A gastric cancer three-dimensional tumor microenvironment biomarker discovery system and method

Abstract

The application discloses a gastric cancer three-dimensional tumor microenvironment biomarker discovery system and method, and belongs to the technical field of digital pathology and artificial intelligence assisted medical image analysis. The system comprises five modules of data acquisition, data preprocessing, three-dimensional cell graph construction, three-dimensional neighborhood and topological feature extraction and biomarker candidate sorting. The system takes gastric cancer H&E staining continuous sections as input, and outputs a cross-section stable and space-specific candidate biomarker sorting table through a heterogeneous undirected three-dimensional cell graph formed by 40-micron adjacent edges in the section and 10-micron matching edges between the sections, three-dimensional neighborhood division selected by stability-z continuity joint score, and seven-group three-dimensional topological features containing barrier scores, which can be used for gastric cancer tumor microenvironment research and clinical auxiliary analysis.

Description

Technical Field

[0001] This invention relates to the fields of digital pathological image analysis, artificial intelligence-assisted medical image analysis, and spatial tumor omics technology, specifically to a system and method for discovering three-dimensional tumor microenvironment biomarkers in gastric cancer. Background Technology

[0002] Spatial heterogeneity of the tumor microenvironment has a significant impact on the prognosis and treatment response of gastric cancer. Spatial molecular pathology techniques such as multiplex immunofluorescence, multiplex immunohistochemistry, and CODEX can characterize the spatial interactions between tumor cells, immune cells, and stromal cells at the single-cell level. However, these techniques are costly and difficult to adapt to the massive archives of routine pathology slides in clinical practice.

[0003] Conventional hematoxylin and eosin stained sections offer advantages such as large inventory, low cost, and high clinical accessibility. However, existing methods for spatial analysis of the tumor microenvironment based on conventional H&E sections still suffer from the following technical problems: First, existing methods cannot reflect the continuous distribution characteristics of cells along the tissue thickness direction, i.e., the Z-axis. Due to differences in sampling locations between consecutive sections, the cell neighborhood and spatial features obtained by two-dimensional methods are highly sensitive to section location, resulting in poor cross-section repeatability. Second, existing cell neighborhood clustering methods typically use only a single clustering quality index such as the silhouette coefficient to select the number of clusters, without simultaneously considering the repeatability stability of the clustering results and the spatial continuity along the Z-axis, leading to fragmentation of the obtained cell neighborhoods between consecutive sections. Third, existing spatial pathological features are mainly oriented towards two-dimensional planar topological relationships, lacking quantitative indicators adapted to three-dimensional tissues. In particular, there is a lack of quantitative indicators for cross-section cell interactions, as well as a lack of three-dimensional barrier indicators that can characterize the matrix barrier relationship between tumor cells and immune cells, resulting in the loss of important microenvironment interaction information along the tissue thickness direction.

[0004] Therefore, how to construct three-dimensional cell maps using routine clinical H&E serial slices, obtain stable three-dimensional cell neighborhoods across slices, and extract topological features containing cross-slice interaction and three-dimensional barrier information is a technical problem that urgently needs to be solved in the field of digital pathology spatial analysis. Summary of the Invention

[0005] The purpose of this invention is to provide a system and method for discovering three-dimensional tumor microenvironment biomarkers in gastric cancer, so as to solve the technical problems of poor cross-slice consistency, single cluster evaluation, and lack of three-dimensional specific microenvironment quantitative indicators in existing two-dimensional pathological spatial analysis methods.

[0006] The system and method provided in this invention use in vitro pathological slide images as the processing object for automatic screening and spatial feature analysis of candidate biomarkers. They do not use living human or animal bodies as the direct object, nor do they aim to obtain disease diagnosis results.

[0007] This invention provides a three-dimensional tumor microenvironment biomarker discovery system for gastric cancer, including a data acquisition module, a data preprocessing module, a three-dimensional cell map construction module, a three-dimensional neighborhood and topological feature extraction module, and a candidate biomarker ranking module.

[0008] The data acquisition module is used to acquire sequential hematoxylin-eosin stained whole-section image sequences and section registration information of the same gastric cancer tissue specimen. The section registration information includes registration transformation parameters and registration dependencies between sections.

[0009] The data preprocessing module is used to reconstruct the physical Z-axis order of the whole slice image sequence based on the slice registration information, and to perform tissue mask generation, staining normalization, cell nucleus detection and cell type assignment on each slice to obtain single-cell data containing cell two-dimensional coordinates, physical Z-axis sequence number and cell type.

[0010] The 3D cell graph construction module is used to construct a heterogeneous undirected 3D cell graph using cells from the single-cell data as nodes. The heterogeneous undirected 3D cell graph includes intra-slice adjacent edges within the same slice, and cross-slice matching edges between adjacent slices in the physical Z-axis order. The intra-slice adjacent edges and the cross-slice matching edges are distinguished by edge type labels.

[0011] The three-dimensional neighborhood and topological feature extraction module is used to cluster each cell based on the neighborhood composition vector in the heterogeneous undirected three-dimensional cell graph, and to determine the optimal number of cell neighborhoods by a joint score obtained by weighted fusion of four components: silhouette coefficient, Davidson-Bolding index, double random seed K-means clustering consistency index, and Z-axis continuous run length.

[0012] Based on the cell neighborhood division corresponding to the optimal number of cell neighborhoods, multiple three-dimensional topological features are extracted. These multiple three-dimensional topological features include at least: cross-slice cell interaction features based on the Z-score obtained from the cell type label permutation test, and tumor-immune matrix barrier features characterizing the proportion of stromal cells on the shortest path from tumor epithelial cells to immune cells.

