A method and system for spatial quantitative analysis of the tumor immune microenvironment in pathological images
By automatically identifying multiple immune cells and adaptively segmenting tumor functional regions in H&E-stained whole-section digital images, the high cost and reliance on manual annotation for region segmentation in existing technologies are solved, achieving efficient generation and accurate quantification of multi-cell type spatial density matrices.
Patent Information
- Application Number
- CN202611302341.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-08-26
- Publication Date
- 2026-11-10
AI Technical Summary
Existing technologies for tumor immune microenvironment analysis require additional immunohistochemical staining sections and reagents, which are costly and have long turnaround times. They cannot accurately quantify multiple immune cell subtypes, and region division relies on manual annotation or pixel discrimination, failing to generate a complete spatial density matrix of multiple cell types and multifunctional regions.
Using whole-slice digital images based on conventional H&E staining, a cell segmentation and classification model is used to automatically identify multiple immune cells. Tumor functional regions are divided using an adaptive threshold method, and a spatial density matrix of multiple cell types is generated to achieve fully automated analysis.
It reduces detection costs and turnaround time, achieves cell-level segmentation and classification of six cell types, automates region division, and outputs a complete density matrix of multiple cell types multiplied by multifunctional regions, significantly improving information dimensionality and prognostic accuracy.
Smart Images

Figure CN123072017A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the fields of computational pathology and medical image analysis, specifically to a method and system for spatial quantitative analysis of multiple cell types in the tumor immune microenvironment based on digital images of whole-slice pathological sections stained with conventional hematoxylin and eosin (H&E). More specifically, this invention relates to a computer-based method and apparatus for automatically identifying multiple immune cell subtypes, automatically dividing tumor functional regions, and generating a spatial density matrix of the intersection of multiple cell types and multifunctional regions from digital images of whole-slice sections stained with conventional H&E using deep learning technology. Background Technology
[0002] The tumor immune microenvironment (TIME) refers to the sum of various immune cells, stromal cells, cytokines, and extracellular matrix components contained in tumor tissue and its surrounding areas. Numerous studies have shown that the composition and spatial distribution of TIME are closely related to tumor occurrence, development, metastasis, and treatment response. Within TIME, different types of immune cells play distinctly different functions: T lymphocytes and B lymphocytes participate in adaptive immune responses, with T cells exerting anti-tumor effects by directly killing tumor cells, and B lymphocytes differentiating into plasma cells and secreting antibodies to participate in humoral immunity; eosinophils and neutrophils are important components of the innate immune system, playing pro-inflammatory or immunomodulatory roles in the tumor microenvironment. Importantly, the spatial distribution of immune cells in tumor tissue exhibits significant heterogeneity—immune cells located in the tumor core differ fundamentally in number, phenotype, and function from those located in the invasion front or distal interstitial regions. Therefore, accurate quantitative analysis of the spatial distribution of different immune cell subtypes in different functional regions of the tumor has significant clinical value for tumor immunophenotypic classification, prognostic assessment, and treatment planning. Hematoxylin and eosin (H&E) staining is the standard method for pathological examination. Almost all tumor biopsies and surgical specimens are stained with H&E. It is inexpensive, easy to operate, and has a wide coverage, making it the cornerstone of clinical pathological diagnosis.
[0003] Among existing technologies, immunohistochemistry (IHC) is the most mature method for quantifying the tumor immune microenvironment. EP3195191B1 (Ventana Medical Systems Inc., applied for in 2015, authorized in 2020, "Systems and Methods for Calculating Immune Scores") discloses a system and method for calculating immune scores. This system uses two immunohistochemical markers, CD3 and CD8, to stain serial sections. It automatically quantifies the density of CD3-positive T cells and CD8-positive cytotoxic T cells in the tumor core and invasion front regions through digital scanning and image analysis, and calculates the Immunoscore immune score accordingly. This immune scoring system has been used for prognostic stratification in colorectal cancer. The technical implementation of this IHC route includes: preparing serial sections and performing CD3 and CD8 immunohistochemical staining separately; digitizing the stained sections using a whole-section scanner; labeling the tumor core and invasion front regions; counting CD3 and CD8-positive cells in the labeled areas; and calculating the immune score based on density percentiles. Regarding the acquisition of tumor regions, the document states that tumor masks can be obtained from tumor marker channels after spectral demixing of multiple stained images, or from single-stained sections stained with tumor-specific staining. It also states that they can be obtained from hematoxylin-eosin stained images using a tumor segmentation algorithm, wherein the segmentation algorithm is a thresholding method based on pixel intensity or a learning-based method. After obtaining the tumor mask, the region recognition module automatically identifies the internal regions of the tumor and the regions located around the tumor.
[0004] Another technical approach is the detection of tumor-infiltrating lymphocytes (TILs) based on H&E images. US11164312B2 (Emory University et al., filed in 2018, authorized in 2021, "System and Method to Quantify Tumor-Infiltrating Lymphocytes") discloses a system and method for quantifying TILs from H&E images. This system segments whole-slide digital images (WSI) into fixed-size image blocks and uses a convolutional neural network (CNN) to binary classify each block as TIL-positive or TIL-negative, thereby generating a TIL spatial density map. Saltz et al. (Cell Reports, 2018) used this method to generate TIL mapping data on H&E slides of 13 tumor types in the TCGA database. The technical implementation of this scheme is as follows: the WSI is segmented into image blocks ranging from 50 micrometers to 100 micrometers, a trained CNN classification network is used to determine whether there is significant TIL infiltration in each block, and the classification results are mapped back to spatial coordinates to generate a TIL heatmap.
[0005] Furthermore, CN111417958B (Ventana Medical Systems Inc., filed in 2018, authorized in 2025, "Deep Learning System and Method for Joint Cell and Region Classification in Biological Images") discloses a deep learning method for joint cell type and region classification. This method uses a conventional CNN architecture (such as VGG16 or ResNet) on H&E images to simultaneously predict the cell type and tissue region type of each pixel, improving classification accuracy by eliminating biologically impossible combinations. The region labels in this method rely on manually labeled training data from pathologists.
[0006] CN115428039A (Ventana Medical Systems Inc., 2021 application, "A Method for Efficiently Determining the Density and Spatial Relationship Characteristics of Multiple Cell Types in a Tissue Region") discloses a method for determining the density of multiple cell types in a tissue image. This method takes an identified tumor region as input, defines a boundary region surrounding each region in the image, and determines whether each region belongs to a tumor based on the geometric inclusion or intersection relationship between the boundary region and the tumor region. Then, it accesses a metric characterizing biological observations in the regions belonging to the tumor and compares it with a reference value.
[0007] WO2019149817A1 and WO2020161125A1 (Ventana Medical Systems Inc. et al., filed in 2019 and 2020, respectively) disclose methods for assessing immune cell infiltration in stage III and IV colorectal cancer. Both methods are based on immunohistochemical markers, using the density of CD3-positive cells in the invasion front zone and the density of CD8-positive cells in the tumor core zone as characteristic measures, respectively, and outputting an immune context score through a scoring function.
[0008] The aforementioned existing technologies have the following shortcomings:
[0009] First, the method described in EP3195191B1 uses immunohistochemical markers as the sole source for quantifying immune cells. It measures CD3-positive T cells and CD8-positive cytotoxic T cells, requiring additional immunohistochemical slides beyond routine H&E staining. Each marker requires a dedicated primary antibody and detection reagent, increasing material and manual costs. Furthermore, each IHC slide typically detects only one or two markers; to comprehensively cover multiple immune cell subtypes, multiple consecutive slides are needed, further increasing cost and turnaround time. Its output is limited to the density of two markers in two regions, failing to constitute a density index system covering multiple cell types and functional regions. Regarding tumor region acquisition, while the document states that tumor masks can also be obtained from hematoxylin-eosin stained images using segmentation algorithms, the segmentation algorithm directly distinguishes between tumor and non-tumor cells based on image pixels, using pixels and their local features as the discrimination unit. The document does not describe a scheme that first classifies cells individually and then automatically determines the segmentation threshold based on the statistical distribution of the proportion of specific cell types within a spatial grid.
[0010] Second, the H&E-based TIL detection approach (such as the method described in US11164312B2) only performs TIL binary classification (positive or negative) at the image block level. Its spatial resolution is limited by the block size, making precise quantification at the cell level impossible. More importantly, this method uniformly classifies all immune cells as TILs, failing to distinguish between immune cell subtypes with different prognostic significance, such as lymphocytes, plasma cells, and eosinophils. Furthermore, this method does not address the functional partitioning of tumor tissue, only outputting the global TIL proportion, resulting in insufficient utilization of spatial information.
[0011] Third, while joint classification approaches (such as the method described in CN111417958B) can simultaneously classify cell types and regions, their region labels rely on manually labeled training data, requiring time-consuming manual re-labeling when extended to new tumor types or datasets. Furthermore, this method uses a traditional CNN architecture, limiting its ability to capture long-range contextual information. Additionally, this method lacks a quantitative density index system for the intersection of multiple cell types and multifunctional regions, thus failing to output a complete spatial immune spectrum.
[0012] Fourth, the region attribution method that uses known tumor areas as input (such as the method described in CN115428039A) requires that the tumor areas have been identified and provided as input before the method is executed. Its technical contribution lies in the determination of the attribution of each region after the tumor areas are known, rather than the generation of the tumor areas themselves; and its determination is based on the geometric inclusion or intersection relationship between the boundary area and the tumor area, without involving the automatic determination of the segmentation threshold based on the statistical distribution of the cell composition ratio within the region.
