[0038] Figure 5 This is an example diagram of a structured report generated by a system that is interpretable, provided as an embodiment of this application. Detailed Implementation

[0039] Specific embodiments of the invention will now be described in detail. Although the invention is described in conjunction with these specific embodiments, it should be understood that it is not intended to limit the invention to these specific embodiments. Rather, these embodiments are intended to cover alternative, modified, or equivalent embodiments that may be included within the spirit and scope of the invention as defined by the claims. In the following description, numerous specific details are set forth in order to provide a thorough understanding of the invention. The invention may be practiced without some or all of these specific details.

[0040] When used in conjunction with the terms "comprising," "method comprising," or similar language in this specification and appended claims, the singular forms "a," "some," and "the" include plural references unless the context clearly indicates otherwise. Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0041] The following is in conjunction with the appendix Figure 1-5 The preferred embodiments of the present invention will be described in detail so that those skilled in the art can implement the present invention accordingly.

[0042] like Figure 1 As shown, Figure 1 This is a block diagram of an immunohistochemical efficacy prediction system based on dynamic analysis, provided as an embodiment of the present invention. The system includes: an image processing module, a dynamic registration and analysis module, a spatial heterogeneity quantification module, a fusion prediction module, and an interpretability report generation module.

[0043] Specifically, the overall system solution follows the technical path of "dynamic registration → spatiotemporal feature extraction → multimodal fusion → survival prediction", aiming to transform clinical pathological images into quantifiable predictive indicators.

[0044] The following will describe in detail the various modules of the immunohistochemical efficacy prediction system based on dynamic analysis in the embodiments of this application.

[0045] The image processing module is used to acquire and preprocess paired immunohistochemical digital pathology slides of the same patient before and after treatment.

[0046] Furthermore, the image processing module also includes a staining normalization unit, which performs color deconvolution or deep learning-driven color normalization on immunohistochemical slices to eliminate differences between different staining batches.

[0047] Specifically, the image processing module is the data entry point for this system. Its task is to process raw digital pathology images acquired from the clinical environment, which contain various technical variations (such as staining differences and slide damage), into high-quality, standardized image data, providing reliable input for subsequent dynamic registration and quantitative analysis. The implementation of this module follows the process of "acquisition-quality control-standardization-annotation".

[0048] The first step of this module is to acquire multi-source data and standardize its format. The image processing module acquires data from various sources through standardized medical data interfaces (such as the DICOM WG-26 Pathology Supplemental Protocol compliant with DICOM standards, or the HL7 FHIR interface commonly used in hospital information systems). These sources include: digital slide scanner workstations in hospital pathology departments, regional pathology image data centers, or declassified clinical research databases.

[0049] It should be noted that the system receives not only image files, but also structured metadata, such as: patient anonymization ID, slide unique identifier, sampling site, staining markers (e.g., PD-L1, clone number 22C3), treatment stage (baseline / week X after treatment), scanner model (e.g., Leica Aperio GT 450, Hamamatsu NanoZoomer S360), and objective magnification (usually 40x objective with a resolution of approximately 0.25 micrometers / pixel).

[0050] Because different scanners generate files in various formats (e.g., .svs, .ndpi, .mrxs, .tif), the module incorporates a universal WSI decoder library (based on the OpenSlide or libvips open-source library) to uniformly decode images of all formats into multi-resolution pyramid-structured memory objects and extract key metadata. It outputs a unified internal image object and its associated metadata table.

[0051] The second step involves automatically identifying and removing substandard images to ensure the input quality of the analysis process. This module integrates the open-source tool HistoQC as its core quality control engine, executing a series of parallel image quality analysis pipelines.

[0052] In one embodiment of this step, tissue region detection and area screening are performed on the received image. Specifically, the Otsu thresholding method or a color clustering-based algorithm is applied to the lowest resolution layer of the WSI (e.g., a 1.25x objective lens) to distinguish tissue regions from the background (white slide). Then, the proportion of tissue regions is calculated. If the proportion is lower than a set threshold (e.g., <30%), it indicates that the tissue is too small or the slice is invalid, and the image will be marked as "unacceptable." The formula for calculating the proportion of tissue regions is:

[0053]

[0054] in, For organizational coverage ratio, The total number of pixels in the target organization. This represents the total number of pixels in the entire analysis area.

[0055] In another embodiment of this step, the received image is evaluated for focus sharpness. This involves extracting high-resolution image patches (e.g., 512x512 pixels under a 20x objective lens) from multiple sampling regions. The local sharpness of each image patch is calculated, typically using the variance of Laplacian as a metric.

[0056] in, This is a grayscale image of an image patch. To score the focus quality of an image patch, Areas below the empirical threshold are considered out of focus.

[0057] If the out-of-focus area of ​​the entire slice exceeds a certain proportion (e.g., >15%), it will be marked as "image blurred".

[0058] In another embodiment of this step, tissue folding, tearing, and bubble detection are performed on the received image. Since tissue folds and bubbles appear as abnormal bright white bands or circular areas in the image, continuous regions exceeding the normal tissue brightness range can be detected by analyzing the distribution of the brightness channels in the H&E color space (or color deconvolution for IHC images). Morphological operations and connected component analysis are then used to quantify the area and number of these artifacts.

