A method for predicting regulatory t cell infiltration based on digitized pathology sections

Abstract

The application belongs to the technical field of artificial intelligence and digital pathology, and relates to a regulatory T cell infiltration prediction method based on digital pathological sections. The method takes a conventional hematoxylin-eosin staining whole section digital image as input, limits the analysis range through tissue region segmentation and tumor region positioning, extracts multi-scale morphological features on a field unit in the tumor region, and constructs a graph neural network to encode local-global spatial context. Under the premise of taking adjacent section immunohistochemical quantification or transcriptome immunodeconvolution results as a supervision signal, regression learning is adopted, combined with adversarial regularization and transfer learning to improve the robustness of cross-staining batches and cross-scanning conditions. The output is the regulatory T cell infiltration density or grade and a heat map visualization, and further uncertainty and spatial heterogeneity indicators can be given for review and decision support.

Description

Technical Field

[0001] This application belongs to the field of interdisciplinary technology of artificial intelligence and digital pathology, specifically involving a method for predicting regulatory T cell infiltration based on digital pathological slides. Background Technology

[0002] With the continuous advancement of tumor immunotherapy, the infiltration level of immunosuppressive cells such as regulatory T cells (Tregs) in the tumor microenvironment has become a key biomarker for assessing treatment response, predicting prognosis, and developing personalized immune intervention strategies. Traditional detection methods mainly rely on multiplex immunohistochemistry or flow cytometry, which, while possessing high specificity, suffer from high costs, limited throughput, and complex operation, making them difficult to integrate into routine pathological diagnostic procedures. In recent years, artificial intelligence analysis methods based on digital pathological slides have provided a new pathway for label-free prediction of specific immune cell infiltration from widely available hematoxylin-eosin (HE) stained images, significantly improving the accessibility of the technology and its clinical translational potential.

[0003] However, existing deep learning-based pathological image analysis methods still face significant challenges in modeling the specific immune subset of regulatory T cells. On the one hand, regulatory T cells lack clear morphological markers under conventional HE staining, and their nuclei are highly similar to ordinary lymphocytes, making it difficult for models to effectively distinguish them using visual features at a single scale. On the other hand, most current methods focus on the overall identification of broadly defined tumor-infiltrating lymphocytes, without addressing specific needs. While specific optimizations have been made for immunosuppressive subpopulations such as Treg cells, there is a lack of modeling capabilities for their spatial distribution patterns and interactions with the microenvironment. Furthermore, although some high-precision protocols utilize multiplex fluorescence immunohistochemistry to achieve accurate quantification of cell subtypes, their reliance on specialized staining techniques makes them unsuitable for routine pathology workflows, which primarily rely on HE staining, severely hindering their widespread application in primary healthcare institutions.

[0004] Therefore, there is an urgent need for a computational method that can predict the level of regulatory T cell infiltration using only conventional HE-stained whole-slice images during the inference phase. This method can construct a supervision signal during the training phase using quantitative immunohistochemical results from adjacent slices or transcriptomic immunodeconvolution results. By fusing multi-scale morphological cues with local-global spatial context information, a deep learning representation system for Tregs can be constructed. This provides a scalable and low-cost technical support for the quantitative analysis of the tumor immune microenvironment while maintaining clinical compatibility. Summary of the Invention