[0013] Fifth, the stage-specific assessment methods based on immunohistochemical markers (such as those described in WO2019149817A1 and WO2020161125A1) rely on staining with immunohistochemical markers such as CD3 and CD8. Their characteristic measurement is limited to the density of cells positive for one or two markers in a single functional region, and a density index system that crosses multiple cell types and multiple functional regions has not been established. Moreover, the above two methods are limited to stage III and stage IV colorectal cancer, respectively.
[0014] Therefore, there is an urgent need for a fully automated analysis method that can directly extract spatial distribution information of multiple cell types from conventional H&E stained sections, automatically divide tumor functional regions, and generate a complete spatial immune density matrix covering multiple cell types and multifunctional regions. Summary of the Invention
[0015] To address the aforementioned problems in existing technologies, the present invention aims to provide a method and system for spatial quantitative analysis of the tumor immune microenvironment in pathological images. This invention aims to solve the following three levels of technical problems: First, it addresses the issues of existing IHC routes requiring additional immunohistochemical staining sections and reagents, resulting in high costs, long turnaround times, and limited coverage of biomarkers in a single detection. Second, it addresses the problem that existing H&E-based deep learning methods can only perform image block-level TIL binary classification and cannot distinguish between immune cell subtypes with different prognostic significance, such as lymphocytes, plasma cells, and eosinophils. Third, it addresses the problems of existing tumor region segmentation methods either relying on manually labeled training data or directly distinguishing tumors from non-tumor cells using image pixels and their local features, thus failing to utilize cellular composition information. This invention provides a fully automated method for multi-cell type spatial immunoassay using only conventional H&E-stained whole-section digital images. It can automatically extract spatial distribution information of multiple immune cell types from H&E-stained whole-section digital images and generate a quantitative density matrix covering a multifunctional region.
[0016] According to a first aspect of the present invention, a method for spatial quantitative analysis of the tumor immune microenvironment in pathological images is provided, the method comprising the following steps:
[0017] Obtain digital images of whole sections stained with hematoxylin and eosin, and use a cell segmentation and classification model to perform cell-level segmentation and classification on the whole section digital images, outputting the spatial coordinates and cell type of each cell. The cell types include at least six categories: epithelial cells, lymphocytes, plasma cells, eosinophils, neutrophils, and connective tissue cells.
[0018] The tissue region of the whole-slice digital image is divided into a spatial grid of a preset size. The proportion of epithelial cells in each grid to the total number of cells in that grid is counted as the epithelial cell fraction. Based on the distribution of epithelial cell fractions in all cell-containing grids, an adaptive thresholding method is applied to determine the segmentation threshold, thereby distinguishing between tumor and non-tumor regions. The tumor region mask is then morphologically processed to generate a tumor mask. Based on the tumor mask, the tumor core region is defined by an erosion operation, the invasion front region is defined by an expansion operation, and the remaining tissue regions containing cells are defined as the stroma region.
[0019] The number of cells of each type and the area of each functional region are counted. The density of N cell types in M functional regions is calculated, and an N-by-M density matrix is generated. The unit of density is cells per square millimeter.
[0020] In some embodiments, the method further includes: dividing the functional region into intraepithelial and nonepithelial subregions based on whether epithelial cells are aggregated, and then statistically analyzing the density of each cell type within each subregion. This secondary division allows for the separate quantification of cell density in different histological microenvironments within the same functional region.
[0021] In some embodiments, the method further includes: using the lymphocyte density in the non-epithelial subregion of the tumor core region as a selected density index, comparing the index with a preset threshold or inputting it into a statistical model, and outputting an in vitro risk stratification identifier. The risk stratification identifier is an intermediate result for subsequent analysis.
[0022] In some embodiments, the cell segmentation and classification model is a Vision Transformer (ViT) based model, employing an encoder-decoder architecture based on a multi-head self-attention mechanism. The encoder divides the input image into multiple image blocks and extracts global contextual features through a multi-layer Transformer encoder. The decoder outputs pixel-by-pixel cell instance segmentation masks and class predictions through feature upsampling and refinement. In some embodiments, the whole-slice digital image is processed block-by-block, with each block ranging from 256 x 256 pixels to 1024 x 1024 pixels in size. Overlapping regions are set between adjacent blocks to eliminate boundary effects. In some embodiments, the size of the preset-size spatial grid is 200 to 300 micrometers. In some embodiments, the adaptive thresholding method is the Otsu adaptive thresholding method, which automatically determines the optimal threshold for distinguishing tumor and non-tumor regions by maximizing inter-class variance based on the epithelial cell fraction distribution of all cell-containing grids within the tissue region. In some embodiments, the morphological processing includes: sequentially performing a closing operation to fill gaps within the tumor region, performing an opening operation to remove isolated noise regions, and performing connected component area filtering to remove small regions with areas smaller than a preset area threshold. In some embodiments, the tumor core region is obtained by eroding the tumor mask inward by a distance of 200 to 300 micrometers; the infiltration front region is the annular region between the area where the tumor mask expands outward by 400 to 600 micrometers and the tumor core region.
[0023] According to a second aspect of the present invention, a spatial quantitative analysis system for the tumor immune microenvironment in pathological images is provided. The system includes: a cell segmentation and classification module for acquiring a digital image of a whole slide stained with hematoxylin and eosin, performing cell-level segmentation and classification on the whole slide digital image using a cell segmentation and classification model, and outputting the spatial coordinates and cell type of each cell, wherein the cell types include at least six categories: epithelial cells, lymphocytes, plasma cells, eosinophils, neutrophils, and connective tissue cells; and an automatic region division module for dividing the tissue region of the whole slide digital image into a spatial grid of a preset size, calculating the proportion of epithelial cells in each grid to the total number of cells in that grid as the epithelial cell score, determining a segmentation threshold based on the distribution of the epithelial cell score across all cell-containing grids using an adaptive threshold method, thereby distinguishing between tumor and non-tumor regions, and performing morphological processing on the mask of the tumor region to generate a tumor mask; based on the tumor mask, defining the tumor core region through an erosion operation, defining the invasion front region through an expansion operation, and defining the remaining tissue region containing cells as the stroma region; The density matrix generation module is used to count the number of cells of each type and the area of each functional region, calculate the density of N cell types in M functional regions, and generate an N-by-M density matrix, wherein the unit of density is cells per square millimeter.
[0024] In some embodiments, the automatic region segmentation module is further used to perform secondary segmentation of the functional region according to whether epithelial cells are aggregated, to obtain an intraepithelial subregion and a non-epithelial subregion; the system also includes a risk stratification module, used to output an in vitro risk stratification identifier using the lymphocyte density in the non-epithelial subregion of the tumor core area as a selected density index.
[0025] In some embodiments, the above method may be implemented by an electronic device including a processor and a memory, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the above method.
[0026] In some embodiments, the computer program described above may be stored in a computer-readable storage medium, and when executed by a processor, the computer program implements the steps of the method described above.
[0027] Compared with the prior art, the present invention has the following beneficial effects:
[0028] First, the immunohistochemical substitution effect. This invention only requires routine H&E stained sections to complete spatial immunoassay of multiple cell types, without the need for additional immunohistochemical staining, significantly reducing detection costs and turnaround time, making large-scale retrospective cohort analysis and routine clinical applications possible.
[0029] Second, the accuracy of cell classification is improved. This invention achieves cell-level segmentation and classification of six types of cells: epithelial cells, lymphocytes, plasma cells, eosinophils, neutrophils, and connective tissue cells. This is an improvement over the existing image block-level TIL binary classification to cell-level six-class classification, with significant improvements in both resolution and information content.
[0030] Third, the region division is automated. The bottom-up region division method based on epithelial cell density proposed in this invention is fully automated, requiring no immunohistochemical markers, no manual delineation by pathologists, and no reliance on prior knowledge or simple geometric rules, thus eliminating subjectivity and labeling dependence.
[0031] Fourth, the information dimension is expanded. This invention outputs a complete density matrix of multiple cell types multiplied by multiple functional regions (6 cell types multiplied by 3 functional regions equals 18 density indicators), expanding the information dimension from 2x2 (CD3 / CD8 multiplied by CT / IM equals 4 indicators) in existing IHC methods to 6x3, significantly improving the coverage.
[0032] Fifth, spatial partitioning carries irreplaceable information. In the TCGA-COADREAD cohort (524 cases), lymphocyte density in the non-epithelial subregion of the tumor core was significantly associated with overall survival (multivariate adjusted hazard ratio 0.931, 95% confidence interval 0.882–0.983, P = 0.010); while two global indicators available without partitioning—total lymphocyte density in the whole slice and the proportion of lymphocytes to the total number of cells in the whole slice—were not significantly associated with overall survival (P = 0.386 and 0.385, respectively). Further incremental information testing showed that adding the above partitioning indicators to the model based on total lymphocyte density in the whole slice significantly improved the model fit (likelihood ratio test P = 0.004); while the reverse operation, i.e., adding total lymphocyte density in the whole slice to the model that already included partitioning indicators, did not bring significant improvement (P = 0.207). The above comparison confirms that prognostic information comes from the spatial distribution of lymphocytes rather than their total amount, thus proving that the partitioning step is a necessary step to produce the technical effect.