[0059] In another embodiment of this step, the received image undergoes a staining integrity check. For IHC slices, the image is deconvolved from the RGB color space to the hematoxylin (H) and DAB (D) staining channels. The pixel intensity histogram of the DAB channel is analyzed. If the proportion of high-intensity (positive signal) pixels is extremely low (e.g., <0.1%), it may indicate staining failure or antibody inactivation. This step finally generates a QC report, including the pass / fail status of each slice, specific quality indicators (tissue percentage, average sharpness, artifact area, etc.), and a visual annotation map (highlighting out-of-focus or folded areas). Only slices that "pass" proceed to the next stage.

[0060] The third step, to overcome staining differences between different centers and batches and improve the model's generalization ability, provides two parallel standardization methods, with the classic physics-based method used by default.

[0061] The first method is the default normalization method, which is color deconvolution-based normalization. This method is based on the Lambert-Beer law, assuming that at a specific wavelength, optical density (OD) is linearly related to dye concentration. The RGB image is converted to optical density space and deconvolved to the basis vector of the specific dye. The steps of this default normalization method are as follows:

[0062] First, convert the RGB image to optical density. The conversion formula is:

[0063]

[0064] in, For optical density, For RGB pixel intensity, The background light intensity of the glass slide (usually taken as the value of the brightest pixel in the image).

[0065] Furthermore, using predefined staining matrix vectors (For example, for IHC:) , Solve the staining concentration map Solve the system of linear equations. The least squares solution is obtained by inverting C. The formula for calculating the pseudo-inverse is:

[0066]

[0067] in, This is a staining concentration graph. For staining matrix, Optical density.

[0068] Furthermore, from the density map of the image to be processed Concentration map of target reference image In the process, the hematoxylin channel concentrations are extracted separately. The hematoxylin channel concentrations of the image to be processed are then replaced with the corresponding values ​​of the target reference image, while the DAB channel concentrations (i.e., biomarker information) are retained unchanged.

[0069] Furthermore, using a new concentration map and the same staining matrix vector... Reconstruct the RGB image:

[0070]

[0071] in, For the standardized RGB image, The incident light intensity is a reference value (1.0 in this embodiment). For staining matrix, This is a standardized staining concentration diagram.

[0072] Understandably, after standardization, the color of the cell nuclei (hematoxylin staining) in all sections became highly consistent, while the brown DAB positive signal was faithfully preserved.

[0073] This application also provides a second standardized method, namely a color transfer method based on deep learning.

[0074] This method is used for cases with exceptionally complex color distributions, or in the few cases where the above methods are ineffective. The module integrates a lightweight recurrent generative adversarial network (such as a variant of CycleGAN). This network is pre-trained on a large number of unpaired pathological images, learning to map images with arbitrary staining styles to standard staining styles while maintaining absolute invariance in the histological morphology.

[0075] The fourth step aims to provide spatial priors for subsequent dynamic registration and weakly supervised learning, significantly reducing the reliance on fine-grained annotation. This module embeds a simplified, web-based annotation interface or provides an integration script with the open-source software QuPath. A pathologist (or a trained annotator) is invited to annotate a baseline (pre-treatment) standardized IHC slide. Using polygon or pen tools, a rough outline is drawn along the outer contour of the tumor cell-rich region, forming one or more closed polygonal regions. This step does not require precise delineation of individual tumor cells or differentiation between the tumor core and stroma. It only needs to ensure that the delineation covers the majority of the tumor area, allowing for a small amount of adjacent non-tumor tissue. Finally, a binary mask file (such as JSON or GeoJSON format) aligned with the image pixel coordinates is generated, where the tumor region pixel value is 1 and the background is 0.

[0076] After the above four steps, the image processing module provides downstream modules with a fully standardized and analyzable "data package". This data package includes at least: a pair of color-normalized full-view digital pathology images (WSI) (before and after treatment), a coarse binary mask of the tumor region corresponding to the baseline image, and a structured data file containing all acquisition parameters and quality control indicators.

[0077] Understandably, this module, through a fully automated pipeline, ensures high quality and comparability of input data, laying an indispensable foundation for subsequent accurate dynamic registration, reliable quantitative feature extraction, and even the predictive performance and generalization ability of the entire system.

[0078] Next, we will introduce the dynamic registration and analysis module. This module, connected to the image processing module, is used for high-precision image registration of paired slices and to extract dynamic quantitative parameters showing changes in biomarker expression with treatment.

[0079] The high-precision image registration includes, in turn, global rigid registration based on feature points and local non-rigid registration based on a B-spline free deformation model. Dynamic quantization parameters include at least an index calculating the change in positive cell density in the same anatomical region before and after treatment.

[0080] Furthermore, in one feasible implementation, the dynamic quantification parameter is the immune cell infiltration dynamic index DII, which is calculated using the following formula:

[0081]

[0082] in, and The positive cell densities were calculated in the tumor invasion border area before and after treatment, respectively.

[0083] Specifically, the dynamic registration and analysis module is a crucial step in the system's "dynamic analysis" process. It is responsible for precisely aligning the standardized paired slices output by the image processing module in space and, based on the aligned image pairs, quantifying the changes in biomarker expression over time. Its implementation process strictly follows the logic of "global rigid initial registration → local non-rigid fine-tuning → dynamic parameter extraction".

[0084] The goal of high-precision image registration is to find a spatial transformation that allows every point on the post-treatment slice (floating image F) to find its accurate anatomical counterpart on the pre-treatment slice (reference image R). A two-stage progressive strategy is employed to achieve cell-level alignment.