[0005] The purpose of this invention is to provide a method for predicting regulatory T cell infiltration based on digital pathological sections, which can effectively solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0007] A method for predicting regulatory T cell infiltration based on digital pathological slides includes the following specific steps: Step (1) Obtaining digital images of whole slides stained with conventional hematoxylin and eosin: Retrieving the whole slide images digitized by the scanner from the pathology department information system. The image pixel size is not greater than the preset pixel size threshold, the color space is RGB three channels, and the image size is within the predetermined pixel range; Step (2) Performing tissue region segmentation and tumor region localization on the whole slide images: Using a semantic segmentation model based on a U-shaped convolutional neural network, pixel-level classification is performed on the tissue region, background region, and tumor parenchyma region in the image, and a tumor region mask map is output; Step (3) Extracting multi-scale morphological feature maps within the tumor region: The tumor region is divided into multiple non-overlapping visual units, each with a predetermined size. Morphological features at low, medium, and high scales are extracted using cascaded convolution modules to generate multi-scale morphological feature tensors. (4) Constructing a local-global spatial context encoder: Based on a graph neural network architecture, each visual unit is treated as a graph node. Adjacency relationships are established based on its spatial coordinates. The attention mechanism is used to aggregate the multi-scale features of nodes in the local neighborhood. The spatial distribution information of the entire tumor region is fused through global pooling operations to output a context encoder with spatial awareness. Embedded vector; Step (5) Train and deploy the regulatory T cell infiltration prediction model: Using the context embedding vector as input, predict the regulatory T cell infiltration intensity value through a fully connected regression head; Construct supervision labels during the training phase: First, obtain adjacent tissue slices corresponding to the HE slices and perform FOXP3 immunohistochemical staining, scan to obtain FOXP3 digital images, and use coarse-to-fine cross-slice registration (e.g., first perform global affine registration, then perform local non-rigid registration) to project the immunohistochemical results onto the HE coordinate system, and then detect and count FOXP3 positive cells (the image processing flow of color decomposition and threshold segmentation can be used, or the method of color decomposition and threshold segmentation can be used). The cell nucleus detection network based on deep learning is used to calculate the density of FOXP3 positive cells in a unit area of ​​tumor as the regression true value; secondly, when matching transcriptome data is available, the immune deconvolution method can be used to estimate the proportion or score of regulatory T cell infiltration as a weak supervision signal; the model training uses the mean squared error loss function and introduces an adversarial regularization term to improve the robustness to HE staining variations; step (6) outputs the evaluation results of regulatory T cell infiltration level: the infiltration level is divided according to the predicted infiltration intensity value, including low infiltration, medium infiltration and high infiltration, and the results are superimposed on the original whole slice image in the form of a heat map for pathologists to review and refer to.

[0008] Preferably, the U-shaped convolutional neural network in step (2) may include Each downsampling stage and Each upsampling stage has a convolutional kernel size of [size missing]. The activation function is ReLU, and the loss function is weighted cross-entropy, where the weight coefficient for the tumor region can be set to... .

[0009] Preferably, in step (3), the cascaded convolution module consists of multiple parallel branches, corresponding to the low, medium and high scales respectively. Each branch contains several convolutional layers, and the number of convolutional kernels increases sequentially. The outputs of each branch are spliced ​​and fused through channels to form a multi-scale feature vector with a predetermined dimension.

[0010] Preferably, in step (4), the node feature dimension of the graph neural network is a predetermined value, the adjacency matrix is ​​constructed based on the Euclidean distance between the center points of the vision units, the distance threshold is set to a preset distance threshold, the attention weight calculation adopts the scaling dot product mechanism, and the global pooling adopts a combination of average pooling and max pooling.

[0011] Preferably, the adversarial regularization term in step (5) is implemented by introducing a discriminator network, which is used to distinguish HE image batches from different hospitals or different scanners. During the training process, the backbone prediction model needs to minimize the prediction loss while maximizing the confusion of the discriminator, thereby decoupling the staining style from the biological signal.

[0012] Preferably, the regulatory T cell infiltration prediction model can employ a transfer learning strategy during the training phase, with initial weights derived from a backbone network pre-trained on large-scale publicly available pathological image data, and only the final weights fine-tuned. A Transformer block and regression header parameters.

[0013] Preferably, in step (6), the heatmap generation uses bilinear interpolation to map the predicted density values ​​back to the full slice coordinate system, the color mapping uses a blue-white-red gradient, and the heatmap transparency is set to... This ensures that the underlying organizational structure remains clearly identifiable.

[0014] Preferably, the present invention further includes uncertainty quantification: maintaining the activation of the Dropout layer during the inference phase using the Monte Carlo Dropout method, and performing uncertainty quantification on the same input. The forward propagation was obtained Calculate the standard deviation ;when Greater than the threshold When this happens, the area is marked as a low-confidence area.

[0015] Preferably, the present invention also includes integration with a clinical decision support system: the regulatory T cell infiltration level, spatial distribution heterogeneity index and uncertainty indicators are packaged into structured data and pushed to the hospital's electronic medical record system through a standard medical communication protocol, providing oncologists with decision support reference information related to immune checkpoint inhibitor therapy.