[0033] Sixth, feasibility for clinical application. This invention is based on the analysis of routine H&E stained slides in clinical practice, without changing the existing pathology workflow. Processing a single whole slide digital image only takes 5 to 15 minutes, and it has good prospects for clinical application. Attached Figure Description
[0034] Figure 1 This is a schematic diagram of the complete processing flow of the method of the present invention, showing the processing steps from inputting a digital image of a hematoxylin-eosin stained whole slide to outputting a multi-cell type spatial density matrix, as well as an optional risk stratification step.
[0035] Figure 2 This is a schematic diagram of the bottom-up region division based on epithelial cell density in this invention, showing the complete process from spatial grid division, epithelial cell fraction calculation, adaptive threshold binarization, morphological processing to the definition of three functional regions: tumor core region, invasion front region and stromal region, as well as the process of secondary division within the tumor core region according to whether epithelial cells are aggregated to obtain the intraepithelial subregion and non-epithelial subregion.
[0036] Figure 3 This is a schematic diagram of the multi-cell type detection results of the present invention, showing the segmentation and classification results of six types of cells: epithelial cells, lymphocytes, plasma cells, eosinophils, neutrophils, and connective tissue cells in the hematoxylin-eosin staining image.
[0037] Figure 4 This is a schematic diagram of the density matrix data structure of the present invention, showing a density matrix structure with six cell types as rows and three functional regions as columns. Each element in the matrix represents the density of the corresponding cell type in the corresponding functional region, with the unit being the number of cells per square millimeter.
[0038] Figure 5 This is a system architecture diagram of the present invention, showing the various functional modules of the system and their data flow.
[0039] Figure 6 This is a timing diagram for the processing of the present invention, which illustrates the calling timing and data transfer relationships between various functional modules during the whole-slice digital image processing.
[0040] The meanings of the markings in the attached diagram are as follows:
[0041] mark meaning mark meaning 101 Digital images of whole sections stained with hematoxylin and eosin 201 Tissue regions in a full-slice digital image 102 Cell list (coordinates and type) 202 Spatial grid division 103 Functional area mask 203 Epithelial cell fraction 104 density matrix 204 Adaptive threshold determination 105 Risk stratification identification 205 Binarized tumor candidate region S1 Cell-level segmentation and classification steps 206 Morphological closing operation S2 Bottom-up region division steps 207 Morphological opening operation S3 Density matrix generation steps 208 Connected domain area filtering S4 Risk stratification steps (optional) 209 Tumor mask CT Tumor core area 210 Intraepithelial subregion of the tumor core IM Immersion Frontier Area 211 Non-epithelial subregion of the tumor core ST interstitial area 301 Hematoxylin-eosin stained tissue blocks 302 epithelial cells 305 eosinophils 303 lymphocytes 306 neutrophils 304 plasma cells 307 connective tissue cells 401 Cell list input 200 Spatial quantitative analysis system 402 Functional area mask input M1 Cell segmentation and classification module 403 Density calculation M2 Automatic region division module M3 Density matrix generation module M4 Risk stratification module (optional) Detailed Implementation
[0042] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. The following embodiments are used to illustrate the technical solutions of this invention, but do not limit the invention to the scope of the embodiments described. All technical solutions implemented based on the above content of this invention are within the protection scope of this invention. The validation data used in the following embodiments comes from the colorectal adenocarcinoma (COAD) and rectal adenocarcinoma (READ) cohorts of The Cancer Genome Atlas (TCGA) project. After deduplication and screening for complete follow-up information, digital images of hematoxylin and eosin stained whole sections of 524 colorectal cancer patients and corresponding clinical follow-up data were included. The sections came from 35 different submitting institutions. Those skilled in the art will understand that the above dataset is for illustrative purposes only, and the methods and systems described in this invention are applicable to digital images of H&E stained whole sections of any type of solid tumor.
[0043] Example 1: Complete Method Flow
[0044] like Figure 1 As shown, the spatial quantitative analysis method for the tumor immune microenvironment in pathological images provided by this invention includes three core steps, S1 to S3, and an optional step, S4. Each step is described in detail below.
[0045] Step S1: Cell-level segmentation and multi-type classification
[0046] like Figure 1 Middle step S1 and Figure 3 As shown, the purpose of this step is to automatically detect and classify all cells from the digital image of the whole slide stained with hematoxylin and eosin, and output the spatial coordinates and cell type of each cell.
[0047] S1.1 Acquisition and Specification of Full-Slice Digital Images
[0048] The input data used in this invention are H&E-stained tissue slide images digitized by a whole-slide scanner. In this embodiment, the whole-slide digital images are scanned at 40x optical magnification, corresponding to a spatial resolution of approximately 0.25 micrometers per pixel (μm / pixel). A typical whole-slide digital image is approximately 40,000 pixels by 40,000 pixels to 100,000 pixels by 100,000 pixels, corresponding to physical dimensions of approximately 10 mm by 10 mm to 25 mm by 25 mm. Whole-slide digital images are typically stored in multi-resolution pyramid formats (such as SVS, NDPI, MRXS, etc.) to support efficient multi-scale reading. In this embodiment, cell segmentation and classification are processed using image data from the highest resolution layer (40x, level 0) to ensure sufficient cell morphological detail is obtained.
[0049] S1.2 Preprocessing: Tissue region detection and background removal
[0050] Before cell segmentation, the whole-slice digital image is preprocessed to identify areas containing effective tissue and remove blank backgrounds and non-tissue areas (such as slide edges, bubbles, marker pen marks, and other artifacts), thereby reducing the computational load of subsequent processing and avoiding artifact interference.
[0051] Specifically, the preprocessing steps include: First, reading the image from a low-resolution thumbnail of the full-slice digital image (such as the 4th or 5th level pyramid layer, corresponding to approximately 1.25x or 2.5x magnification); then, converting the RGB image to a grayscale image or HSV color space representation; next, binarizing the image based on a grayscale threshold or a saturation threshold, where regions with grayscale values higher than a preset threshold (e.g., 220, grayscale range 0-255) are determined as background, and connected components with grayscale values lower than the threshold and an area greater than the minimum tissue region threshold are determined as tissue regions; finally, mapping the tissue region mask detected at low resolution back to the coordinate system of the highest resolution layer to guide subsequent block-by-block processing, i.e., only performing cell segmentation analysis on blocks containing tissue, skipping blank background blocks.
[0052] Through the above preprocessing steps, for a typical 40,000-pixel by 40,000-pixel full-slice digital image, the tissue region usually accounts for 30% to 70% of the total area. Therefore, preprocessing can effectively reduce the amount of unnecessary computation by 40% to 70%.
[0053] S1.3 Block-by-block processing strategy
[0054] Because the size of a full-slice digital image is extremely large (usually exceeding one billion pixels), it cannot be loaded into the memory of a computing device for processing all at once. Therefore, this invention adopts a patch-by-patch processing strategy, which divides the tissue region of the full-slice digital image into multiple fixed-size blocks, and each block is independently segmented and classified into cells.
[0055] In this embodiment, each block is sized to 256 pixels by 256 pixels, corresponding to a physical size of approximately 64 micrometers by 64 micrometers at 40x magnification. A 64-pixel (approximately 16 micrometers) overlap region is established between adjacent blocks to eliminate incomplete segmentation and boundary effects caused by cell truncation at block boundaries. Within the overlap region, duplicate detection results from adjacent blocks are merged: when the center-to-center distance between two detected cell instances is less than a preset merging threshold (e.g., 5 micrometers), the detection result with higher confidence is retained, and the other is discarded, thereby avoiding duplicate cell counting.
[0056] The choice of block size requires a balance between segmentation accuracy and computational efficiency. Smaller blocks (e.g., 256 pixels by 256 pixels) provide finer local cell morphology information, suitable for densely cell-rich tissue regions; larger blocks (e.g., 512 pixels by 512 pixels or 1024 pixels by 1024 pixels) provide greater spatial context, helping to distinguish morphologically similar cell types, but requiring more video memory. In this embodiment, 256 pixels by 256 pixels is chosen as the default block size to balance accuracy and efficiency. Those skilled in the art can adjust the block size within the range of 256 pixels to 1024 pixels according to the actual hardware configuration and tissue type.
[0057] S1.4 Cell Segmentation and Classification Model
[0058] In this embodiment, the cell segmentation and classification model uses the publicly available CellViT++ model, which is a cell nucleus instance segmentation and classification model based on the visual Transformer. Its model weights and inference code have been publicly released. The classification head used is trained on the Lizard dataset, a publicly labeled dataset for segmenting and classifying colorectal tissue cell nucleus instances. Its labeling system corresponds to the six cell categories described in this invention: neutrophils, epithelial cells, lymphocytes, plasma cells, eosinophils, and connective tissue cells. Those skilled in the art can obtain this model and reproduce step S1 according to the aforementioned publicly available sources. Alternatively, other segmentation and classification models capable of outputting cell spatial coordinates and labels for the aforementioned six cell types can be used. This invention does not limit the specific implementation of the model used in step S1; its output only needs to satisfy the following interface conventions: each cell has a unique identifier, spatial coordinates, a cell type label, and a classification confidence score.
[0059] This invention employs a cell segmentation model based on the Vision Transformer (ViT) to perform cell-level segmentation and classification for each image block. Compared with traditional convolutional neural networks (CNNs), the ViT model, through its multi-head self-attention mechanism, can effectively capture long-distance spatial dependencies and has a stronger modeling ability for spatially uneven and morphologically variable cell populations in pathological images.