[0085] The first stage is a global rigid registration based on feature points. First, it is performed on a lower-resolution pyramid layer (e.g., corresponding to a 5x objective lens) for the reference image R and the floating image F. Next, a Gaussian difference pyramid is constructed to search for local extrema at continuous scales and spatial locations; these points are scale- and rotation-invariant. Then, keypoints are precisely located, and a principal orientation is assigned to each point based on its local image gradient direction. A 16x16 region centered on the keypoint is divided into 4x4 sub-blocks, and the gradient histograms for eight directions within each sub-block are calculated, ultimately forming a 128-dimensional (4*4*8) feature vector, i.e., the feature descriptor. This descriptor is robust to changes in illumination and slight deformations.

[0086] Furthermore, using the k-nearest neighbor algorithm, for each descriptor in the floating image F, two nearest neighbor descriptors are found in the reference image R. Lowe's ratio test is then applied for filtering; a match is accepted only if the ratio of the nearest neighbor distance to the second nearest neighbor distance is less than a threshold T (T=0.75 in this system). This effectively eliminates approximately 90% of false matches. Let the set of filtered matching point pairs be... ,in and These are the two-dimensional coordinates on R and F, respectively.

[0087] Furthermore, the RANSAC algorithm is employed for transformation model estimation and global alignment. The global transformation is assumed to be an affine transformation model, with its matrix form H containing translation, rotation, scaling, and shearing. This transformation maps a point q in F to the coordinate system R: The RANSAC algorithm process is as follows:

[0088] a. Random sampling: Randomly select 3 pairs from the matched point pairs (to solve the minimum sample set for affine transformation).

[0089] b. Model Calculation: Calculate a candidate affine transformation matrix using these 3 pairs of points. .

[0090] c. Interior point statistics: using Transform all matching points in F Calculate its corresponding point in R. Euclidean distance .like (This system is designed) If the pixel (at 5x objective scale) is used, then the point is marked as an "interior point".

[0091] d. Iteration and Selection: Repeat step ac N times (N is determined by an adaptive algorithm to ensure that a sample consisting entirely of interior points is drawn at least once with a high probability). Select the sample with the most interior points. As the best model .

[0092] Use all For matching points determined to be interior points, a more accurate global affine transformation matrix is ​​recalculated using the least squares method. Using bilinear interpolation, the floating image F is... Resampling is performed to obtain a globally aligned image. At this point, the overall anatomical structure of the two images is basically aligned, and the Dice coefficient of the tumor region can usually reach 0.70-0.78.

[0093] The second level is a local non-rigid fine registration based on a B-spline free deformation model. Above, a uniform two-dimensional control point grid is applied. The initial spacing between control points is (For example, 256 pixels under a 20x objective lens). Each control point Having a two-dimensional displacement vector Any point in the image The displacement is obtained by weighted interpolation of the displacements of 16 control points within its 4x4 neighborhood using cubic B-spline basis functions:

[0094] in, , , , . arrive The basis functions are standard cubic B-spline. This model guarantees the continuity and smoothness of the deformation field.

[0095] Furthermore, find a set of control point displacements. This makes the deformed image It is most similar to the reference image R, and the deformation itself is not too drastic. Loss function Defined as:

[0096] in, For similarity measurement, To smooth the regularization term, It is a hyperparameter that balances similarity and smoothness, and is usually determined on the validation set through grid search.

[0097] Furthermore, gradient descent or its variants (such as momentum-based SGD) are employed for optimization. In each iteration, the loss function is calculated. Displacement relative to each control point The gradient is calculated, and the displacement is updated in the opposite direction of the gradient. To prevent getting trapped in local optima, a multi-resolution strategy is often adopted, that is, first, the control points are spaced relatively far apart (...). The initial deformation field is obtained by optimizing on a relatively coarse mesh (larger); then the mesh is progressively refined to optimize deformation details at a finer scale. When the loss function... The optimization stops when the decrease in consecutive iterations is less than a threshold, or when the maximum number of iterations is reached. This yields the optimal local non-rigid deformation field. Combine it with global transformation Combining the results yields the final transformation: .application The floating image is resampled to output the final registered image that is precisely aligned with the reference image. .

[0098] Furthermore, in obtaining precisely registered image pairs Then, dynamic change parameters can be calculated within the defined anatomical region. The parameters are automatically generated using the output of the spatial heterogeneity quantification module (see other sections), or based on a coarse tumor mask from the reference image R within this module, through morphological operations (such as erosion followed by dilation). Specifically: the inner edge of the core region is defined by contracting 50 micrometers inward from the tumor region boundary, and the outer edge of the invasion margin region is defined by expanding 100 micrometers outward; the annular region between these two is the "tumor invasion margin region." This distance is set based on histological common sense and can be fine-tuned according to the specific cancer type. (The image shows the reference image R and the post-treatment image.) The above steps involve performing cell nucleus segmentation, positive signal interpretation, and density calculation on the invasion edge region within the same spatial coordinate range defined above, to obtain the positive cell density. Then, the DII is calculated using the formula for the immune cell infiltration dynamic index (DII). The formula is:

[0099]

[0100] in, and The positive cell densities were calculated in the tumor invasion border area before and after treatment, respectively.

[0101] Understandably, this module seamlessly integrates with the upstream image processing module and the downstream multimodal fusion module. It receives standardized image pairs and a coarse tumor mask, and through the aforementioned complex but automated calculation process, ultimately outputs: registered image pairs, key dynamic parameters, and a registration quality report. The registered image pairs are used for visual verification. The key dynamic parameters are primarily the DII (Digital Indicator Index), as well as other optional parameters such as the rate of change in the number of positive cells and the rate of change in the area of ​​positive regions. The registration quality report includes the Dice coefficients for each stage, registration time, etc., and is used for process monitoring and quality backtracking.