[0016] Preferably, the method supports batch processing mode, and can achieve minute-level processing speed through pyramid resolution reading, batch processing and GPU parallel acceleration; the number of concurrent processing cases and the average processing time can be set according to the full slice size, the number of field units and hardware configuration to meet the routine workload requirements of the pathology department.

[0017] Compared with the prior art, the present invention has the following beneficial effects:

[0018] 1. Label-free prediction capability for Tregs. By fusing multi-scale morphological features and spatial contextual information, the model of this invention can learn microenvironment association patterns related to regulatory T cell infiltration, rather than relying solely on single cell morphology; in some embodiments, it achieves better predictive performance compared to baseline methods based solely on generalized lymphocyte segmentation or simple statistical features.

[0019] Inference assessment can be achieved without special staining: During the model deployment / inference phase, only routine HE-stained whole-slice images are needed to output infiltration density, grading, and thermograms, which can reduce the cost and process complexity of additional immunostaining or multiple labeling, making it easier to integrate regulatory T cell infiltration assessment into existing pathology workflows.

[0020] 2. Powerful spatial modeling and clinical compatibility. Finely characterizing the spatial heterogeneity of the immune microenvironment: The graph neural network-driven local-global context encoder can model the distribution differences of regulatory T cells in different regions such as the tumor core, the invasion front, and tertiary lymphoid structures, providing richer spatial information than the global average density, offering a new dimension for predicting immunotherapy responses;

[0021] High robustness and cross-platform adaptability: Through adversarial regularization and transfer learning strategies, the model has a stronger ability to adapt to image variations from different scanners, staining batches and hospital sources; its generalization ability can be quantitatively evaluated through independent validation sets and external cohorts, and combined with uncertainty quantification mechanisms to indicate areas that need to be reviewed.

[0022] 3. Intelligent and efficient integration capabilities. Automated processing and verifiable output: The entire process from image input to result output is automated, and uncertainty indicators highlight areas requiring focused verification; it supports batch concurrent and structured data output, facilitating integrated deployment in large medical centers or regional pathology centers.

[0023] Clinical decision-making closed-loop support: By connecting structured data output with the electronic medical record system, it provides oncologists with decision support reference information related to immunotherapy, shortens the time window from pathological assessment to treatment decision, and promotes the standardized implementation of individualized immunotherapy. Attached Figure Description

[0024] Figure 1 This is a schematic diagram of the overall technical solution architecture of a regulatory T cell infiltration prediction method based on digital pathological sections proposed in this invention.

[0025] Figure 2 This is a schematic diagram of the distribution of the training / validation dataset sources and the distribution of regulatory T cell infiltration labels in this invention. (A) is a schematic diagram of the distribution of the number of samples from different sources, and (B) is a schematic diagram of the distribution of regulatory T cell infiltration intensity (e.g., FOXP3 positive cell density). Relevant example statistics can be found in Table 1.

[0026] Figure 3 This is a schematic diagram of the regression prediction results and error distribution of the infiltration intensity of the present invention, wherein (A) is a scatter plot of the comparison between the predicted value and the actual value, and (B) is a schematic diagram of the prediction error distribution and statistics; relevant example statistical data can be found in Table 2.

[0027] Figure 4 This is a schematic diagram of the ablation experiment and cross-domain robustness assessment of the key module of the present invention, wherein (A) is a schematic diagram of the control experiment on the performance impact of the key module, and (B) is a schematic diagram of the robustness assessment under cross-center / cross-scan conditions.

[0028] Figure 5 This is a flowchart illustrating the logical flow of the generation of regulatory T cell infiltration level assessment results and the overlay of a heatmap in this invention.

[0029] Figure 6 This is a schematic diagram of the generalization performance evaluation of the present invention in an independent dataset or external queue, wherein (A) can show the error distribution under different centers / different scanning conditions, and (B) can show the performance comparison with the baseline method.

[0030] Figure 7 This is an example image of common histological region categories in digital pathology whole slides, including fat (ADI), background (BACK), debris / necrosis (DEB), lymphoid tissue (LYM), mucus (MUC), muscle (MUS), normal epithelium (NORM), stroma (STR), and tumor (TUM), etc.