[0060] The ViT model adopts an encoder-decoder architecture, the structure of which is described as follows:
[0061] Encoder Section: The input 256-pixel multi-256-pixel block image is first divided into multiple non-overlapping image tokens, each with a size of 16 pixels by 16 pixels. Therefore, a 256-pixel multi-256-pixel input block is divided into 256 16-16 image tokens. Each image token is first mapped to a fixed-dimensional feature vector (e.g., 768-dimensional) through a linear projection layer. The linear projection operation can be described as follows: For the pixel matrix X_p of the p-th image token (dimension 16 x 16 x 3, where 3 represents the three RGB color channels), it is flattened into a 768-dimensional (i.e., 16 x 16 x 3) one-dimensional vector, and then linearly transformed through a learnable weight matrix W_proj (dimension 768 x 768) to obtain the embedding vector z_p = W_proj × flatten(X_p) + b_proj. Subsequently, a learnable positional embedding is appended to each embedding vector to preserve the spatial location information of the image token. The positional encoding employs a learnable one-dimensional positional embedding approach, assigning a trainable 768-dimensional vector to each position index, which is jointly optimized with the model parameters during training. These feature vector sequences are fed into a multi-layer Transformer encoder, each layer containing a multi-head self-attention module and a feedforward neural network module. In the multi-head self-attention module, the input sequence is transformed through three linear transformations to generate query (Q), key (K), and value (V) matrices, respectively. The attention weights are calculated by the dot product of Q and K and then normalized using softmax: Attention(Q, K, V) = softmax(Q × K^T / sqrt(d_k)) × V, where d_k is the feature dimension of each attention head. The multi-head mechanism divides the 768-dimensional feature space into multiple parallel attention heads. Each head independently calculates its attention, and the resulting concatenation and linear transformation yield the final output. This mechanism enables the model to simultaneously focus on feature relationships at different spatial locations and semantic levels, effectively capturing long-range spatial dependencies between cells in pathological images. The feedforward neural network module contains two linear transformation layers and a GELU activation function, which performs a non-linear transformation on the output of the attention layer to enhance the expressive power of the features. In this embodiment, the encoder contains 12 Transformer layers, each using 12 attention heads, with a hidden layer dimension of 768, and the intermediate layer dimension of the feedforward network is 3072 (i.e., 4 times the hidden layer dimension).
[0062] Decoder Section: The decoder receives the feature map output from the encoder and restores the low-resolution feature map to a predicted map with the same resolution as the input image through progressive upsampling and feature refinement operations. The upsampling process includes bilinear interpolation and convolutional refinement operations, and fuses feature information from different levels in the encoder through skip connections to preserve fine-grained spatial details. The final output of the decoder consists of two parts: first, a pixel-by-pixel cell instance segmentation mask, used to identify the cell instance to which each pixel belongs; second, a pixel-by-pixel class probability map, containing six channels, corresponding to the predicted probabilities of six classes: epithelial cells, lymphocytes, plasma cells, eosinophils, neutrophils, and connective tissue cells.
[0063] S1.5 Post-processing: Instance segmentation and cell information extraction
[0064] The pixel-by-pixel segmentation mask output by the model needs to be converted into individual cell instances through post-processing steps. These post-processing steps include: First, applying a watershed algorithm or connected component analysis to the segmentation mask to separate continuous segmentation regions into individual cell instances. The watershed algorithm is particularly effective for closely packed cell populations, segmenting along energy valleys between cells. Then, calculating the morphological parameters of each cell instance, including its center coordinates (x, y), contour area, and bounding box. Next, determining its final class based on the average class probability map within the cell instance's coverage area, i.e., selecting the class with the highest average probability as the cell type. Finally, filtering detection results with areas that are too small (e.g., less than 10 square pixels, approximately 0.625 square micrometers) or too large (e.g., greater than 5000 square pixels, approximately 312.5 square micrometers) to eliminate noise and anomaly detections.
[0065] S1.6 Six-Cell Classification System
[0066] like Figure 3 As shown, this invention defines a six-class cell classification system, covering the main identifiable cell types in H&E-stained tissue sections. The morphological characteristics of each cell type are as follows:
[0067] Epithelial cells: These cells have relatively large, round or oval nuclei that stain deeply with hematoxylin. They are densely packed and typically form glandular structures or mucosal layers. In tumor tissue, epithelial cells include both neoplastic and normal epithelial cells, and their proportion within the spatial grid is a core indicator for subsequent region division.
[0068] Lymphocytes: The nucleus is small and round, approximately 6 to 8 micrometers in diameter, and deeply stained with hematoxylin. The cytoplasm is minimal and appears dark blue. Lymphocytes include T lymphocytes and B lymphocytes, and are the main effector cells of the adaptive immune response, playing a central role in anti-tumor function within the tumor immune microenvironment.
[0069] Plasma cells: The nucleus is eccentrically distributed (located on one side of the cell), and a clear halo is visible around the nucleus (perinuclear Golgi apparatus region). The chromatin is arranged in a characteristic "clock face" (wheel-like) pattern, and the cytoplasm is abundant and basophilic (blue-purple). Plasma cells differentiate from B lymphocytes and are the main effector cells for antibody secretion, and are enriched in tertiary lymphoid structures.
[0070] Eosinophils: The nucleus is usually bilobed (the two lobes are connected by thin filaments), and the cytoplasm is filled with numerous eosinophilic (orange-red in color) granules of uniform size. Eosinophil infiltration in certain tumors (such as colorectal cancer and gastric cancer) is associated with a better prognosis.
[0071] Neutrophils: The nucleus is lobed, usually with 3 to 5 lobes connected by thin filaments, and the cytoplasm contains small neutrophilic granules (appearing pale pink to colorless in H&E staining). Neutrophils are an important component of innate immunity and can play a dual role of pro-inflammatory or immunosuppressive effects in the tumor microenvironment.
[0072] Connective tissue cells: These include fibroblasts and other stromal cells. Their nuclei are elongated or spindle-shaped, and the intercellular spaces are relatively large. They are usually distributed alongside extracellular matrix structures such as collagen fibers. Connective tissue cells are the main component cells of the tumor stroma.
[0073] S1.7 Output Data Format
[0074] The output of step S1 is a structured list of cells, each entry containing the following fields: a unique cell identifier (cell_id), the x-coordinate of the cell center (in micrometers), the y-coordinate of the cell center (in micrometers), the cell type (one of six classes), the cell outline area (in square micrometers), and the classification confidence (a floating-point number between 0 and 1). For a typical whole-slice digital image, step S1 typically detects tens of thousands to hundreds of thousands of cell instances.
[0075] Step S2: Bottom-up region segmentation based on epithelial cell density
[0076] like Figure 1 Middle step S2 and Figure 2As shown, the purpose of this step is to automatically divide the tumor functional regions in the whole-slice digital image based on the cell spatial distribution information output in step S1, including three functional regions: the tumor core region (CT), the invasion front region (IM), and the stroma region (ST). The core innovation of this step lies in adopting a "bottom-up" strategy, that is, starting from the cell-level analysis results, inferring the location of the tumor region by statistically analyzing the distribution ratio of epithelial cells in the spatial grid, rather than relying on additional annotations or prior knowledge.
[0077] S2.1 Spatial Grid Generation
[0078] The tissue region of the full-slice digital image is divided into a spatial grid of equal area. In this embodiment, the grid size is set to 250 micrometers by 250 micrometers, which corresponds to approximately 1000 pixels by 1000 pixels at a magnification of 40x (0.25 micrometers per pixel). The starting coordinates of the grid are aligned with the boundary of the tissue region. For incomplete grids located at the tissue edge, the calculation is based on the actual tissue area they contain.
[0079] For each grid cell, the number of each type of cell detected within it is counted. Specifically, the cell list output in step S1 is traversed, and each cell is assigned to a corresponding spatial grid cell based on its center coordinates (x, y). Let N_epi_k be the number of epithelial cells in the k-th grid cell, and N_total_k be the total number of cells. Only when N_total_k is greater than the minimum cell count threshold (e.g., 5 cells) is the grid cell considered a valid "cell-containing grid" for subsequent calculations, to avoid statistical instability due to insufficient cell count within the grid cell.
[0080] The choice of grid size is a trade-off between accuracy and robustness: smaller grids (e.g., 100 micrometers) offer higher spatial resolution, but the number of cells within a single grid may be insufficient to provide reliable statistics; larger grids (e.g., 500 micrometers) provide more stable statistics, but spatial resolution is reduced. In this embodiment, a 250-micrometer grid size corresponds to approximately 1000 pixels by 1000 pixels at 40x magnification, typically containing tens to hundreds of cells, providing sufficient statistical reliability. Those skilled in the art can adjust the grid size according to different tissue types and cell densities, preferably from 200 to 300 micrometers; a wider range of options is described in the Alternatives section below.
[0081] S2.2 Calculation of Epithelial Cell Fraction
[0082] For each cell-containing grid, the epithelial cell fraction (ECF) is calculated and defined as:
[0083] ECF_k = N_epi_k / N_total_k
[0084] Where ECF_k is the epithelial cell fraction of the k-th grid, with a value ranging from 0 to 1. ECF_k close to 1 indicates that the grid is dominated by epithelial cells, which usually corresponds to areas with dense tumor cells in solid tumors (because most solid tumors are of epithelial origin); ECF_k close to 0 indicates that the grid is dominated by non-epithelial cells (such as lymphocytes, connective tissue cells, etc.), which usually corresponds to areas of tumor stroma or immune infiltration.