[0013] The candidate biomarker ranking module is used to calculate the cross-slice stability index and spatial specificity index for each candidate feature along the physical Z-axis, determine a comprehensive score based on the cross-slice stability index and the spatial specificity index, and output the candidate biomarker ranking result according to the comprehensive score.

[0014] This invention also provides a method for discovering three-dimensional tumor microenvironment biomarkers in gastric cancer. The method is a non-diagnostic in vitro pathological image data processing method, which includes acquiring continuous H&E whole-slice image sequences and slice registration information, reconstructing the physical Z-axis order, constructing a three-dimensional cell map, determining the optimal number of cell neighborhoods, extracting three-dimensional topological features, and outputting the ranking results of candidate biomarkers based on cross-slice stability index and spatial specificity index.

[0015] The present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the above-described method.

[0016] The present invention has at least the following advantages over the prior art:

[0017] To address the first problem mentioned in the background art, namely the poor reproducibility of two-dimensional methods across slices, this invention constructs a heterogeneous undirected three-dimensional cell map that includes adjacent edges within a slice and matching edges across slices. This enables cell neighborhood partitioning to not only depend on the planar adjacency relationship within a single slice, but also to express the cell correspondence between adjacent slices along the tissue thickness direction, thereby reducing the sensitivity of two-dimensional methods to the location of slice sampling.

[0018] To address the second issue, namely the fragmentation of neighborhoods caused by a single clustering metric, this invention employs a four-component weighted joint score—profile coefficient, Davidson-Bolding index, double-random seed clustering consistency, and Z-axis continuous run length—when determining the optimal number of cell neighborhoods. The resulting cell neighborhoods are constrained in terms of both statistical clustering stability and three-dimensional tissue continuity, reducing neighborhood fragmentation along the Z-axis.

[0019] To address the third issue, namely the lack of specific three-dimensional quantitative indicators, the three-dimensional topological features extracted in this invention include at least cross-slice cell interaction features and tumor-immune matrix barrier features. Specifically, cross-slice cell interaction features quantify the spatial interaction relationships between cell types between adjacent slices through edge counting and label permutation tests on the three-dimensional cell map; tumor-immune matrix barrier features quantify the degree of spatial barrier formed by stromal cells by calculating the proportion of stromal cells in the shortest path between tumor cells and immune cells on the three-dimensional cell map.

[0020] Furthermore, this invention employs cross-slice stability and spatial specificity indices to rank candidate features, enabling the screening of candidate biomarkers that possess both cross-slice stability and spatial specificity without relying on external prognostic labels. This makes it suitable for basic pathological cohorts and research data analysis. This invention can achieve three-dimensional tumor microenvironment spatial analysis using only conventional H&E serial sections, without relying on costly multiplex immunofluorescence staining, which is beneficial for conducting large-scale retrospective studies using existing clinical pathology sections. Attached Figure Description

[0021] Figure 1 This is a schematic diagram of the overall process of a three-dimensional tumor microenvironment biomarker discovery system for gastric cancer according to the present invention;

[0022] Figure 2 This is a three-dimensional cell diagram of the present invention. Detailed Implementation

[0023] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments. The described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. It should be noted that the processing object in the embodiments of the present invention is ex vivo tissue slice image data. The system and method provided in the embodiments of the present invention are used for pathological image data processing and candidate biomarker screening, and do not take living human or animal bodies as the direct object, nor is obtaining disease diagnosis results the direct purpose.

[0024] I. System Overall Architecture

[0025] This invention provides a three-dimensional tumor microenvironment biomarker discovery system for gastric cancer. Figure 1 The system includes a data acquisition module, a data preprocessing module, a 3D cell map construction module, a 3D neighborhood and topological feature extraction module, and a candidate biomarker ranking module. Optionally, the system also includes a slice registration module, a visualization module, a storage module, and a computing resource scheduling module. The slice registration module can be used to generate the slice registration information; the slice registration information can also be generated by an external registration device or a pre-registration process and then input to the data acquisition module. In one embodiment, the processing flow of the system includes: S1, acquiring a sequence of consecutive H&E whole-slice images of the same gastric cancer tissue specimen and slice registration information; S2, reconstructing the physical Z-axis order and performing tissue mask generation, staining normalization, cell nucleus detection, and cell type assignment; S3, constructing a heterogeneous undirected 3D cell map; S4, performing 3D neighborhood division and extracting 3D topological features; S5, ranking candidate features based on cross-slice stability and spatial specificity, and outputting the candidate biomarker ranking results.

[0026] II. Data Acquisition and Registration Information

[0027] The data acquisition module is used to acquire consecutive H&E whole-section image sequences and section registration information of the same gastric cancer tissue specimen. In one embodiment, the consecutive H&E whole-section image sequences come from the same paraffin block of the same surgically resected gastric adenocarcinoma specimen.

[0028] The continuous H&E whole-section image sequence comprises multiple tissue sections continuously sliced ​​along the tissue thickness direction. Preferably, the continuous H&E whole-section image sequence includes at least 20 valid sections. When the number of valid sections is insufficient, the system can indicate a decrease in the confidence level of 3D modeling, or downgrade to 2D analysis mode.

[0029] The slice registration information includes at least one of the following: affine registration parameters between slices; non-rigid registration parameters between slices; thin-plate spline transformation parameters; registration dependencies between slices; registration error index; and shared two-dimensional coordinate system information. In one embodiment, slice registration can adopt a progressive registration method between adjacent slices. Using the anchor slice as a global reference, affine coarse alignment is first performed on adjacent slices, followed by non-rigid fine alignment using thin-plate splines. In one embodiment, the system determines the registration quality. If the average alignment deviation of the tissue region of the slice is greater than a preset registration error threshold, the slice registration is determined to be invalid.