[0031] Figure 8This is an example diagram of immunohistochemical / multiplex immunofluorescence staining of regulatory T cell-related markers to illustrate the feasible implementation of surveillance signal construction and positive cell counting, exemplarily showing channels such as Hematoxylin, CD3, CD8, FOXP3, and PanCK. Detailed Implementation

[0032] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments.

[0033] Currently, with the in-depth development of tumor immunotherapy, the infiltration degree of immunosuppressive cells such as regulatory T cells (Tregs) in the tumor microenvironment has become an important biomarker for assessing immunotherapy response, prognosis, and the formulation of personalized treatment strategies. Traditionally, the detection of regulatory T cells relies on multiplex immunohistochemistry or flow cytometry, which suffers from high cost, low throughput, and difficulty in widespread adoption in routine pathology procedures. In recent years, artificial intelligence analysis methods based on digital pathology slides have provided new insights into predicting the infiltration of specific immune cells from routine hematoxylin-eosin (HE) staining images. However, existing technologies still have significant limitations in the accurate identification and spatial distribution modeling of the specific immune subset of regulatory T cells.

[0034] A search revealed a patent, CN116469513B, entitled "Deep Learning-Based Personalized Treatment and Prognostic Information Prediction System for Colorectal Cancer," published on November 14, 2023. This patent utilizes deep learning to finely segment and identify tumor-infiltrating lymphocytes in routine pathological sections, and combines this with whole-exome sequencing to construct a recurrence risk prediction model. However, while this technical solution mentions "tumor-infiltrating lymphocytes," it does not specifically distinguish regulatory T cells (such as...). Treg cells (Tregs) and other effector or memory T cell subsets are trained on a broad lymphocyte region, lacking specific modeling of regulatory T cell phenotypic characteristics. Furthermore, they rely on conventional HE staining images, but regulatory T cells exhibit atypical morphological features under HE staining, making them difficult to distinguish from ordinary lymphocytes. This limits the model's predictive specificity and sensitivity for Treg infiltration, failing to meet the needs for precise assessment of the immune microenvironment.

[0035] A search revealed the publication CN120089385A, entitled "A Diagnostic Model for Predicting the Prognosis of CAR-T Cell Immunotherapy in Gastric Cancer and Its Application." This method integrates multiplex fluorescence immunohistochemistry (mIHC) with whole-slice digital imaging to analyze the tumor microenvironment components (including...) of CLDN18.2-positive gastric cancer samples. Treg et al. were used to quantify and spatially analyze the distribution of immune cells and establish prognostic prediction models. While they can accurately quantify specific immune cell subsets, their core workflow relies on multiplex fluorescent labeling / immunostaining platforms, making them difficult to directly adapt to pathological workflows primarily based on conventional HE staining. Furthermore, they are costly and complex to operate. Examples of FOXP3 immunohistochemistry and multiplex immunolabeling channels can be found in [link to relevant documentation]. Figure 8 .

[0036] The aforementioned problems indicate that existing technologies, either lacking the ability to specifically model regulatory T cells or relying on special staining methods, are difficult to apply in routine pathology scenarios. Therefore, they struggle to achieve stable and generalizable prediction of regulatory T cell infiltration levels in digital pathology workflows primarily based on conventional HE staining. This invention provides a method for predicting regulatory T cell infiltration based on digital pathological sections. It aims to utilize only conventional HE-stained whole-section images during the inference phase, constructing a deep learning model by fusing multi-scale morphological features and spatial contextual information to achieve label-free prediction of regulatory T cell infiltration levels. This provides low-cost, highly accessible technical support for tumor immune microenvironment assessment and personalized immunotherapy decision-making. To address the aforementioned technical problems, this invention proposes a method for predicting regulatory T cell infiltration based on digital pathological sections.