[0102] The spatial heterogeneity quantification module, connected to the image processing module, is used to divide tumor functional sub-regions within a single slice and quantify the spatial distribution heterogeneity parameters of biomarkers in different functional sub-regions.

[0103] Among them, the tumor functional sub-regions divided by the spatial heterogeneity quantification module include at least the tumor core region and the tumor invasion edge region; the spatial distribution heterogeneity parameters include at least the edge-core polarization index ECPI; and the biomarkers include PD-L1, Ki-67 or HER2.

[0104] Furthermore, the formula for calculating the edge-core polarization index ECPI is as follows:

[0105]

[0106] or

[0107]

[0108] in, and These represent the positive cell densities in the tumor invasion periphery and the tumor core, respectively. To prevent extremely small constants from being divided by zero; the tumor invasion margin is a region extending at a predetermined distance from the tumor-stromal boundary, as determined by the histological profile.

[0109] Specifically, this module is responsible for finely dividing functional regions with different biological significance within a single pathological slide and quantitatively calculating the distribution patterns of biomarkers within these regions. Its implementation follows a progressive logic of "functional zoning → cell identification → parameter calculation".

[0110] Precisely delineating the tumor core and invasive border regions is fundamental to all calculations in this module. In one possible embodiment, an automated algorithm based on morphological and distance transformations is used to delineate the tumor core and invasive border regions. This module uses a single, stain-normalized immunohistochemical (IHC) slice (e.g., PD-L1) from the image processing module and its corresponding coarse binary mask of the tumor region (roughly outlined by a pathologist, with a pixel value of 1 for the tumor region and 0 for the background). Morphological closing operations (dilation followed by erosion) are performed on the tumor mask to smooth the contours, fill tiny voids, and obtain a continuous and complete tumor region. .

[0111] Further calculation The contour of the image is used to obtain the initial tumor-stromal interface (TSI). The resolution of each pixel in the image is calculated. The directed Euclidean distance of the invasion edge (TSI). Based on broad histological consensus, the invasion edge region is defined as a ring-shaped area within and outside the TSI. This system sets this distance to 100 micrometers (µm). Therefore, the invasion edge region mask... It can be generated through the following logical conditions:

[0112] That is, it includes all pixels within 100µm inward and 100µm outward from the tumor boundary.

[0113] Furthermore, the tumor core region is defined as the internal area of ​​the tumor far from its boundaries. From the intact tumor mask... In the middle, subtract the mask of the invasion edge region. This allows us to obtain the mask of the tumor core area. :

[0114] To ensure the core area is connected, it is possible to... Perform morphological opening operations (erosion followed by dilation) to remove small islands that may be generated due to the subtraction of edge regions.

[0115] At this point, the system has completed the automated, repeatable, and precise division of the two key functional sub-regions. The results do not rely on subjective judgment and provide a stable spatial framework for subsequent quantitative analysis.

[0116] The identification and density calculation of biomarker-positive cells aims to quantify the expression levels of specific biomarkers (such as PD-L1, Ki-67, and HER2) within functional subregions. First, a pre-trained deep learning model, HoVer-Net, is used to perform semantic segmentation, instance segmentation, and preliminary classification on the entire IHC slice. The model outputs a set containing the coordinates of all detected cell nuclear contours. .

[0117] For DAB (brown) color markers such as PD-L1 and Ki-67, the system first uses color deconvolution (such as the H-DAB method) to separate the pure DAB coloring channels from the RGB image. In this channel image, the pixel grayscale value is directly proportional to the DAB staining intensity. Next, cell-level positivity assessment is performed. For each segmented cell nucleus... Calculate its contour The median pixel intensity (using the median is less sensitive to outliers and better than the mean). If If the median DAB intensity is greater than or equal to the threshold, the cell is marked as a positive cell.

[0118] The threshold can be set in two ways. The first is to determine an empirical threshold T by analyzing known negative / positive control tissue sections. The second is to automatically determine the optimal threshold for distinguishing between positive and negative results by performing cluster analysis (such as the Otsu algorithm) on the median DAB intensity of all cell nuclei within the entire section or tumor region. .

[0119] For any functional sub-region (such as the invasion edge region) The formula for calculating the positive cell density D is as follows:

[0120] in, To provide a mask for the area The total number of positive cells in the body This represents the actual area of ​​the organization in the region.

[0121] Based on the above method, the density of positive cells in the tumor invasion border area was calculated in parallel. (number of cells / mm²) and the density of positive cells in the tumor core area .

[0122] Furthermore, based on the above calculation data, the spatial heterogeneity parameter is calculated. The system provides two methods to calculate the edge-core polarization index ECPI, and the calculation formula is as follows:

[0123]

[0124] or

[0125]

[0126] in, and These represent the positive cell densities in the tumor invasion periphery and the tumor core, respectively. To prevent extremely small constants from being divided by zero; the tumor invasion margin is a region extending at a predetermined distance from the tumor-stromal boundary, as determined by the histological profile.

[0127] It should be noted that, The explanation is more intuitive and suitable for clinical reports; It exhibits superior properties in mathematical models and statistical analyses. The system can calculate and store both by default.