[0030] The preset registration error threshold can be 5 micrometers, or it can be adaptively determined based on the slice thickness, tissue deformation degree, and image resolution. For slices with registration failure, the system can remove them from the continuous slice sequence and regenerate registration information and physical Z-axis order based on the remaining slices.

[0031] III. Physical Z-axis Sequence Reconstruction

[0032] The data preprocessing module reconstructs the physical Z-axis order based on the slice registration information. In one embodiment, the registration information includes a registration summary table. The registration summary table records each row of source slices. To target slice The system establishes the registration dependencies based on the registration summary table. The system then constructs a directed registration graph.

[0033]

[0034] in, Each vertex in the image corresponds to a slice. Each directed edge in the graph represents the direction of registration from the source slice to the target slice. The system further determines anchor slices in the registration directed graph. The anchor slice is the reference root node in the registration dependency relationship. In one embodiment, the anchor slice is a slice with an empty target slice number and an empty registration index. In another embodiment, the anchor slice can also be determined by an explicit root node identifier, a reference slice identifier, or a node without a registration target relationship. Starting from the anchor slice, the system traverses the registration directed graph in two directions to obtain the first slice chain:

[0035]

[0036] Second slice chain:

[0037]

[0038] By splicing the two slice chains with the anchor slice, the physical Z-axis order is obtained:

[0039]

[0040] in, For anchoring slices. The physical Z-axis distance between two adjacent slices is determined by the slice thickness parameter. The physical Z-axis coordinates of each slice are given below:

[0041]

[0042] in, This is the sequence number of the slice in the physical Z-axis order. In this embodiment, .

[0043] IV. Tissue Masking and Staining Normalization

[0044] For each slice in the physical Z-axis sequence, the system generates a tissue mask. In one embodiment, the system converts the slice image to the HSV color space and performs Otsu thresholding on the saturation channel to obtain an initial tissue mask. A morphological opening operation is performed on the initial tissue mask to remove isolated noise; then a morphological closing operation is performed to fill small holes within the tissue region, resulting in a final binary tissue mask. The system further performs staining normalization. Using the first slice in the physical Z-axis sequence as the staining reference slice, the remaining slices are normalized to the staining reference slice using the Macenko staining normalization method. After staining normalization, the system outputs a hematoxylin channel concentration map and an eosin channel concentration map. Through the above tissue masking and staining normalization processes, non-biological disturbances caused by tissue damage, differences in staining depth, and differences in scanning conditions between consecutive slices can be reduced.

[0045] V. Cell nuclear detection

[0046] The data preprocessing module also includes a cell nucleus detection unit. In one embodiment, cell nucleus detection is performed based on a staining-normalized hematoxylin channel concentration map. The system uses the Laplacian Gaussian operator to perform dot matrix response detection on the hematoxylin channel concentration map to obtain candidate cell nucleus response maps:

[0047]

[0048] in, The concentration graph of sappanwood is shown. The system employs a Gaussian scale to match the radius of the target cell nucleus. Local extrema are extracted from the candidate cell nucleus response maps, and a watershed segmentation algorithm is used to separate adherent cell nuclei. Finally, the system outputs the two-dimensional coordinates of the center of each cell. Cell area and eccentricity can be output. Optionally, morphological features such as cell nucleus roundness, chromatin texture characteristics, and local optical density characteristics can also be output.

[0049] VI. Cell Type Assignment

[0050] The cell type assignment unit assigns each cell to a preset gastric cancer tumor microenvironment cell type system. In one embodiment, the cell type system includes at least the following nine categories: tumor epithelial cells; normal epithelial cells; lymphocytes; plasma cells; macrophages; neutrophils; fibroblasts; endothelial cells; and unclassified cells (i.e., cells whose morphological characteristics do not belong to the aforementioned eight categories).

[0051] Cell type classification employs a two-stage strategy. In the first stage, cells are coarsely classified morphologically based on cell area, eccentricity, nuclear morphology parameters, and local texture statistics. The classification criteria are as follows: lymphocytes have small, round nuclei with low cytoplasm content; plasma cells have eccentric nuclei and basophilic cytoplasm; macrophages are large with abundant cytoplasm; neutrophils have lobed nuclei; fibroblasts are long and spindle-shaped; and endothelial cells are spindle-shaped with accompanying tubular structures. Based on these morphological characteristics, cells are divided into epithelial candidates, lymphocytes, plasma cells, macrophages, neutrophils, fibroblasts, endothelial cells, and unclassified cells.

[0052] In the second stage, for epithelial candidates, image patches are extracted with the cell nucleus as the origin. The image patches cover a single cell nucleus and the surrounding cytoplasm and adjacent mesenchymal regions, ensuring that the encoded information includes nuclear morphological atypia and local microenvironment texture features. The image patches are then input into a pre-trained pathological image encoder.

[0053] This preferred embodiment uses a pathological pre-trained ViT-B / 16 as the pathological image encoder; the input image patch size is 224×224 pixels; the encoder is pre-trained based on a subset of publicly available gastric cancer pathological slides from TCGA, with the training dataset containing over 10,000 gastric cancer regional pathological images; the training loss function is InfoNCE contrastive learning loss; the training objective is to learn the morphology of cell nuclei, nuclear atypia, and tissue microenvironment morphology features within the pathological image patch; the output is a 512-dimensional image embedding vector. This pre-trained encoder, through pathological image contrastive learning, can encode atypia information such as cell nucleus area variation and nuclear polymorphism into a 512-dimensional embedding vector space; image patches corresponding to tumor epithelium will form clusters in the embedding space, while normal epithelial samples are distributed in other clusters. Therefore, the two types of epithelium can be distinguished through HDBSCAN density clustering.