[0085] All ECF values containing cell grids were mapped to a two-dimensional heatmap based on their spatial location, namely, an epithelial cell fraction spatial heatmap. This heatmap visually reflects the spatial distribution pattern of epithelial cells in the whole-slice digital image: within the tumor region, the epithelial cell fraction typically exhibits higher values and spatial continuity; in the stromal region and immune infiltration region, the epithelial cell fraction is typically lower. This feature based on the spatial distribution of epithelial cells provides the foundational data for subsequent adaptive threshold binarization.
[0086] S2.3 Adaptive Threshold Binarization
[0087] The Otsu adaptive thresholding method was applied to the fractional space heatmap of epithelial cells to automatically determine the optimal threshold T for distinguishing tumor and non-tumor regions. The basic principle of the Otsu method is: for a distribution of grayscale (or fractional) values, find a threshold T such that when all data points are divided into two classes (foreground and background) with T as the boundary, the inter-class variance between the two classes is maximized. The formula for calculating the inter-class variance is:
[0088] σ_B^2(T) = ω_0(T) × ω_1(T) × [μ_0(T) - μ_1(T)]^2
[0089] Where ω_0(T) and ω_1(T) are the percentages of pixels (in this case, grids) on either side of the threshold T, and μ_0(T) and μ_1(T) are the mean values on both sides. The optimal threshold T* is the T value that maximizes σ_B^2.
[0090] In the application scenario of this invention, the input data is a set of epithelial cell scores {ECF_1, ECF_2, …, ECF_K} containing cell grids, where K is the total number of cell grids. The Otsu method processes this set, classifying the grids into two categories: "high epithelial cell score" (tumor candidate region) and "low epithelial cell score" (non-tumor candidate region). Grids with an epithelial cell score greater than or equal to the optimal threshold T* are marked as tumor candidate regions, and grids with an epithelial cell score less than T* are marked as non-tumor candidate regions, generating a preliminary binarized tumor candidate mask.
[0091] The advantage of choosing the Otsu method lies in its completely data-driven nature, requiring no pre-set fixed thresholds or prior distribution assumptions, and its ability to adaptively adapt to differences in epithelial cell fraction distribution between different slices. This characteristic is particularly important for processing whole-slice digital images from different institutions and different staining batches, as the staining intensity, tissue processing methods, and cell densities may vary significantly between different slices. In terms of computational efficiency, the Otsu method has a time complexity of O(L), where L is the quantization order of the fraction values (L=256 in this embodiment), which is negligible for whole-slice digital images containing thousands of grids. Furthermore, the Otsu method theoretically has optimal segmentation performance for data following a bimodal distribution. In solid tumor H&E slices, the distribution of epithelial cell fractions typically exhibits a clear bimodal characteristic—one peak corresponds to a high epithelial cell fraction in the tumor region, and the other peak corresponds to a low epithelial cell fraction in the mesenchymal region—therefore, the applicability of the Otsu method in this application scenario is supported by both theory and practice.
[0092] S2.4 Morphological treatment
[0093] The initial tumor candidate mask generated by the Otsu threshold may contain small internal voids (due to a temporary decrease in the epithelial cell fraction caused by local immune infiltration) and isolated small noisy regions (due to statistical fluctuations within a single grid). Therefore, the following morphological processing operations are performed sequentially on the initial tumor candidate mask:
[0094] The first step is morphological closing. A circular structuring element is used to perform the closing operation, the purpose of which is to fill small gaps and cavities within the tumor candidate region. The closing operation consists of an expansion operation followed by an erosion operation: first, expansion fills the gaps, then erosion restores the boundaries. In this embodiment, the radius R_close of the circular structuring element for the closing operation is set to 2 grid units (i.e., 500 micrometers), which can effectively fill internal gaps with a diameter less than 1000 micrometers while maintaining the overall outline of the tumor region. The recommended value range for R_close is 1 to 3 grid units (250 to 750 micrometers), which can be adjusted by those skilled in the art according to the typical interstitial space width in the tissue type.
[0095] The second step is morphological opening. This operation uses a circular structuring element to remove isolated small noise regions surrounding the tumor candidate mask. The opening operation consists of an erosion operation followed by a dilation operation: first, the small region is eroded away, then the boundary of the remaining region is restored by dilation. In this embodiment, the radius R_open of the circular structuring element for the opening operation is set to 1 grid unit (i.e., 250 micrometers), which can remove isolated noise regions with a diameter less than 500 micrometers. The recommended value for R_open is 1 to 2 grid units (250 to 500 micrometers).
[0096] The third step is connected component area filtering. Connected component analysis is performed on the mask after closing and opening operations, marking all independent connected components and calculating the area of each. Connected components with areas smaller than a preset minimum area threshold A_min are deleted. In this embodiment, A_min is set to 10 grid units (i.e., 10 x 250 micrometers x 250 micrometers equals 0.625 square millimeters). The recommended range for A_min is 5 to 20 grid units. The purpose of this step is to remove small, biologically insignificant "tumor" regions caused by statistical noise.
[0097] After the three morphological processing steps described above, a final binary tumor mask is generated. This mask accurately marks the spatial distribution of tumor tissue in the whole-slice digital image, with continuous boundaries and a regular shape.
[0098] S2.5 Three-Functional-Area Definition
[0099] Three functional regions were defined based on the morphologically processed tumor mask. For example... Figure 2 As shown, the definition process of the three functional areas is as follows:
[0100] Tumor Core Region (CT): An erosion operation is performed on the tumor mask, with the erosion distance D_erode set to 250 micrometers (i.e., 1 grid unit). The remaining internal area after erosion is the tumor core region. In this embodiment, the recommended value for D_erode is 200 to 300 micrometers, which can be adjusted by those skilled in the art according to tumor size and pathological characteristics. Biologically, the tumor core region represents the area where tumor cells are most densely packed and furthest from the tumor-stromal boundary. The infiltrating immune cells in this region usually represent effector cells that have successfully penetrated the tumor defenses and penetrated deep into the tumor parenchyma, which has important prognostic significance.
[0101] Infiltration Front (IM): An expansion operation is performed on the original tumor mask, with the expansion distance D_dilate set to 500 micrometers (i.e., 2 grid units). The annular region between the expanded area and the tumor core is the Infiltration Front, which covers approximately 750 micrometers on both sides of the tumor boundary (an inner 250-micrometer erosion zone plus an outer 500-micrometer expansion zone). The recommended value for D_dilate is 400 to 600 micrometers. Biologically, the Infiltration Front is the frontier zone where tumor cells directly interact with the host immune system. It is a key entry point for immune cell infiltration into the tumor and the most active area for local tumor invasion and micrometastasis.
[0102] The interstitial zone (ST) is defined as the remaining tissue region in a whole-section digital image that contains cells but is not part of the tumor core or invasion front. The ST includes normal interstitial tissue far from the tumor, tertiary lymphoid structures, and other areas of immune infiltration. Biologically, the immune cells in the ST reflect the state of the host's systemic immune response.
[0103] The definition of the above three functional regions is entirely based on the cell-level analysis results of step S1 and the calculation processing of step S2, without any manual annotation or external prior knowledge, thus realizing the full automation of tumor region division.
[0104] Step S3: Generation of spatial density matrix for multiple cell types
[0105] like Figure 1 Middle step S3 and Figure 4 As shown, the purpose of this step is to calculate the density of each cell type in each functional region based on the cell list output in step S1 and the three-functional region mask output in step S2, and generate a complete spatial immune density matrix.
[0106] S3.1 Cell statistics within the region
[0107] For each functional region j (j can be CT, IM, or ST), iterate through the cell list output in step S1, determine the functional region to which each cell belongs based on its center coordinates, and count the number of each cell type in that region. Let C_ij be the number of cells of the i-th cell type in the j-th functional region, where i ranges from {epithelial cells, lymphocytes, plasma cells, eosinophils, neutrophils, connective tissue cells}, and j ranges from {CT, IM, ST}.
[0108] S3.2 Area Calculation
[0109] Calculate the area A_j (in square millimeters, mm^2) of each functional region. The area of a functional region is obtained by multiplying the number of grids contained in the mask of that region by the area of a single grid: A_j = N_grid_j × (250 × 10^-3)^2 mm^2, where N_grid_j is the number of grids contained in the j-th functional region. For incomplete grids located at the edge of the region, a weighted calculation is performed based on the actual tissue coverage area.
[0110] S3.3 Density Calculation and Matrix Generation
[0111] Calculate the value of each element in the density matrix:
[0112] D_ij = C_ij / A_j
[0113] Where D_ij is the density of the i-th cell type in the j-th functional region, expressed in cells per square millimeter (cells / mm^2). Figure 4 As shown, the final generated density matrix is a 6-row by 3-column data structure containing 18 density indices:
[0114] CT scan (tumor core area) IM (Immersion Frontier Zone) ST (interstitial zone) epithelial cells D_epi_CT D_epi_IM D_epi_ST lymphocytes D_lym_CT D_lym_IM D_lym_ST plasma cells D_pla_CT D_pla_IM D_pla_ST eosinophils D_eos_CT D_eos_IM D_eos_ST neutrophils D_neu_CT D_neu_IM D_neu_ST connective tissue cells D_con_CT D_con_IM D_con_ST
[0115] This density matrix constitutes a complete spatial immune context, comprehensively reflecting the quantitative distribution information of multiple cell types in different functional regions within the tumor immune microenvironment. The output format is structured data (such as JSON or CSV), containing a statistical summary of each density value, corresponding cell count, region area, and classification confidence score.