[0128] In another possible implementation, the spatial distribution heterogeneity parameter is the spatial clustering index. This parameter quantifies whether positive cells are clustered, dispersed, or randomly distributed in two-dimensional space, going beyond a simple density measure. Given the coordinates of all positive cells within a given region (such as the entire tumor or core area), this parameter... Its empirical K-function estimate is:

[0129] in, The area is the region. This represents the total number of positive cells within the region. Let be the distance between cell i and cell j; This is an indicator function; it returns 1 when the condition is true and 0 otherwise. The radius of the region; This is the boundary correction weight, used to correct for counting bias caused by the boundaries of the study area.

[0130] The calculated Theoretical values ​​under completely spatial stochastic processes Comparison is typically performed using its transformed form. .like >0 indicates that the scale The space above is clustered; if <0 indicates spatial regularity or dispersion; if ≈0 indicates a random distribution. The system can output values ​​at different spatial scales. (e.g., 10µm, 20µm, ..., 100µm) Values ​​are used as feature vectors, or extracted The maximum value is used as the global clustering index.

[0131] Understandably, this module ultimately outputs spatial partitioning results and quantification parameters. Through this module, the system successfully transforms the abstract pathological concept of "tumor spatial heterogeneity" into a series of objective numerical indicators that can be accurately calculated and repeatedly compared, providing crucial spatial dimensional features for building models with high predictive performance.

[0132] The fusion prediction module, connected to the dynamic registration and analysis module and the spatial heterogeneity quantification module, is used to fuse dynamic quantification parameters and spatial distribution heterogeneity parameters, and output clinical efficacy prediction results based on the fused features through a survival analysis model.

[0133] The survival analysis model is either a random survival forest model or a Cox proportional hazards model, used to output the patient's risk score or the probability of progression-free survival at a specific time point.

[0134] Specifically, the fusion prediction module is the intelligent decision-making hub of this system. Its core task is to intelligently fuse the multi-dimensional and multi-scale features (dynamic, spatial, morphological, and clinical) extracted by the preceding modules; and based on the fused features, to directly output quantitative prediction results that are instructive for clinical decision-making using a survival analysis model. Its implementation process follows a precise logic of "feature aggregation → intelligent fusion → survival modeling → result output".

[0135] Since features from different modules and sources have different dimensions, scales, and distributions, direct fusion will introduce bias. Therefore, the first step is to standardize and vectorize the features.

[0136] The input features for this module include: dynamic indices of immune cell infiltration from the dynamic registration and analysis module, including the rate of change in the area of ​​positive regions before and after treatment; and edge-core polarization indices and spatial clustering indices from the spatial heterogeneity quantification module. The data includes values ​​at different scales; slice-level deep feature vectors extracted using an attention-based multi-instance learning model; and patient age, sex, clinical stage (TNM), ECOG performance status score, and tumor mutational burden obtained from laboratory tests. All continuous features (such as DII, ECPI, TMB, and age) are Z-score standardized to a mean of 0 and a standard deviation of 1. Categorical variables (such as sex and T stage) are then one-hot encoded, resulting in an initial fused feature vector. ,in This represents the total dimension of the features.

[0137] Furthermore, a dynamic attention fusion network is employed to simulate the weighing process that clinicians use when synthesizing different pieces of evidence.

[0138] Will The features in the image are divided into K feature subgroups according to their source or type (such as "image dynamic features", "image spatial features", "deep semantic features", "clinical features"). .

[0139] Each feature subgroup Each query vector is generated by mapping through an independent, learnable linear layer. Key vector Sum value vector :

[0140]

[0141] in, , , This is a trainable weight matrix. This allows features from different sources to enter different representation spaces.

[0142] To establish dynamic relationships between feature subgroups, calculate subgroups query vector Key vectors for all subgroups (including itself) Attention weights.

[0143] Attention Score Attention is calculated by scaling the dot product:

[0144] in, yes Dimensions Used for scaling to prevent gradient vanishing. Operation ensures for each ,all The sum of their weights is 1. This reflects the construction of the current feature subgroup. When considering contextual information, feature subgroups The information should be given relative importance.

[0145] For each subgroup Its context-aware features It is the weighted sum of the value vectors of all subgroups:

[0146] in, No longer isolated It is not a representation that integrates global information; This is the value vector corresponding to the feature of the j-th subgroup.

[0147] A gating mechanism is introduced to determine the extent to which the final output depends on this new contextual feature. rather than original features Gating value Determined by both original features and contextual features:

[0148] in, This represents vector concatenation. It is the Sigmoid function. and These are trainable parameters. .

[0149] subgroup The final output feature is a weighted sum of the original feature and the context feature:

[0150] Finally, the final output of all subgroups The features are concatenated again and passed through a fully connected layer to generate the final unified fused feature vector. The dimensions can be set according to the design (e.g., 256 dimensions) for dimensionality reduction and abstraction.

[0151] This module offers two industry-standard survival analysis model options, both of which can handle right-censored data commonly found in clinical studies.

[0152] In one possible embodiment, the survival analysis model is a random survival forest. The algorithmic logic of this model involves building multiple decision trees, each growing based on bootstrap samples and a random subset of features, improving predictive stability and accuracy through ensemble learning. It can predict and output a risk score and a survival probability. The risk score is typically defined as the cumulative risk value at a fixed time point (e.g., 24 months), or the comprehensive risk obtained by integrating over all time points; the survival probability... ,in It is the integrated cumulative risk function. You can directly query the progression-free survival probability at a specific time point (such as 12 months or 18 months).