[0054] HDBSCAN density clustering is performed on the cell image embedding vectors. The preferred clustering parameters are: minimum cluster size (min_cluster_size) = 50, minimum number of samples (min_samples) = 10. The atypia thresholds are: a fixed and preferred threshold of 0.22 for the normalized nuclear area variance; and a fixed and preferred threshold of 0.18 for the nuclear outline roundness standard deviation, which is an indicator of polymorphism. An adaptive threshold scheme is also supported: the nuclear area variance and polymorphism index of all epithelial candidate samples are statistically analyzed, and the 75th quantile of all samples is used as the dynamic threshold. That is, the first preset threshold (area variance threshold) and / or the second preset threshold (polymorphism threshold) can be fixed values ​​or adaptively determined values ​​based on the statistical distribution of the samples. Clusters that meet the preset atypia conditions in the clustering results are labeled as tumor epithelial cells, and the remaining epithelial candidate clusters are labeled as normal epithelial cells.

[0055] Through the above two-stage strategy, the system can obtain the spatial distribution of major cell types in the tumor microenvironment from conventional H&E sections without relying on multiple immunostaining.

[0056] VII. Construction of 3D Cell Map

[0057] The three-dimensional cell map construction module uses all detected cells as nodes. Figure 2 Constructing a heterogeneous, undirected 3D cell map:

[0058]

[0059] Each node It has cell numbering and two-dimensional coordinates Attributes include physical Z-axis index and cell type. The edge set includes two types of edges: the set of adjacent edges within the slice. Cross-slice matching edge set The two types of edges are distinguished by a binary edge type label. For example, flag=0 indicates an adjacent edge within a slice, and flag=1 indicates a matching edge across slices.

[0060] For any slice Any cell within The system searches for candidate neighbor cells within the same slice that are at a planar distance of no more than 40 micrometers:

[0061]

[0062] in, This is the conversion factor from pixels to micrometers. The system sorts data by planar distance from smallest to largest and retains the most frequent distances. Find the nearest neighbor. Then establish the intra-slice adjacency edges.

[0063] Explanation of reasonable parameter range: The reasonable working range for the adjacency radius is 30-50 micrometers; less than 30 micrometers is insufficient for capturing neighborhood information, while greater than 50 micrometers introduces too many irrelevant distant cells.

[0064] For two adjacent slices in the physical Z-axis sequence and The system treats them as being in the same shared two-dimensional coordinate system. For slices any cell on The system slices Query its nearest neighbor cells If the planar distance between the two is no more than 10 micrometers, then a cross-slice matching edge is established:

[0065]

[0066] The 3D distance property of the matching edge across a slice is determined by both the planar distance and the slice thickness.

[0067]

[0068] in, .

[0069] Explanation of reasonable parameter range: The reasonable working range for cross-slice matching threshold is 7-13 micrometers; if the threshold is too high, it will incorrectly match non-homologous cells, and if the threshold is too low, it will lose true cross-slice pairing.

[0070] Through the two types of edges mentioned above, the three-dimensional cell map can simultaneously express the close symbiotic relationship between cells within the same tissue slice, as well as the cell correspondence between adjacent slices along the tissue thickness direction.

[0071] VIII. Three-dimensional neighborhood partitioning

[0072] The 3D neighborhood and topological feature extraction module includes a neighborhood discovery unit. For each cell... The system obtains a set of neighbors on the three-dimensional cell map whose three-dimensional Euclidean distance does not exceed the three-dimensional neighborhood radius. The three-dimensional neighborhood radius is preferably 50 micrometers, and the reasonable working range is 40-60 micrometers.

[0073]

[0074] The neighbor set includes neighbors within the same slice and cross-slice neighbors on adjacent slices. Reasonable parameter range explanation: a 3D neighborhood radius less than 40 micrometers cannot capture complete microenvironment information, while a radius greater than 60 micrometers will introduce distant irrelevant cells, diluting neighborhood specificity.

[0075] The system counts the number of each cell type in the neighbor set, including the central cell itself, to obtain the neighborhood composition vector:

[0076]

[0077] in, For cells The type is a one-hot vector. In one embodiment, the neighborhood composition vector can be batch-computed using sparse matrix multiplication:

[0078]

[0079] in, It is a symmetric binary adjacency matrix of a three-dimensional cell graph, containing intra-slice adjacency edges and cross-slice matching edges; It is an identity matrix used to include the central cell itself in the composition; For cell type unique heat matrix; This is a row-normalized diagonal matrix. For the matrix... Perform principal component analysis for dimensionality reduction. In one embodiment, retain a cumulative variance contribution rate of at least [value missing]. The first few principal components yield the reduced-dimensional neighborhood composition matrix. The system is in Number of candidate clusters Traversal, The reasonable working range is 4-25; for each candidate cluster number, K-means clustering is used to obtain the cell neighborhood division result.

[0080] IX. Determining the Optimal Number of Cell Neighborhoods through Joint Scoring

[0081] To simultaneously evaluate clustering quality, clustering stability, and Z-axis continuity, this invention introduces a joint scoring method:

[0082]

[0083] in: The profile coefficient; The Davidson-Bolding index is used; the lower the value, the better. It serves as a consistency index between K-means clustering results under two different random seeds; The Z-axis continuous run length component; the weighting parameter takes the value of The sum of the weights equals 1.

[0084] Explanation of reasonable parameter range: The weights can be slightly adjusted under the condition that the sum is 1; the number of label replacements is preferably 200 times, and the reasonable range is 100-500 times; the Z-axis scaling weight (Ripley) is preferably 0.8, and the reasonable range is 0.6-1.0.