[0037] In the aforementioned method for predicting regulatory T cell infiltration based on digital pathological sections, step (1) includes acquiring a digital image of a whole section stained with conventional hematoxylin and eosin. Preferably, the pixel size of the whole section image is no larger than […]. The color space is RGB three-channel, and the image size is usually located in... to Pixels. After reading the image, the system performs data integrity and quality control, including header verification, focus blur detection, background ratio threshold screening, and obvious scanning artifact detection; color normalization is performed on the staining style when necessary. To train the infiltration prediction model, supervision signals (annotations) need to be constructed: firstly, adjacent slices of the same wax block are... Immunohistochemical staining and digitization were performed, and cross-slice registration was executed to project the positive cell detection results onto the HE coordinate system to obtain the true infiltration density of each field of view. Secondly, Treg infiltration levels are estimated using matched RNA sequencing data via an immune deconvolution method, serving as weak supervision tags. Subsequent embodiments cite statistical data from publicly available cohort literature (see Tables 1-3) to illustrate the existence of verifiable validation results in the publicly available data, but do not constitute a limitation on the effectiveness of this invention.

[0038] In the above method, step (2) involves tissue region segmentation and tumor region localization. Specifically, a pre-trained U-shaped convolutional neural network (U-Net) semantic segmentation model is used. This U-Net model includes four downsampling stages and four corresponding upsampling stages. In each downsampling stage, the model uses 2×2 max pooling to halve the feature map size and extracts features through two consecutive 3×3 convolutional kernels; in each upsampling stage, the model doubles the feature map size through transposed convolution and performs skip connections with the feature map of the corresponding level of the encoder before performing 3×3 convolution reconstruction. The model outputs a mask map of the same size as the input, and each pixel is classified into background region, normal tissue region, or tumor solid region. To improve the accuracy of tumor region segmentation, weighted cross-entropy can be used as the loss function, and the tumor region weight coefficient can be set. .

[0039] In the above method, step (3) involves extracting a multi-scale morphological feature map within the tumor region. Specifically, based on the tumor region mask map output in step (2), the tumor region is divided into multiple non-overlapping visual field units, with each visual field unit having a fixed size. Pixels; for the remaining parts of a complete unit with insufficient edges, zero padding can be performed to ensure consistent input size. Subsequently, each receptive unit is fed into a cascaded convolutional module to extract features. This module consists of... It consists of several parallel branches, each corresponding to a low-scale... Mesoscale With high scale Three levels of analysis. Each branch contains Convolutional layers, with the number of convolutional kernels as follows: , , Each branch output is upsampled to... After unifying the spatial resolution, channel stitching and fusion are performed to finally obtain the multi-scale feature vector dimensions. .

[0040] In the above method, step (4) involves constructing a local-global spatial context encoder. This step is implemented based on a graph neural network (GNN): each visual unit is considered as a node in the graph, and its multi-scale feature vector is used as the initial feature of the node. Adjacency relationships are established based on the spatial coordinates of the nodes in the original whole-slice image: the Euclidean distance between the center points of any two nodes is calculated, and when the distance does not exceed a preset threshold... Undirected edges are established at the time. During message passing, scaled dot product attention is used to aggregate neighborhood features, and the attention weight coefficients are defined as follows:

[0041]

[0042] in:

[0043]

[0044]

[0045]

[0046] in, For nodes For nodes Attention weights For nodes The neighborhood set, For the dimension of the attention subspace; For nodes The input feature vector, This is the aggregated output feature vector; This is the learnable parameter matrix.

[0047] When the Euclidean distance of the node center point satisfy Connect the edges, among which This is a preset distance threshold.

[0048] Optionally, the attention mechanism can be extended to a multi-head form, and the final representation can be obtained by splicing or summing the heads to enhance the ability to characterize different spatial patterns.

[0049] Through the above mechanism, the model can dynamically learn the importance of different neighboring nodes to the central node.

[0050] After multiple rounds of message passing, a global pooling operation is performed to fuse the overall spatial distribution information of the entire tumor region. The global average pooling vector is... The global max-pooling vector is The context embedding vector is defined as follows: ,in It is a set of nodes.

[0051] In the above method, step (5) involves training and deploying a regulatory T cell infiltration prediction model. The context embedding vector output from step (4) is used as the basis for this process. As input, the regression head outputs the infiltration prediction value. .in The label can be generated from adjacent slices. The results were obtained from immunohistochemical quantification or transcriptomic immunodeconvolution. During training, mean squared error was used as the main loss, and adversarial regularization was introduced to reduce the impact of differences in staining style and scanning conditions on feature representation. At the same time, a transfer learning strategy was adopted, using a backbone network pre-trained on large-scale public pathological image data as initialization, and only fine-tuning some high-level parameters to improve convergence speed and generalization ability.