[0153] In another possible embodiment, the survival analysis model is a deep Cox proportional hazards model. The model structure involves fusing feature vectors. The input is a multi-layer fully connected neural network, with no activation function in the last layer, and the output is a scalar called the prognostic index. This model can directly predict the output risk score and survival probability, where the model directly outputs... This is a continuous risk score; the survival probability requires additional estimation of the baseline survival function. After training, the Breslow estimator can be used on the training set to obtain... For a new sample, its survival probability is: .

[0154] It should be noted that the above model needs to be trained, validated, and ultimately output a structured report. The structured report includes continuous risk scores, risk stratification, survival probabilities at specific time points, as well as interpretability auxiliary information and performance metrics.

[0155] Understandably, this module completes the "last mile" from raw data to clinical insights. It not only outputs a predicted figure, but also provides doctors with a transparent, multi-dimensional evidence-based decision support tool through its internal interpretable mechanisms and structured output.

[0156] The interpretability report generation module is used to generate structured reports that integrate efficacy prediction results, key quantitative parameters, and their visual mapping on pathological slides.

[0157] In one possible embodiment, the module automatically aggregates structured data from upstream modules, including patient clinical information, risk scores and survival probabilities output by the fusion prediction module, pre- and post-treatment comparison maps from the dynamic registration module, parameters such as ECPI and DII calculated by the spatial heterogeneity quantification module, and the attention weight matrix generated by the AB-MIL network. Subsequently, the module utilizes a built-in medical assertion rule engine to automatically convert the quantified parameters into clinical findings described in natural language (e.g., "ECPI > 0.3 suggests a marginal enrichment pattern of PD-L1 expression"). In the visualization rendering stage, the module precisely overlays the AB-MIL attention weights onto the original pathological slide in the form of a semi-transparent heatmap using interpolation and alpha mixing techniques, intuitively highlighting the key areas upon which the AI's judgment is based; simultaneously, it integrates cell nucleus segmentation contours, functional subregion division boundaries, and comparison views of pre- and post-treatment registration areas. Finally, all text assertions, visualizations, parameter dashboards, and feature importance rankings provided by the model are automatically formatted by an HTML / JavaScript-based template engine and synthesized into an interactive HTML report that can be viewed in a browser (or exported as a PDF). This report supports image zooming, heatmap switching, and data hover tips, and is directly pushed to the doctor's workstation through the hospital information system interface. This transforms the complex AI analysis process into a data-driven, evidence-complete, and directly applicable structured narrative document for clinical decision-making.

[0158] It is important to note that the generated reports are stored on the server as encrypted files and are accessible to doctors via a unique link or directly embedded in the hospital's pathology information system / electronic medical record system interface. The system can be configured with automatic alerts; when a "high-risk" report is generated, a notification will be sent to the attending physician via the hospital's internal messaging system.

[0159] Understandably, this module does not produce a simple list of numbers, but rather a data-driven, visually intuitive, and clearly defined AI-assisted diagnostic narrative. It "translates" complex algorithmic outputs into clinical language and uses visual evidence to construct a complete and traceable logical chain from original pathological slides to the final predicted conclusion. This greatly enhances doctors' trust in AI decision-making and ultimately improves the adoption rate and practicality of AI tools in real-world clinical scenarios.

[0160] This application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the functions of the above-described system.

[0161] The following will illustrate a dynamic analysis-based immunohistochemical efficacy prediction system according to specific embodiments of this application.

[0162] This embodiment uses the prediction of the efficacy of PD-1 / PD-L1 inhibitor immunotherapy in patients with advanced non-small cell lung cancer (NSCLC) as an example, but the present invention is not limited to this and is also applicable to the efficacy analysis of other cancer types and other immunohistochemical markers (such as Ki-67, HER2).

[0163] This embodiment is deployed on a computing server equipped with a high-performance GPU (such as NVIDIA A100), and the software environment is based on Python 3.8 and the PyTorch 1.12 deep learning framework.

[0164] The first step is data preparation and image preprocessing.

[0165] First, we collected full-field digital images (WSI) of paired PD-L1 (22C3 antibody) immunohistochemical sections from the same patient before (baseline) and at 8–12 weeks after the start of treatment (first efficacy assessment period), subject to ethical review and approval. Simultaneously, we collected the patient's baseline H&E-stained WSI, tumor mutational burden (TMB) value obtained from blood or tissue sequencing, age, clinical stage (e.g., TNM stage), and performance status score (ECOG). The efficacy endpoint was defined as objective response (OR) or disease progression (PD) as assessed according to RECIST v1.1, and progression-free survival (PFS).

[0166] Furthermore, open-source toolkits (such as HistoQC) were used to perform quality screening on WSIs, excluding sections with severe folding, blurring, or failed staining. Subsequently, the Macenko color normalization method was used to map all IHC sections to the same standard staining space. This step effectively mitigated color deviations caused by differences in staining procedures across different hospital pathology departments and is a key preprocessing step for improving the model's generalization ability.

[0167] Furthermore, a pathologist with intermediate or higher qualifications uses digital pathology image viewing software (such as QuPath) to roughly outline the tumor region on the baseline IHC-WSI. This annotation only needs to define the general area, without needing to be precise down to individual cells, greatly reducing the annotation burden and meeting the data requirements of weakly supervised learning.

[0168] The second step involves high-precision dynamic registration of the pre- and post-treatment slides. This step is a crucial technical step in achieving "dynamic analysis," with the goal of ensuring that every anatomical structure on the post-treatment slide has a precise corresponding counterpart on the baseline slide.