[0085] The Z-axis continuous run length component is used to characterize the degree of consecutive occurrence of the same cell neighborhood along the physical Z-axis on consecutive slices. For each cell neighborhood... Get the set of slice numbers that appear in the physical Z-axis order. Calculate the average consecutive run length:

[0086]

[0087] in, For cell neighborhood The number of slices covered; This is an indicator function; it takes a value of 1 when the condition is met and 0 otherwise, used to count the number of Z-axis jumps. After normalization, we get The system selects to make Take the maximum value As the optimal number of cell neighborhoods ,Right now:

[0088]

[0089] use Perform final clustering on all cells to obtain cell neighborhood labels for all cells.

[0090] 10. Three-dimensional topological feature extraction

[0091] In the optimal cell neighborhood number Given a cell neighborhood partitioning, the system extracts three-dimensional topological features. In one embodiment, the three-dimensional topological features include seven families of features.

[0092] Cellular composition characteristics

[0093] For each cell neighborhood The proportion vectors of each cell type within the cell and the total proportion of the cell neighborhood are statistically analyzed to form the cell neighborhood composition characteristics. These cell composition characteristics can reflect the enrichment degree of tumor cells, immune cells, and stromal cells in different microenvironment regions.

[0094] Cellular diversity characteristics

[0095] The neighborhood of each cell forms a vector. The diversity index is calculated, including at least one of Shannon entropy, Simpson's index, inverse Simpson's index, and evenness. Shannon entropy is defined as:

[0096]

[0097] The Simpson index is defined as:

[0098]

[0099] For the number of cell types, It is a very small constant used to prevent taking the logarithm with respect to zero. Uniformity: .

[0100] 3D Ripley Spatial Scattering Features

[0101] Z-axis scaling weight The 3D distance scaled along the Z-axis is defined as follows:

[0102]

[0103] in, The distance between two cells in the same shared two-dimensional coordinate system. The physical Z-axis distance between the slices containing the two cells. The Z-axis scaling weight is used to adjust the contribution of the tissue thickness direction to the spatial scattering analysis.

[0104]

[0105]

[0106] This indicates that the cells are clustered together. This indicates that the cells are distributed in a repulsive manner; during calculation, the tissue mask boundary correction method can be used to reduce boundary estimation bias.

[0107] Cross-slice cell interaction characteristics

[0108]

[0109] Number of random label permutations ; This counts the edges of corresponding cell type pairs within the 3D cell graph, including cross-slice matching edges.

[0110] 3D nearest neighbor distance features

[0111]

[0112]

[0113] in, For type The collection of all cells. Conditions Indicates limited cells With cells Located on the same slice; condition Indicates limited cells Located in cells On the next adjacent slice of the slice.

[0114] Cell Neighborhood Transfer Entropy Features

[0115]

[0116] in, For cells The neighbor set belongs to the cell neighborhood The number of cells; For cells The total number of neighbors; It is the set of all cell neighborhoods.

[0117] Tumor-immune matrix barrier score characteristics

[0118] The system samples tumor epithelial cells and immune cells separately to form cell pairs. In this embodiment, 1000 pairs of tumor epithelial cells and immune cells were randomly selected to calculate the average barrier score; the sampling size can be adaptively adjusted according to the total number of cells in the tissue, with a minimum of 200 pairs to ensure statistical stability. Tumor epithelial cells were calculated on a three-dimensional cell map or its sub-map. to immune cells shortest path The barrier score is defined as the proportion of stromal cells in the shortest path, excluding the start and end points.

[0119]

[0120] It is a collection of stromal cell types, including fibroblasts and macrophages; the barrier score can also be calculated based on the three-dimensional shortest path across slices to more completely characterize the matrix barrier effect in the Z-axis direction.

[0121] XI. Ranking of Candidate Biomarkers

[0122] The candidate biomarker ranking module performs unsupervised ranking of three-dimensional topological features. For each candidate feature... Obtaining the value sequence by slicing along the physical Z-axis .

[0123]

[0124] Stability metrics: For cross-slice cell interaction features, the spatial specificity index is the absolute value of the interaction Z-score; for non-interaction features, the spatial specificity index is the absolute deviation of the feature value relative to the global mean divided by the global average absolute deviation; the global refers to the statistical set of all values ​​of the feature within the entire tissue region corresponding to the entire three-dimensional cell map.

[0125]

[0126] in, The global mean of the candidate features. Overall score:

[0127]

[0128] Output a ranking table of candidate biomarkers in descending order of their comprehensive scores.

[0129] 12. Visual Output

[0130] Optional visualization module outputs include: 3D cell cross-section, cell neighborhood distribution map, Ripley curve, cross-slice heatmap, barrier path diagram, sorting table, and feature Z-axis variation curve, for scientific research review and auditing.

[0131] XIII. Method Examples

[0132] This invention also provides a method for discovering three-dimensional tumor microenvironment biomarkers in gastric cancer. The method is a non-diagnostic method for processing ex vivo pathological image data, comprising the following steps: S1, acquiring a sequence of consecutive H&E whole-slice images and slice registration information from the same gastric cancer tissue specimen; S2, reconstructing the physical Z-axis order based on the slice registration information, and performing tissue mask generation, staining normalization, cell nucleus detection, and cell type assignment; S3, constructing a heterogeneous undirected three-dimensional cell map with cells as nodes, wherein the heterogeneous undirected three-dimensional cell map includes intra-slice adjacent edges and cross-slice matching edges; S 4. Clustering is performed based on the cell neighborhood composition vector, and the optimal number of cell neighborhoods is determined by a joint score obtained by weighted fusion of four components: silhouette coefficient, Davidson-Bolding index, double random seed K-means clustering consistency index, and Z-axis continuous run length; S5. Three-dimensional topological features are extracted based on the optimal cell neighborhood partitioning; S6. Cross-slice stability index and spatial specificity index are calculated for each candidate feature along the physical Z-axis order; S7. A comprehensive score is determined based on the cross-slice stability index and the spatial specificity index, and the candidate biomarker ranking results are output according to the comprehensive score.