[0052]

[0053]

[0054] in, The number of training samples, These are the weighting coefficients; The discriminator cross-entropy loss can be used, and the backbone model can be made insensitive to coloring / device domain information through gradient reversal or adversarial training.

[0055] In the above method, step (6) is to output the assessment results of the regulatory T cell infiltration level. First, based on the predicted value... And based on preset thresholds, the infiltration is divided into low, medium, and high levels; then... Map back to the full slice coordinate system and generate a heatmap overlay display. The heatmap transparency can be set to [value missing]. Furthermore, Monte Carlo Dropout can be used to perform [further processing] on the same input during the inference phase. The second forward propagation calculates the mean and standard deviation to quantify uncertainty:

[0056]

[0057]

[0058] when When the threshold is exceeded, the corresponding area is marked as a low-confidence area to prompt for focused review. Furthermore, spatial heterogeneity indices, such as the coefficient of variation, can be used to represent the degree of unevenness in infiltration distribution. .

[0059] Example (Example of Publicly Available Cohort Data): To avoid introducing unverifiable fictitious values ​​into the specification, the following uses statistical data from publicly available preprint studies (see Non-Patent Document 1) to illustrate the existence of verifiable comparative results in the publicly available pathological data; this example is for illustrative purposes only and does not constitute a limitation on the implementation effect of the present invention. This example selects IDH wild-type glioblastoma patients from the TCGA cohort. ,according to Proportional random partitioning of the training set With the validation set The cells were divided into two groups based on the level of regulatory T cell infiltration: low-infiltration and high-infiltration (see Table 1). In this publicly available example, ROC analysis showed that the training set AUC was [value missing]. The validation set AUC is The Hosmer-Lemeshow goodness-of-fit test p-values ​​were respectively and (See Table 2). Furthermore, publicly available research provides examples of risk stratification associated with infiltration: using scoring thresholds. Will Divided into high ratings With low rating Two groups, median survival time in the high-scoring group Months, low-scoring group Significant differences over months In Cox regression, higher scores were associated with worse overall survival. Figures 2-6 The accompanying drawings are for illustrating the process and exemplary results of this invention and do not need to correspond one-to-one with the figures in the cited preprints.

[0060] Table 1. Training / Validation Division and Clinicopathological Features Statistics ( )

[0061]

[0062] Table 2 Predictive Performance Statistics

[0063]

[0064] Table 3. Prognostic correlation statistics for public cohort examples (Cox regression, scoring threshold 0.299)

[0065]

[0066] Based on the foregoing embodiments, the present invention also provides an optional lightweight implementation scheme to adapt to deployment scenarios with limited computing resources. This scheme, while maintaining consistency between the inputs and feature extraction in steps (1) to (3), replaces the graph modeling in step (4) with lightweight spatial encoding: reshaping the multi-scale morphological feature tensor into a two-dimensional feature map. It employs multi-layer dilated convolution for long-range dependency modeling, and the dilation rate can be calculated according to... The settings are configured step-by-step; then global average pooling is performed to obtain the context embedding vector, which is then fed into the regression head. This alternative approach can reduce memory usage and inference latency; its specific performance can be evaluated based on the target hospital's equipment conditions and external validation data.

[0067] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

[0068] Non-patent literature 1: Research Square preprint (DOI: 10.21203 / rs.3.rs-3274237 / v1).