[0116] Step S4: Prognostic risk assessment (optional)
[0117] like Figure 1 As shown in step S4, this step is optional. Its purpose is to output an in vitro risk stratification identifier based on the key indicators in the density matrix output in step S3, using a statistical model or a preset threshold. This identifier is an intermediate result for subsequent analysis or manual interpretation. This invention does not involve making a disease diagnosis or providing a treatment plan based directly on this identifier.
[0118] S4.1 Density Index Selection
[0119] From the 18 density indicators in the density matrix, key prognostic indicators were selected for risk assessment. In this embodiment, based on the validation results of the TCGA-COADREAD cohort, the lymphocyte density in the non-epithelial subregion of the tumor core (D_lym_CT-stroma) was selected as the key indicator; the lymphocyte density in the intraepithelial subregion within the same tumor core (D_lym_CT-epi) served as a negative control to confirm that the extracted information indeed came from spatial partitioning rather than total cell count. The selection of density indicators can be based on prior clinical knowledge, screening results from univariate survival analysis, or automatic screening from all 18 indicators using multivariate feature selection methods (such as LASSO-regularized Cox regression).
[0120] S4.2 Threshold Determination and Risk Stratification
[0121] For each selected density index, a threshold for risk stratification is determined. Threshold determination methods include: (1) median segmentation: using the median of the index in the training cohort as the threshold, patients are divided into high-density and low-density groups; (2) optimal cut-off point method (e.g., X-tile method): iterating through all possible cut-off point values, selecting the cut-off point that maximizes the survival difference between the two groups (e.g., minimizing the P-value in the log-rank test) as the optimal threshold. In this embodiment, the median segmentation method is used to determine the threshold.
[0122] S4.3 Statistical Model Construction
[0123] In this embodiment, a Cox proportional hazards regression model is used for prognostic assessment. The Cox model takes the following form:
[0124] h(t) = h_0(t) × exp(β_1 × X_1 + β_2 × X_2 + … + β_p × X_p)
[0125] Where h(t) is the hazard function at time t, h_0(t) is the baseline hazard function, X_1 to X_p are covariates (including selected density indicators and clinical variables that need to be adjusted, such as stage and age), and β_1 to β_p are regression coefficients. The model parameters are fitted using the maximum partial likelihood estimation method, and the hazard ratio (HR) and corresponding p-value for each density indicator are calculated.
[0126] S4.4 Report Generation
[0127] The output of this step is an in vitro risk stratification identifier, that is, the group (high-density group or low-density group) to which the selected density index belongs relative to a preset threshold. Complete density matrix data (values of each density index and corresponding cell counts and area), model parameters, and corresponding visualization charts can also be output as supplementary presentation. All of the above outputs are intermediate results for subsequent analysis or manual interpretation.
[0128] Example 2: TCGA-COADREAD queue verification
[0129] To verify the effectiveness and clinical application value of the method described in this invention, a systematic validation analysis was performed using the colorectal adenocarcinoma (COAD) and rectal adenocarcinoma (READ) cohorts from The Cancer Genome Atlas (TCGA) project.
[0130] Dataset Description
[0131] The original data consisted of 556 digital images of hematoxylin-eosin stained whole sections from the TCGA-COADREAD cohort. After deduplication by retaining the section with the largest tumor core area for each patient, 549 cases were obtained. Further, cases with complete overall survival and progression-free interval information and a follow-up time greater than zero were retained, resulting in a final cohort of 524 cases (392 cases of colon adenocarcinoma and 132 cases of rectal adenocarcinoma). This cohort had 104 overall survival events, 138 progression-free interval events, and 66 disease-specific survival events (evaluable in 506 cases). Pathological stages were: stage I (95 cases), stage II (184 cases), stage III (157 cases), and stage IV (71 cases), with 17 cases lacking stage information. The median age was 68 years (interquartile range 58–75 years), and the median overall survival follow-up time was 22.7 months.
[0132] The 524 slides came from 35 different testing institutions, each with different sampling, slide preparation, staining, and scanning conditions. The scanning resolution ranged from 0.2325 micrometers to 0.2520 micrometers per pixel. This heterogeneity provided stringent conditions for verifying the robustness of the method of the present invention.
[0133] Full process results
[0134] The 524 whole-slice digital images were processed in a fully automated manner according to steps S1 to S3 described in Example 1.
[0135] In step S1, the cell segmentation and classification detected a median total number of cell instances in each whole-slice digital image of 67,784 (interquartile range 42,898 to 106,544, minimum 5,007, maximum 442,373), with the detection scale varying with tissue area and cell density.
[0136] Step S2 successfully generated three functional regions—the tumor core, the invasion front, and the stroma—in all 524 cases. The median threshold, adaptively determined based on the epithelial cell fraction distribution, was 0.385 (interquartile range 0.303 to 0.572). The significant variation in this threshold across different cases underscores the necessity of using an adaptive threshold rather than a fixed one: a uniform fixed threshold would systematically misjudge tumor boundaries in cases with excessively high or low epithelial cell proportions. The median areas of the three functional regions were 8.8 mm² (interquartile range 4.0 to 18.5), 9.7 mm² (interquartile range 6.4 to 14.6), and 16.4 mm² (interquartile range 4.4 to 49.5), respectively.
[0137] Step S3 generates a 6x3 density matrix for each patient. Taking a typical case from the cohort (ID TCGA-AA-A01R, 74211 cells detected in the whole slide, adaptive threshold 0.475, areas of the three functional regions 3.6, 7.5, and 33.4 square millimeters respectively) as an example, its density matrix is shown in the table below (unit: cells per square millimeter):
[0138] Cell types Tumor core area Immersion Frontier Area interstitial area epithelial cells 1137.7 875.9 726.0 lymphocytes 232.6 336.5 451.4 plasma cells 78.3 80.4 104.0 neutrophils 1.7 2.1 0.7 eosinophils 9.7 11.7 15.0 connective tissue cells 136.6 230.9 401.1
[0139] In the entire cohort of 524 cases, the spatial distribution pattern of each element in the density matrix consistently showed that the density of epithelial cells monotonically decreased from the tumor core area to the stromal area, while the density of lymphocytes and connective tissue cells showed the opposite trend. This pattern conforms to the known histological patterns of colorectal cancer, indicating that the regional division steps of this invention do indeed separate histologically distinct regions, rather than arbitrarily dividing them.
[0140] Prognostic association analysis
[0141] Based on the density matrix, each functional region was further subdivided according to the presence or absence of epithelial cell aggregation, resulting in intraepithelial and non-epithelial subregions within the tumor core. Lymphocyte density in the non-epithelial subregion of the tumor core (denoted as D_lym_CT-stroma) was used as a pre-defined assessment indicator, and its association with clinical outcomes was evaluated using a Cox proportional hazards regression model, with the hazard ratio calculated per 100 cells per square millimeter.
[0142] The number of cases and endpoint events included in the analysis were as follows: overall survival: 524 cases, 104 events; progression-free survival: 524 cases, 138 events; disease-specific survival: 506 cases, 66 events. In the table below, hazard ratios are denoted as HR, and their 95% confidence intervals are denoted as 95% CI. Multivariate models were adjusted for pathological stage, age, and primary site. The results are shown in the table below:
[0143] end Model HR 95% CI p-value Total Survival univariate 0.936 0.889-0.985 0.011 Total Survival Multivariable 0.931 0.882-0.983 0.010 Progression-free interval univariate 0.948 0.908-0.990 0.016 Progression-free interval Multivariable 0.950 0.907-0.994 0.028 Disease-specific survival univariate 0.934 0.876-0.997 0.041 Disease-specific survival Multivariable 0.929 0.866-0.997 0.041
[0144] This indicator was consistent in direction and statistically significant across the three endpoints of overall survival, progression-free interval, and disease-specific survival, and maintained independent prognostic value after adjusting for pathological stage, age, and primary site.
[0145] Negative control: Verification of the necessity of the partitioning step
[0146] To verify the technical contribution of the region division step (step S2) of this invention, rather than simply repeating the known rule that "more lymphocytes mean a better prognosis", two levels of controls were set up.
[0147] The first level is the non-partitioned control, which represents the indicators obtained without performing step S2. The total lymphocyte density across the entire slice (weighted by the cell count and area of the three functional regions, denoted as D_lym_WS) was used as the first non-partitioned indicator; the proportion of lymphocytes to the total number of cells in the entire slice (denoted as F_lym_WS, corresponding to the global proportion output by existing block-level tumor-infiltrating lymphocyte detection methods) was used as the second non-partitioned indicator. The analyses of the three endpoints were repeated. The results are shown in the table below:
[0148] In the table below, D_lym_WS refers to the total lymphocyte density of the whole section, F_lym_WS refers to the proportion of lymphocytes in the whole section, and the number of cases and events are the same as in the table above.
[0149] index end HR 95% CI p-value D_lym_WS Total Survival 0.967 0.897-1.043 0.386 D_lym_WS Progression-free interval 0.972 0.912-1.037 0.394 D_lym_WS Disease-specific survival 0.929 0.841-1.026 0.144 F_lym_WS Total Survival 0.943 0.827-1.076 0.385 F_lym_WS Progression-free interval 0.987 0.882-1.104 0.813 F_lym_WS Disease-specific survival 0.874 0.740-1.031 0.110
[0150] Neither of the two non-regional indicators reached statistical significance at any of the three endpoints.