[0133] Optional steps: Remove damaged or poorly stained sections; remove sections with registration failures; update the physical Z-axis order; calculate features for tumor core / infiltration front / normal mucosa partitions; adaptive parameter adjustment.

[0134] XIV. Storage Media

[0135] This invention also provides a computer-readable storage medium storing a computer program thereon. When the computer program is executed by a processor, it implements all or part of the steps in the above method embodiments. The storage medium includes a hard disk, a solid-state drive, flash memory, an optical disk, and a cloud storage medium.

[0136] XV. Supplementing and Optimizing Implementation Methods

[0137] Partition modeling optimization

[0138] The system can automatically divide the tumor core area, invasion front area, and normal mucosa area, and independently calculate the three-dimensional topological features of each region to adapt to the spatial heterogeneity of gastric cancer.

[0139] 3D barrier score optimization

[0140] Barrier scores can be calculated based on the shortest path across slices in three dimensions, characterizing the matrix barrier effect in the Z-axis direction.

[0141] Adaptive parameter strategy

[0142] The Z-axis weight can be adjusted according to the slice thickness, and the matching threshold can be adjusted according to the registration error.

[0143] Large-scale computation optimization

[0144] WSI block partitioning, sparse matrix storage, KD-Tree nearest neighbor search, permutation downsampling, batch Ripley computation, and GPU parallel acceleration.

[0145] XVI. Experimental Data and Technical Results

[0146] Table 1 shows the comparison between the present invention and the two-dimensional baseline method. The technical solution proposed in this invention is significantly superior to the two-dimensional baseline method.

[0147] Table 1 Comparison of the effects of the present invention and the two-dimensional baseline method.

[0148] Cell Neighborhood Cross-Slice ICC 0.842 0.581 Average coefficient of variation of candidate features 0.183 0.362 Cross-Z-axis interaction Z-score peak 12.6 not applicable Barrier score recall 0.774 not applicable Mean stability of the top 10 candidate features 0.897 0.712

[0149] Furthermore, ablation experiments have demonstrated that the continuous Z-axis run component and cross-slice matching edge are key technical features of this invention, and their removal significantly reduces cross-slice stability (Table 2).

[0150] Table 2 Ablation Experiment

[0151] Complete invention 0.842 0.183 0.91 Remove Z consecutive run components 0.615 0.327 0.54 Remove cross-slice matching edges 0.593 0.351 0.48 Two-dimensional baseline method 0.581 0.362 Uncomputable

[0152] XVII. Explanation of Clinical Validation Data

[0153] In a retrospective cohort analysis for scientific research purposes, the system of this invention was applied to ex vivo slide images corresponding to 120 follow-up specimens of gastric adenocarcinoma, outputting a list of candidate biomarkers. After obtaining the list of candidate biomarkers, researchers conducted post-hoc correlation analysis using existing follow-up survival data outside of this system. They observed statistically significant differences in overall survival between groups with high and low overall scores for the top 10 three-dimensional biomarkers, as shown in the Log-rank analysis. The AUC was 0.81; the optimal feature AUC of the two-dimensional baseline was only 0.64.

[0154] Important limitations: The survival and AUC indicators mentioned above are not output results of this system; this system itself does not perform prognostic prediction or disease diagnosis, but only outputs a ranking list of candidate biomarkers; follow-up correlation is only for post-scientific research to support the output products of this system, used to illustrate that the output candidate biomarkers have potential biological relevance, and does not represent that the system of this invention is used for disease diagnosis or prognostic judgment.

[0155] XVIII. Terminology Explanation

[0156] H&E staining: Hematoxylin-eosin staining. WSI: Whole-section image. CN: Cell neighborhood. Physical Z-axis order: The actual stacking order of consecutive sections reconstructed along the tissue thickness direction. Anchored section: The reference root section in the registration dependency relationship. Intra-slice adjacent edge: An edge between cells within the same slice based on planar distance. Cross-slice matching edge: An edge between adjacent slices based on nearest neighbor matching in a shared coordinate system. Joint score: A multi-index score used to select the optimal number of cell neighborhoods. Z-axis continuous run length: The average length of consecutive occurrences of the same cell neighborhood on consecutive slices. Barrier score: The proportion of stromal cells in the shortest path from tumor cells to immune cells. Cross-slice interaction Z-score: A significant indicator of cell type interaction based on the number of edges in the 3D cell map and the null distribution of label permutation. : The conversion factor from pixel to micrometer.

[0157] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. 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.