Claims

1. A method for predicting regulatory T cell infiltration based on digital pathological sections, characterized in that, Includes the following steps: Obtain digital images of whole sections stained with hematoxylin and eosin in a conventional manner, with pixel size not exceeding a preset pixel size threshold, color space of RGB three channels, and image size within a predetermined pixel range; Tissue region segmentation and tumor region localization are performed on the whole slice image. A semantic segmentation model based on U-shaped convolutional neural network is used to classify the tissue region, background region and tumor parenchyma region in the image at the pixel level and output the tumor region mask map. Multi-scale morphological feature maps are extracted in the tumor region, and the tumor region is divided into multiple non-overlapping visual field units. Each visual field unit has a predetermined size. The features of cell nuclear morphology, chromatin distribution and intercellular arrangement pattern at low scale, medium scale and high scale are extracted by cascaded convolution module to generate multi-scale morphological feature tensors. A local-global spatial context encoder is constructed based on a graph neural network architecture. Each field of view (PVR) is treated as a graph node with a predetermined node feature dimension. Adjacency relationships are established based on their spatial coordinates. The adjacency matrix is ​​constructed based on the Euclidean distance between the center points of PVRs, and a preset distance threshold is set. An attention mechanism is used to aggregate multi-scale features of nodes within the local neighborhood. The attention weights are calculated using a scaled dot product mechanism. Global pooling is used to fuse spatial distribution information across the entire tumor region. Global pooling employs a combination of average pooling and max pooling to output a spatially aware context embedding vector. A regulatory T cell invasion prediction model is trained and deployed. Using the context embedding vector as input, a fully connected regression head predicts the regulatory T cell invasion intensity value. This invasion intensity value characterizes the degree of regulatory T cell invasion per unit area of ​​the tumor region. Specifically, when using adjacent slices… When immunohistochemical quantification is used as a supervisory signal, the infiltration intensity value is the infiltration density value, and the infiltration density value is defined as the equivalent number of FOXP3-positive cells per unit area of ​​tumor region. When transcriptomic immunodeconvolution results are used as a supervisory signal, the infiltration intensity value is the proportion or score of regulatory T cell infiltration obtained by immunodeconvolution. The model training uses a mean squared error loss function and introduces an adversarial regularization term to improve robustness to HE staining variations. The adversarial regularization term is implemented by introducing a discriminator network, which is used to distinguish HE image batches from different hospitals or different scanners. The regulatory T cell infiltration prediction model needs to minimize the prediction loss while maximizing the discriminator's confusion level during training. The output is the regulatory T cell infiltration level assessment result, which is divided into infiltration grades according to the infiltration intensity value, and the result is superimposed on the original whole slice image in the form of a heatmap.

2. The method for predicting regulatory T cell infiltration based on digital pathological sections according to claim 1, characterized in that, The U-shaped convolutional neural network includes multiple downsampling stages and a corresponding number of upsampling stages. Each stage uses a convolutional kernel of a predetermined size, the activation function is a modified linear unit, and the loss function is a weighted cross-entropy, where the weight coefficient of the tumor region is higher than that of the background region and the normal tissue region.

3. The method for predicting regulatory T cell infiltration based on digital pathological sections according to claim 1, characterized in that, The cascaded convolution module consists of multiple parallel branches, corresponding to low-scale, medium-scale and high-scale respectively. Each branch contains at least two convolutional layers, and the number of convolutional kernels increases sequentially. The outputs of each branch are spliced ​​and fused through channels to form a multi-scale feature vector with a predetermined dimension.

4. The method for predicting regulatory T cell infiltration based on digital pathological sections according to claim 1, characterized in that, The regulatory T cell infiltration prediction model employs a transfer learning strategy during the training phase, with initial weights derived from a visual Transformer backbone network pre-trained on a large-scale public pathological image dataset.

5. The method for predicting regulatory T cell infiltration based on digital pathological sections according to claim 1, characterized in that, The heatmap is generated by bilinear interpolation to map the infiltration intensity value back to the full slice coordinate system. The color mapping uses a blue-white-red gradient, and the heatmap transparency is set to a preset transparency value to ensure that the underlying tissue structure can still be clearly identified. At the same time, the standard deviation of the infiltration intensity value is calculated by multiple forward propagations during the inference phase using the Monte Carlo Dropout method. When the standard deviation is greater than the preset standard deviation threshold, the system automatically marks the area as a low confidence area.

6. The method for predicting regulatory T cell infiltration based on digital pathological sections according to claim 1, characterized in that, The regulatory T cell infiltration level, spatial distribution heterogeneity index, and uncertainty indicators are packaged into structured data and pushed to the hospital's electronic medical record system through standard medical communication protocols, providing oncologists with decision support reference information related to immune checkpoint inhibitor therapy.

Images