[0151] The second level was intra-regional control: within the same tumor core region, the lymphocyte density in the intraepithelial subregion (denoted as D_lym_CT-epi) was used to replace the lymphocyte density in the non-epithelial subregion, and the overall survival analysis was repeated. The results showed no significant association (P=0.510).
[0152] Based on the above two levels of control, further incremental information testing was conducted: using the total lymphocyte density of whole slices as the base model, adding the lymphocyte density of the non-epithelial subregion of the tumor core significantly improved the model fit (likelihood ratio test, overall survival P=0.004, progression-free interval P=0.009); while the reverse operation, i.e., adding the total lymphocyte density of whole slices to the model that already included this partition index, did not bring significant improvement (overall survival P=0.207, progression-free interval P=0.286).
[0153] The above results indicate that the prognostic information provided by the method of this invention cannot be replaced by global indicators obtainable without region partitioning—the total lymphocyte density and proportion across the entire slice have no prognostic value, while the density of the same cell type within a specific partition after partitioning in step S2 has independent prognostic value that cannot be replaced by total indicators. This directly proves that the region partitioning step is a necessary step to produce the technical effect, rather than an optional intermediate step.
[0154] Robustness verification across testing institutions
[0155] Given that the slides in this cohort came from 35 submitting institutions and had not undergone any image-level batch correction, a systematic sensitivity analysis was conducted to rule out the possibility that the aforementioned prognostic associations were driven by differences in submitting institutions. Using overall survival as the endpoint, six different batch control methods were sequentially introduced into the original model: stratification by submitting institution, robust variance estimation by submitting institution clustering, correction for scan resolution, correction for the proportion of epithelial cells in the whole slide (as a batch fingerprint), correction for pathological stage, age, and primary site, and correction for all of the above combined with stratification by submitting institution.
[0156] Under the six batch control methods described above, the hazard ratios remained stable between 0.924 and 0.936, all maintaining statistical significance. This result indicates that the density matrix indices output by the method of this invention are independent of specific slide preparation and scanning conditions, and possess cross-institutional transferability.
[0157] Explanation regarding threshold portability
[0158] The absolute value of the density index described in this invention will drift overall with the scanning equipment and film preparation conditions. Therefore, when applying this method to a new queue, the cut-off point for risk stratification should be redefined within that queue (e.g., taking the median within the queue), rather than directly transplanting the absolute threshold obtained from a certain queue. This invention claims protection for the method of generating the density matrix and the method of region division themselves; the specific cut-off point for risk stratification is determined by parameters during application and does not affect the implementation of the technical solution of this invention.
[0159] Comparative analysis with existing technologies
[0160] The method of this invention is compared with three existing mainstream technical routes:
[0161] Compared with immunohistochemical scoring methods: Methods represented by Immunoscore (such as those described in EP3195191B1, WO2019149817A1, and WO2020161125A1) require additional CD3 and CD8 immunohistochemical staining sections in the immunocellular quantification step, with each section typically only detecting one or two markers. Regarding region segmentation, EP3195191B1 records that tumor masks can be obtained from multiple staining unmixing channels, tumor-specific single-stain sections, or hematoxylin-eosin stained images through segmentation algorithms, with segmentation based on image pixels and their local features as the discrimination units. The difference between this invention and others lies in the following: First, the typing of immune cells does not rely on any additional staining; all six cell types are obtained from the same conventional hematoxylin-eosin stained slide through cell-level segmentation and classification. Second, the tumor mask is not obtained through pixel-level discrimination, but rather by first obtaining the type labels of all cells, then statistically analyzing the proportion of epithelial cells in each spatial grid, and finally applying an adaptive thresholding method to determine the segmentation threshold based on the statistical distribution of this proportion across the entire slide—the discrimination unit is the cellular composition of the grid, not pixels or their local features. Therefore, this invention outputs a density matrix of six cell types multiplied by three functional regions, significantly increasing the information dimensionality compared to scoring systems based on two markers multiplied by two regions.
[0162] Compared with the H&E-based tumor-infiltrating lymphocyte detection route: This route (as described in US11164312B2) binary classifies tissue blocks into lymphocyte-positive or lymphocyte-negative infiltration, outputting a single lymphocyte spatial density map, without distinguishing immune cell subtypes such as lymphocytes, plasma cells, and neutrophils, nor performing functional zoning of the tumor core and invasion front zones. This invention expands the output from a one-dimensional scalar to a two-dimensional matrix through the classification of six cell types and the division into three regions. The above non-zoning comparison results show that the information carried by this subdivision cannot be replaced by the total lymphocyte count index without zoning: the global lymphocyte ratio output by this route was not significantly associated with overall survival in this cohort (P=0.385), while the lymphocyte density in the non-epithelial subregion of the tumor core after regional division by this invention was significantly associated with overall survival (P=0.011), and the latter still brought a significant improvement in model fit on the basis of the former (likelihood ratio test P=0.004).
[0163] Compared to region attribution methods based on known tumor areas: These methods (such as the method described in CN115428039A) use already identified tumor areas as input, and determine region attribution by defining bounding boxes for each tissue region and calculating their geometric inclusion or intersection relationships with the tumor area. The difference in this invention is that the tumor area is not the input but the output—by statistically analyzing the proportion of epithelial cells to total cells within a spatial grid and applying an adaptive thresholding method to this proportion, a tumor mask is generated from the bottom up. This eliminates the need for pre-provided tumor area annotations and does not rely on intersection calculations of geometric bounding boxes.
[0164] Explanation of the scope of statistical analysis in this embodiment
[0165] The prognostic correlation reported in this embodiment is a confirmatory analysis of a pre-defined indicator. This indicator was pre-selected based on the prognostic significance of cytoplasmic lymphocyte infiltration in the tumor core region in previous literature, and was not obtained by screening all elements of the density matrix. The panoramic analysis of the correlation between each element of the density matrix and each clinical endpoint is an exploratory analysis, and its results should not be used as a basis for the technical effect of this invention.
[0166] Example 3: System Implementation
[0167] like Figure 5 As shown, the present invention also provides a spatial quantitative analysis system for the tumor immune microenvironment in pathological images. The system includes four functional modules: a cell segmentation and classification module (M1), an automatic region division module (M2), a density matrix generation module (M3), and a risk stratification module (M4), wherein module M4 is an optional module.
[0168] M1 Cell Segmentation and Classification Module
[0169] like Figure 5 As shown in module M1, the cell segmentation and classification module is responsible for receiving the digital image of the whole slide stained with hematoxylin and eosin, and executing all the processing steps described in step S1 of Example 1, including tissue region detection and background removal, block-by-block processing, cell segmentation and classification based on visual Transformer, and post-processing. The core components of this module are the pre-trained ViT model weight file and the GPU inference engine. The weight file can be obtained from the publicly available sources described in Example 1. The output of module M1 is a structured list of cells, with each record containing a cell identifier, spatial coordinates, cell type, area, and classification confidence.
[0170] In terms of hardware configuration, Module M1 has high GPU computing resource requirements; an NVIDIA A100 (80GB VRAM) or V100 (32GB VRAM) or higher specification GPU is recommended. For a typical full-slice digital image (approximately 40,000 pixels by 40,000 pixels), the GPU inference time of Module M1 is approximately 3 to 10 minutes, depending on the area of the organized region and the GPU model. In batch processing mode, multi-GPU parallel inference is supported to improve throughput.
[0171] M2 Automatic Region Division Module
[0172] like Figure 5 As shown in module M2, the automatic region partitioning module receives the cell list output by module M1 and executes all the processing steps described in step S2 of Example 1, including spatial grid partitioning, epithelial cell fraction calculation, Otsu adaptive threshold binarization, morphological processing (closing operation, opening operation, and connected component area filtering), and definition of three functional regions (CT, IM, ST). This module mainly uses CPU computation and does not rely on GPU acceleration. The processing time of module M2 is typically less than 1 minute.
[0173] The configurable parameters for module M2 include: grid size (default 250 micrometers, range 200 to 300 micrometers), minimum cell count threshold (default 5 cells), closing operation structuring element radius R_close (default 2 grid units), opening operation structuring element radius R_open (default 1 grid unit), minimum connected region area A_min (default 10 grid areas), erosion distance D_erode (default 250 micrometers), and dilation distance D_dilate (default 500 micrometers). The selection of these parameters is based on the following: a grid size of 250 micrometers corresponds to the diameter range of approximately 2 to 5 typical glandular structures in H&E staining, providing sufficient histological context; a closing operation radius of 500 micrometers can bridge sparse epithelial cell areas within the tumor caused by local immune infiltration or necrosis; and an erosion distance of 250 micrometers ensures that the boundary of the tumor core is at least one grid distance away from the tumor-stromal interface, avoiding the inclusion of mixed signals from the boundary transition region in the core region statistics.
[0174] M3 Density Matrix Generation Module
[0175] like Figure 5 As shown in module M3, the density matrix generation module receives the cell list output by module M1 and the three-functional region mask output by module M2, and executes the density calculation process described in step S3 of embodiment one to generate a 6x3 density matrix. This module has extremely low computational complexity, with processing time typically less than a few seconds. Module M3 supports both JSON and CSV structured data formats for its output.