Claims

1. A three-dimensional tumor microenvironment biomarker discovery system for gastric cancer, characterized in that, The system includes a data acquisition module, a data preprocessing module, a 3D cell map construction module, a 3D neighborhood and topological feature extraction module, and a candidate biomarker ranking module. The data acquisition module is used to acquire a sequence of consecutive hematoxylin-eosin stained whole-section images of the same gastric cancer tissue specimen and section registration information. The section registration information includes registration transformation parameters and registration dependencies between sections. The data preprocessing module is used to reconstruct the physical Z-axis order of the whole-slice image sequence based on the slice registration information, and to perform tissue mask generation, staining normalization, cell nucleus detection, and cell type assignment on each slice to obtain single-cell data containing cell two-dimensional coordinates, physical Z-axis sequence number, and cell type. The cell types include the following nine categories: tumor epithelial cells, normal epithelial cells, lymphocytes, plasma cells, macrophages, neutrophils, fibroblasts, endothelial cells, and unclassified cells whose morphology does not belong to the aforementioned eight categories. The three-dimensional cell map construction module is used to construct the single-cell data using cells as nodes. A heterogeneous undirected 3D cell graph is constructed. The heterogeneous undirected 3D cell graph includes intra-slice adjacent edges within the same slice, and cross-slice matching edges between adjacent slices in the physical Z-axis order. The intra-slice adjacent edges and the cross-slice matching edges are distinguished by edge type labels. The 3D neighborhood and topological feature extraction module is used to cluster cells based on the neighborhood composition vector of each cell in the heterogeneous undirected 3D cell graph, and to determine the optimal number of cell neighborhoods by a joint score obtained by weighted fusion of four components: silhouette coefficient, Davidson-Bolding index, double random seed K-means clustering consistency index, and Z-axis continuous run length. Based on the cell neighborhood division corresponding to the optimal number of cell neighborhoods, multiple three-dimensional topological features are extracted. The multiple three-dimensional topological features include at least: cross-slice cell interaction features based on the Z-score obtained by cell type label permutation test, and tumor-immune matrix barrier features that characterize the proportion of stromal cells on the shortest path from tumor epithelial cells to immune cells. The candidate biomarker ranking module is used to calculate the cross-slice stability index and spatial specificity index for each candidate feature along the physical Z-axis, determine a comprehensive score based on the cross-slice stability index and the spatial specificity index, and output the candidate biomarker ranking result according to the comprehensive score.

2. The three-dimensional tumor microenvironment biomarker discovery system for gastric cancer according to claim 1, characterized in that, The data preprocessing module reconstructs the physical Z-axis order based on the slice registration information, including: constructing a registration directed graph according to the slice registration information; wherein each vertex in the registration directed graph corresponds to a slice, and each directed edge represents the direction from which the source slice is registered to the target slice; identifying slices in the registration directed graph that do not have a registration target relationship as anchor slices; starting from the anchor slice, traversing along two directions in the registration directed graph to obtain a first slice chain and a second slice chain; and concatenating the first slice chain, the anchor slice, and the second slice chain to obtain the physical Z-axis order.

3. The three-dimensional tumor microenvironment biomarker discovery system for gastric cancer according to claim 1, characterized in that, The tissue mask generation, staining normalization, and cell nucleus detection include: for each slice, performing Otsu thresholding on the saturation channel of the HSV color space to obtain an initial tissue mask; performing morphological opening operations on the initial tissue mask to remove isolated noise points, and performing morphological closing operations to fill small holes in the tissue area to obtain a binary tissue mask; selecting the first slice in the physical Z-axis sequence as the staining reference slice; using the Macenko staining normalization method, normalizing the staining vectors of the remaining slices to the staining matrix of the staining reference slice, and outputting a hematoxylin channel concentration map and an eosin channel concentration map; using the Laplacian Gaussian operator to perform speckle response detection on the staining normalized hematoxylin channel concentration map to obtain candidate cell nucleus response maps; performing local extremum extraction on the candidate cell nucleus response maps, and using the watershed segmentation algorithm to separate adherent cell nuclei, outputting the center two-dimensional coordinates, cell area, and eccentricity of each cell.

4. The three-dimensional tumor microenvironment biomarker discovery system for gastric cancer according to claim 1, characterized in that, The cell type classification adopts a two-stage classification strategy: In the first stage, based on cell area, cell eccentricity, cell nuclear morphology parameters and local texture statistics, cells are roughly divided into epithelial candidates, lymphocytes, plasma cells, macrophages, neutrophils, fibroblasts, endothelial cells and unclassified cells. In the second stage, for the epithelial candidates, image patches centered on the cell centers are extracted, and the image patches are input into a pre-trained pathological image encoder to obtain cell image embedding vectors; density clustering is performed on the cell image embedding vectors; clusters that meet the preset atypia conditions in the clustering results are labeled as tumor epithelial cells, and the remaining epithelial candidates are labeled as normal epithelial cells; the preset atypia conditions are: the variance of the cell nuclear area is higher than a first preset threshold, or the cell morphology pleomorphism is higher than a second preset threshold; the first preset threshold and the second preset threshold are pre-set fixed thresholds, or adaptive thresholds calculated based on the statistical distribution of all epithelial candidate samples.

5. The three-dimensional tumor microenvironment biomarker discovery system for gastric cancer according to claim 1, characterized in that, The intra-slice adjacent edges and cross-slice matching edges are constructed as follows: for any cell in any slice, candidate neighbor cells with a planar distance not exceeding the adjacent radius are retrieved within the same slice; candidate neighbor cells are retained in order of increasing planar distance, not exceeding the maximum number of nearest neighbors; An intra-slice adjacency edge is established between any of the cells and its retained candidate neighbor cells; The cross-slice matching edge is constructed as follows: for two adjacent slices in the physical Z-axis order, in a shared two-dimensional coordinate system, any cell on the first slice is matched with the cell on the second slice that is closest in planar distance; if the planar distance does not exceed the matching threshold, a cross-slice matching edge is established between any cell and its matching cell; the three-dimensional distance of the cross-slice matching edge is determined based on the planar distance and the slice thickness; the intra-slice adjacent edges and the cross-slice matching edges are distinguished by a binary edge type label.

6. The three-dimensional tumor microenvironment biomarker discovery system for gastric cancer according to claim 5, characterized in that: The adjacency radius is 40 micrometers; the maximum number of nearest neighbors is 25; the matching threshold is 10 micrometers; the slice thickness is 4 micrometers; the three-dimensional distance of the cross-slice matching edge satisfies: （1） in, and These are matching cells from two adjacent slices. and These are the two-dimensional coordinates of the two entities. This is the conversion factor from pixels to micrometers. This is the conversion factor from pixel to micrometer. This represents the slice thickness.