[0169] First, perform global rigid registration. Downsample the baseline slice (as reference image R) and the post-treatment slice (as floating image F) to a lower resolution (e.g., 2x objective resolution).

[0170] The Scale Invariant Feature Transform (SIFT) algorithm was used to detect keypoints on two images and compute their 128-dimensional descriptors.

[0171] A k-nearest neighbor (k-NN) matcher was used for initial matching, and Lowe's ratio test (with a threshold of 0.8) was applied to filter out reliable matching pairs.

[0172] Finally, the Random Sample Consensus (RANSAC) algorithm is used to iteratively estimate the optimal affine transformation matrix M from reliable matching point pairs. This matrix contains translation, rotation, and scaling parameters, which can correct for overall positional differences between images. Applying matrix M to the floating image F completes the initial alignment, at which point the Dice similarity coefficient (calculated based on a binarized tissue region mask) between the two images typically reaches above 0.70.

[0173] Furthermore, local non-rigid fine registration is performed. That is, based on the global registration, the tumor region marked by the pathologist is analyzed in detail. The region is meshed, and non-rigid deformation correction is performed on each local image block (e.g., 512x512 pixels).

[0174] A B-spline free deformation (FFD) model is used to model local deformation. This model defines a smooth deformation field through the displacement of control points on a uniform grid.

[0175] The optimization objective is defined as maximizing the normalized mutual information (NMI) of local regions between the floating image and the reference image, while incorporating the bending energy of the deformation field as a regularization term to prevent excessive deformation. Loss function. Defined as:

[0176]

[0177] in, A larger value indicates that the images are more similar, therefore minimizing... Equivalent to maximizing ; This represents the deformation field controlled by the B-spline; For bending energy; It is the regularization coefficient (set to 0.01 in this embodiment); These are the displacement parameters for the control points.

[0178] The control point positions are iteratively optimized using gradient descent or the L-BFGS optimizer until the loss function converges. After this fine registration step, the Dice coefficient of the slices in the tumor region before and after treatment can be improved to above 0.85, achieving subcellular spatial alignment and meeting the accuracy requirements for subsequent dynamic feature extraction.

[0179] The third step is to extract and quantify spatiotemporal features. This step extracts spatial heterogeneity features and dynamic change features in parallel.

[0180] The spatial heterogeneity feature extraction process based on weakly supervised learning includes: segmenting the tumor region of the registered baseline IHC-WSI into a large number of non-overlapping small image patches (e.g., 224x224 pixels). Each image patch is used to extract features using a ResNet-50 model pre-trained on ImageNet, resulting in a 1024-dimensional feature vector. The features of all image patches in a WSI constitute a feature set. This feature set is fed into an attention-based Multiple Instance Learning (AB-MIL) network. This network learns the attention weights for each image patch through an attention branch. :

[0181] in, and These are learnable parameters. Weights This reflects the importance of this image patch in predicting the therapeutic label (responsive / non-responsive) for the entire slice.

[0182] Finally, slice-level feature representation It is obtained by weighted summation of features from all image patches: This process requires no pixel-level or region-level annotations; training is completed entirely at the slice-level label level.

[0183] The steps for functional sub-region division and spatial parameter quantification include:

[0184] The HoVer-Net cell nucleus instance segmentation model was used to process baseline IHC slices to identify and classify each cell nucleus (e.g., tumor cells, lymphocytes).

[0185] Combined with the attention heatmap generated by the AB-MIL network (by... Based on the mapping back to the original image position and the cell nucleus segmentation results, the tumor core region and the invasion edge region are automatically defined. Specifically, the core portion of the continuous high-attention region is defined as the tumor core region, and the ring-shaped region extending outward 100 micrometers from it is defined as the invasion edge region.

[0186] Calculate the key spatial heterogeneity parameters:

[0187] Edge-core polarization index:

[0188]

[0189] in, and The density (cells / mm²) of PD-L1 positive cells (cells with staining intensity higher than a set threshold) in the invasive border zone and the tumor core zone, respectively. The value range is [-1, 1], with positive values ​​indicating marginal enrichment, which is related to the immune cell infiltration pattern.

[0190] Spatial clustering index: Ripley's K function is used to analyze whether the distribution of positive cells in two-dimensional space exhibits clustering. The index is calculated based on a certain distance radius. The ratio of the actual cell count to the expected count of a random distribution is used; the portion exceeding the confidence interval indicates significant aggregation or dispersion.

[0191] The dynamic feature extraction process includes:

[0192] On the registered slide pairs, the density of PD-L1 positive cells in the same anatomical region (such as the entire tumor region or the divided functional subregion) before and after treatment was calculated.

[0193] On the registered slide pairs, the density of PD-L1 positive cells in the same anatomical region (such as the entire tumor region or the divided functional subregion) before and after treatment was calculated.

[0194] The fourth step is dynamic fusion of multimodal features. This involves integrating the spatiotemporal features extracted from pathological images with other modal data.

[0195] First, feature vectors are constructed. Pathological image features include depth features z extracted by AB-MIL, spatial parameters (ECPI, clustering index), and dynamic parameters (DII). H&E morphological features are extracted from baseline H&E slices using a pre-trained model, including tumor stroma ratio and nuclear polymorphism score. Genomic and clinical features include TMB (log2 transformed), age, and clinical stage (one-hot encoded).