[0176] M4 Risk Stratification Module (Optional)
[0177] like Figure 5 Module M4 and Figure 6 As shown, the risk stratification module is an optional module that receives the density matrix output by module M3 and executes the process described in step S4 of Embodiment 1. This module has built-in Cox proportional hazards regression model parameters and threshold files, and outputs an in vitro risk stratification identifier for the input density index. This identifier is an intermediate result for subsequent analysis or manual interpretation; the module can also output a visual heatmap of the density matrix and a cell distribution map of each functional region as auxiliary presentation, but this invention does not involve the steps of directly drawing disease diagnosis conclusions or providing treatment plans based on the above outputs.
[0178] Overall system configuration
[0179] The recommended hardware configuration for system operation is as follows: GPU server configuration with an NVIDIA A100 (80GB VRAM) or V100 (32GB VRAM) or higher; system memory of at least 64GB; and storage space of at least 1TB SSD for caching and intermediate data storage of the full-slice digital images. The software environment is based on Python 3.8 or later, using the PyTorch deep learning framework for model inference, the OpenSlide library for reading the full-slice digital images, the scikit-image library for morphological processing, and the lifelines library for Cox regression analysis.
[0180] like Figure 6 The processing time shown is approximately 5 to 15 minutes for the complete processing of a single full-slice digital image: preprocessing takes about 30 seconds, GPU inference in module M1 takes about 3 to 10 minutes, region partitioning in module M2 takes about 30 seconds, density calculation in module M3 takes about 5 seconds, and report generation in module M4 takes about 15 seconds. The system supports batch processing mode, which can process multiple full-slice digital images simultaneously to improve overall throughput.
[0181] Alternative solutions description
[0182] Based on the above embodiments, those skilled in the art can make the following substitutions or modifications according to actual needs:
[0183] Grid size alternatives: Spatial grid size can be adjusted from 100 micrometers to 500 micrometers. Smaller grids (100 to 200 micrometers) are suitable for tissues with high cell density and provide higher region division accuracy; larger grids (300 to 500 micrometers) are suitable for tissues with low cell density and provide more stable statistical reliability.
[0184] Cell classification system adjustment: The six-category cell classification system of this invention can be expanded or simplified according to the characteristics of different tumor types. For example, in melanoma analysis, a "melanocyte" category can be added; in scenarios focusing solely on lymphocyte infiltration, it can be simplified to three categories: "epithelial cells," "lymphocytes," and "other cells."
[0185] Cell segmentation model alternatives: In addition to the preferred visual Transformer architecture of this invention, those skilled in the art can also use other cell segmentation models to achieve the function of step S1, including but not limited to HoVer-Net (a CNN model based on horizontal-vertical distance regression), Cellpose (a general cell segmentation model based on gradient flow), and StarDist (a cell nucleus detection model based on star-convex polygons). In comparison, the visual Transformer architecture has advantages in capturing global contextual information and handling morphologically diverse cell populations.
[0186] Region partitioning threshold substitution: In addition to the preferred Otsu adaptive thresholding method of this invention, other adaptive thresholding or clustering methods can also be used, including Gaussian Mixture Model (GMM), Kapur entropy thresholding method, and K-means clustering method. The Otsu method performs best in bimodal distribution scenarios, while the GMM method may provide a more flexible fit in multimodal distribution scenarios.
[0187] Prognostic model alternatives: In addition to the Cox proportional hazards regression model, other survival analysis methods can be used for prognostic assessment, including Random Survival Forest, DeepSurv, and Accelerated Failure Time models. The advantage of Cox regression is that the model is highly interpretable, and the regression coefficients can be directly converted into hazard ratios.
[0188] Supplementary Explanation
[0189] The embodiments described above are merely preferred embodiments of the present invention and do not limit the scope of protection of the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention. Those skilled in the art will understand that the technical features in the above embodiments can be combined arbitrarily; for the sake of brevity, not all possible combinations of these technical features have been described one by one. Furthermore, the TCGA-COAD dataset used in the above embodiments is only a verification example; the method and system described in this invention are applicable to the analysis of digital images of hematoxylin-eosin stained whole sections of all types of solid tumors.
Claims
1. A method for spatial quantitative analysis of the tumor immune microenvironment in pathological images, characterized in that, include: Obtain digital images of whole sections stained with hematoxylin and eosin, and use a cell segmentation and classification model to perform cell-level segmentation and classification on the whole section digital images, outputting the spatial coordinates and cell type of each cell. The cell types include at least six categories: epithelial cells, lymphocytes, plasma cells, eosinophils, neutrophils, and connective tissue cells. The tissue region of the whole-slice digital image is divided into a spatial grid of a preset size. The proportion of epithelial cells in each grid to the total number of cells in that grid is counted as the epithelial cell fraction. Based on the distribution of epithelial cell fractions in all cell-containing grids, an adaptive thresholding method is applied to determine the segmentation threshold, thereby distinguishing between tumor and non-tumor regions. The tumor region mask is then morphologically processed to generate a tumor mask. Based on the tumor mask, the tumor core region is defined by an erosion operation, the invasion front region is defined by an expansion operation, and the remaining tissue regions containing cells are defined as the stroma region. The number of cells of each type and the area of each functional region are counted. The density of N cell types in M functional regions is calculated, and an N-by-M density matrix is generated. The unit of density is cells per square millimeter.
2. The method according to claim 1, characterized in that, The cell segmentation and classification model adopts an encoder-decoder architecture based on a multi-head self-attention mechanism. The encoder divides the input image into multiple image blocks and extracts global contextual features through a multi-layer encoder. The decoder outputs pixel-by-pixel cell instance segmentation masks and class predictions through feature upsampling and refinement.
3. The method according to claim 1, characterized in that, In the cell-level segmentation and classification steps, the whole-slice digital image is processed block by block, with each block ranging in size from 256 pixels by 256 pixels to 1024 pixels by 1024 pixels. Overlapping areas are set between adjacent blocks to eliminate boundary effects. Furthermore, before the cell-level segmentation and classification, tissue region detection is performed on the whole-slice digital image, and background regions are removed based on grayscale thresholding or color space conversion.
4. The method according to claim 1, characterized in that, The size of the preset spatial grid is 200 to 300 micrometers; the adaptive threshold method is the Otsu adaptive threshold method, which automatically determines the optimal threshold for distinguishing tumor areas and non-tumor areas by maximizing the inter-class variance based on the distribution of epithelial cell fractions in all cell-containing grids within the tissue region.
5. The method according to claim 1 or 4, characterized in that, The morphological processing includes: sequentially performing a closing operation to fill the gaps inside the tumor area, performing an opening operation to remove isolated noise regions, and performing connected region area filtering to remove small regions with an area smaller than a preset area threshold.
6. The method according to claim 1, characterized in that, The tumor core region is obtained by eroding the tumor mask inward by a distance of 200 to 300 micrometers; the infiltration front region is the annular area between the area where the tumor mask expands outward by 400 to 600 micrometers and the tumor core region.
7. The method according to claim 1, characterized in that, The method further includes: dividing the functional region into epithelial subregions and non-epithelial subregions based on whether epithelial cells are aggregated, and then calculating the density of each type of cell in each subregion.
8. The method according to claim 7, characterized in that, The method further includes: using the lymphocyte density in the non-epithelial subregion of the tumor core area as a selected density index, comparing the index with a preset threshold or inputting it into a statistical model, and outputting an in vitro risk stratification identifier; the statistical model is a Cox proportional hazards regression model, and the preset threshold is determined based on the median or optimal cut-off point of the selected density index within the sample set to be evaluated.
9. A spatial quantitative analysis system for the tumor immune microenvironment in pathological images, characterized in that, include: The cell segmentation and classification module is used to acquire digital images of whole slides stained with hematoxylin and eosin. The cell segmentation and classification model is used to perform cell-level segmentation and classification on the whole slide digital images, and outputs the spatial coordinates and cell type of each cell. The cell types include at least six categories: epithelial cells, lymphocytes, plasma cells, eosinophils, neutrophils and connective tissue cells. An automatic region segmentation module is used to divide the tissue region of the whole-slice digital image into a spatial grid of a preset size, calculate the proportion of epithelial cells in each grid to the total number of cells in that grid as the epithelial cell score, and determine the segmentation threshold based on the distribution of the epithelial cell score of all cell-containing grids using an adaptive thresholding method. Based on this, the tumor region and non-tumor region are distinguished, and the tumor region is morphologically processed to generate a tumor mask. Based on the tumor mask, the tumor core region is defined by an erosion operation, the invasion front region is defined by an expansion operation, and the remaining tissue regions containing cells are defined as the stroma region. The density matrix generation module is used to count the number of cells of each type and the area of each functional region, calculate the density of N cell types in M functional regions, and generate an N-by-M density matrix, wherein the unit of density is cells per square millimeter.
10. The system according to claim 9, characterized in that, The automatic region division module is also used to perform a secondary division of the functional region according to whether epithelial cells are aggregated, to obtain an intraepithelial subregion and a non-epithelial subregion; the system also includes a risk stratification module, which uses the lymphocyte density in the non-epithelial subregion of the tumor core area as a selected density index to output an in vitro risk stratification identifier.
Citation Information
Patent Citations
Deep learning systems and methods for joint cell and region classification in biological images
CN111417958B
Method for efficiently determining density and spatial relationship characteristics of multiple cell types in region of tissue
CN115428039A
Systems and methods for calculating immune scores
EP3195191B1
System and method to quantify tumor-infiltrating lymphocytes (TILs) for clinical pathology analysis based on prediction, spatial analysis, molecular correlation, and reconstruction of TIL information identified in digitized tissue images
US11164312B2
Methods and systems for evaluation of immune cell infiltrate in stage iii colorectal cancer
WO2019149817A1