7. The three-dimensional tumor microenvironment biomarker discovery system for gastric cancer according to claim 1, characterized in that, The 3D neighborhood and topological feature extraction module includes a neighborhood discovery unit. This neighborhood discovery unit is used to: for each cell, obtain a set of neighbors on the heterogeneous undirected 3D cell graph whose 3D Euclidean distance does not exceed the 3D neighborhood radius, where the 3D neighborhood radius is 50 micrometers; the neighborhood set includes neighbors within the same slice and cross-slice neighbors on adjacent slices; count the number of each cell type in the neighborhood set, including the central cell itself, to obtain the neighborhood composition vector for each cell; perform row normalization on the neighborhood composition vectors of all cells to obtain a normalized neighborhood composition matrix; and perform principal component analysis to reduce the dimensionality of the normalized neighborhood composition matrix to obtain a dimensionality-reduced neighborhood composition matrix. On the reduced-dimensional neighborhood composition matrix, the candidate cluster numbers are traversed, and K-means clustering is used to obtain the cell neighborhood partitioning result for each candidate cluster number; the joint score is obtained by weighted summation of four components: （2） in, The number of candidate clusters; These are the contour coefficient components; The Davidson-Bolding index components, The smaller the value, the better the clustering effect; the negative sign is taken in the joint score calculation. The consistency index component between two sets of cluster labels obtained after performing K-means clustering with two different random seeds; The Z-axis continuous run length component is used to characterize the degree of continuous occurrence of the same cell neighborhood along the physical Z-axis on continuous slices; Let be the weight parameters, and satisfy: （3） The optimal number of cell neighborhoods is to make... Take the maximum value Value, that is The number of candidate clusters ranges from 5 to 20; the weight parameter is set to... , , , .

8. The three-dimensional tumor microenvironment biomarker discovery system for gastric cancer according to claim 1, characterized in that, The multiple three-dimensional topological features are seven families of three-dimensional topological features, including: cell composition features; cell diversity features; three-dimensional Ripley space distribution features; cross-slice cell interaction features; three-dimensional nearest neighbor distance features; cell neighborhood transfer entropy features; and tumor-immune matrix barrier score features. Among them, the cross-slice cell interaction features are obtained by randomly permuting the cell type labels to obtain the interaction Z score, and the number of random permutations is 200. The tumor-immune matrix barrier score features are obtained by calculating the proportion of matrix cells in the shortest path from tumor epithelial cells to immune cells.

9. A three-dimensional tumor microenvironment biomarker discovery system for gastric cancer according to claim 1, characterized in that, The candidate biomarker ranking module is configured to: for each candidate feature, calculate the feature value sequence slice by slice along the physical Z-axis; and calculate the cross-slice coefficient of variation based on the feature value sequence. （4） in, Candidate features The sequence of values ​​along the physical Z-axis. To prevent extremely small constants with a denominator of zero, a stability measure is obtained based on the cross-slice coefficient of variation. （5） For cross-slice cell interaction features, the absolute value of the interaction Z-score is used as the spatial specificity index; for non-interaction features, the absolute deviation of the feature value relative to the global mean is used, divided by the global average absolute deviation, to obtain the spatial specificity index; the spatial specificity index satisfies: （6） in, The global mean of the candidate feature; "global" refers to the statistical set of all values ​​of the feature within all tissue regions corresponding to the entire 3D cell map; the stability measure is multiplied by the spatial specificity index to obtain the comprehensive score: （7） Candidate features are sorted in descending order according to the comprehensive score, and the ranking results of candidate biomarkers are output.

10. A method for discovering three-dimensional tumor microenvironment biomarkers in gastric cancer, characterized in that, The method described is a non-diagnostic in vitro pathological image data processing method, comprising: acquiring a sequence of consecutive hematoxylin-eosin stained whole-section images of the same gastric cancer tissue specimen and section registration information; reconstructing the physical Z-axis order of the whole-section image sequence based on the section registration information; performing tissue mask generation, staining normalization, cell nucleus detection, and cell type assignment on each section to obtain single-cell data containing cell two-dimensional coordinates, physical Z-axis sequence number, and cell type; the cell types include the following nine categories: tumor epithelial cells, normal epithelial cells, lymphocytes, plasma cells, macrophages, neutrophils, fibroblasts, endothelial cells, and unclassified cells whose morphological characteristics do not belong to the aforementioned eight categories; constructing a heterogeneous undirected three-dimensional cell map using cells in the single-cell data as nodes; the heterogeneous undirected three-dimensional cell map includes intra-slice adjacent edges within the same slice, and cross-slice edges between adjacent slices in the physical Z-axis order. The process involves: matching edges between slices; clustering each cell based on its neighborhood composition vector in the heterogeneous undirected 3D cell graph; determining the optimal number of cell neighborhoods using a joint score obtained by weighted fusion of four components: silhouette coefficient, Davidson-Bolding index, double-random seed K-means clustering consistency index, and Z-axis continuous run length; extracting multiple 3D topological features based on the cell neighborhood partitioning corresponding to the optimal number of cell neighborhoods; these features include at least: cross-slice cell interaction features based on Z-scores obtained from cell type label permutation tests, and tumor-immune matrix barrier features characterizing the proportion of stromal cells on the shortest path from tumor epithelial cells to immune cells; calculating cross-slice stability and spatial specificity indices for each candidate feature along the physical Z-axis; determining a comprehensive score based on the cross-slice stability and spatial specificity indices; and outputting the candidate biomarker ranking results according to the comprehensive score.

11. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements all the steps of the method for discovering three-dimensional tumor microenvironment biomarkers for gastric cancer as described in claim 10.

Images