[0196] Furthermore, all features are concatenated into an initial fusion vector. A multi-head cross-attention layer is designed. First, a query (Q), key (K), and value (V) vector is generated for each modality's feature subset through linear transformation. Attention scores between features from different modalities are calculated; for example, the Q of pathological features is multiplied by the K of genomic features, and then normalized using Softmax to obtain attention weights. These weights are then used to sum the V vectors of each modality, allowing the model to dynamically determine whether to focus more on pathological or genomic information, or a specific combination thereof, during prediction. Finally, a unified, context-rich fusion feature vector is output. .

[0197] The fifth step is to generate a predictive and interpretable report on treatment efficacy.

[0198] fuse feature vectors Input a random survival forest (RSF) model. The RSF model can effectively handle right-censored survival data and output a risk score for each patient.

[0199] Training was performed using a dataset containing 300 NSCLC patients (responder to nonresponder ratio approximately 1:1.5). The dataset was randomly split into a training set and an internal test set in a 7:3 ratio.

[0200] On the internal test set, the system achieved an AUC of 0.80 for predicting patient treatment response (OR vs. Non-OR). Patients were divided into high-risk and low-risk groups based on their median risk score, and there was a significant difference in progression-free survival (PFS) between the two groups. .

[0201] The system automatically generates a richly illustrated PDF report, which includes: a patient risk overview, a spatially heterogeneous heat map, a dynamic comparison chart, a list of quantitative parameters, and clinical decision support recommendations.

[0202] The patient risk overview prominently displays the risk score, risk level (high / low), and predicted 12-month progression-free survival probability. The spatial heterogeneity heatmap overlays the attention weights of AB-MIL onto the original pathology slides, highlighting the diagnostic areas that contribute most to prediction. The dynamic comparison map shows the cell distribution in key areas (high-attention areas) before and after treatment. The quantitative parameter list clearly lists the values ​​and clinical reference ranges of core parameters such as ECPI, DII, and TMB in tabular form. Clinical decision support recommendations are based on risk stratification, providing brief suggestions such as "continue current immunotherapy recommended" or "consider combination therapy or close follow-up recommended."

[0203] The detailed description of the above specific embodiments fully illustrates the feasibility, preferred implementation, and technical effects achieved by the technical solution of the present invention. Those skilled in the art can make several modifications and substitutions based on the above description without departing from the principles and spirit of the present invention, and these modifications and substitutions should also be considered within the scope of protection of the present invention.

[0204] The present invention and its embodiments have been described above. This description is not restrictive, and the accompanying drawings are only one embodiment of the present invention. The actual method is not limited to this. In conclusion, if those skilled in the art are inspired by this description and design similar methods and embodiments without departing from the spirit of the present invention, they should all fall within the protection scope of the present invention.

Claims

1. An immunohistochemical efficacy prediction system based on dynamic analysis, characterized in that, include: The image processing module is used to acquire and preprocess paired immunohistochemical digital pathology slides of the same patient before and after treatment. The dynamic registration and analysis module is connected to the image processing module and is used to perform high-precision image registration on the paired immunohistochemical digital pathological slides and extract dynamic quantitative parameters of biomarker expression changes with treatment. A spatial heterogeneity quantification module, connected to the image processing module, is used to divide tumor functional sub-regions within a single slice and quantify the spatial distribution heterogeneity parameters of biomarkers in different functional sub-regions. The fusion prediction module, connected to the dynamic registration and analysis module and the spatial heterogeneity quantification module, is used to fuse the dynamic quantification parameters and the spatial distribution heterogeneity parameters, and output clinical efficacy prediction results based on the fused features through a survival analysis model.

2. The system according to claim 1, characterized in that, The high-precision image registration performed by the dynamic registration and analysis module includes, in turn, global rigid registration based on feature points and local non-rigid registration based on the B-spline free deformation model; the dynamic quantification parameters include at least the change index calculated based on the positive cell density of the same anatomical region before and after treatment.

3. The system according to claim 2, characterized in that, The dynamic quantitative parameter is the immune cell infiltration dynamic index (DII), and its calculation formula is as follows: in, and The positive cell densities were calculated in the tumor invasion border area before and after treatment, respectively.

4. The system according to claim 1, characterized in that, The tumor functional subregions divided by the spatial heterogeneity quantification module include at least the tumor core region and the tumor invasion edge region; the spatial distribution heterogeneity parameters include at least the edge-core polarization index ECPI.

5. The system according to claim 4, characterized in that, The formula for calculating the edge-core polarization index (ECPI) is as follows: or in, and These represent the positive cell densities in the tumor invasion margin region and the tumor core region, respectively, with ε being a minimal constant to prevent division by zero; the tumor invasion margin region is the area extending inward and outward from the tumor-stromal boundary, defined by a predetermined distance based on the histological contour.

6. The system according to claim 1, characterized in that, The fusion prediction module uses either a random survival forest model or a Cox proportional hazards model for survival analysis, which is used to output the patient's risk score or the probability of progression-free survival at a specific time point.

7. The system according to claim 1, characterized in that, The image processing module also includes a staining normalization unit, which is used to perform color deconvolution or deep learning-driven color normalization on immunohistochemical slices to eliminate differences between different staining batches.

8. The system according to claim 1, characterized in that, The system also includes an interpretable report generation module for generating a structured report that integrates the efficacy prediction results, key quantitative parameters, and their visual mapping on pathological sections.

9. The system according to claim 1, characterized in that, The system supports biomarkers including PD-L1, Ki-67, or HER2.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the functions of the system as described in any one of claims 1 to 9.

Images