CN-121662310-B - Regulatory T cell infiltration prediction method based on digital pathological section

CN121662310BCN 121662310 BCN121662310 BCN 121662310BCN-121662310-B

Abstract

The application belongs to the technical field of artificial intelligence and digital pathology intersection, and relates to a regulatory T cell infiltration prediction method based on a digital pathological section. The method takes a conventional hematoxylin-eosin staining full-section digital image as input, limits an analysis range through tissue region segmentation and tumor region positioning, extracts multi-scale morphological characteristics from a visual field unit in a tumor region, constructs a local-global spatial context of image neural network coding, and uses adjacent sections On the premise that the immune histochemical quantitative or transcriptome immune deconvolution result is a supervisory signal, regression learning is adopted and anti-regularization and migration learning are combined to improve robustness of cross-staining batch and cross-scanning conditions, regulatory T cell infiltration density or grade is output, thermodynamic diagram visualization is carried out, and uncertainty and spatial heterogeneity indexes can be further given for rechecking and decision support.

Inventors

YANG MEIYAN
ZHANG MINGWEI
HONG JINSHENG
WANG ZHENGYANG
CHEN JIALEI
XIAO JIAYU
Guo Chenyan

Assignees

福建医科大学

Dates

Publication Date: 20260512
Application Date: 20260209

Claims (6)

1. A regulatory T cell infiltration prediction method based on a digitized pathological section is characterized by comprising the steps of obtaining a conventional hematoxylin-eosin staining full-section digitized image, wherein the pixel size is not larger than a preset pixel size threshold, a color space is RGB three channels, the image size is in a preset pixel range, carrying out tissue region segmentation and tumor region positioning on the full-section image, adopting a semantic segmentation model based on a U-shaped convolutional neural network, carrying out pixel-level classification on a tissue region, a background region and a tumor parenchymal region in the image, outputting a tumor region mask map, extracting a multiscale morphological feature map in the tumor region, dividing the tumor region into a plurality of non-overlapping visual field units, each visual field unit has a preset size, respectively extracting nuclear morphology, chromatin distribution and intercellular arrangement mode features under low scale, medium scale and high scale conditions through a cascade convolution module, constructing a local-global space context encoder, adopting a graph neural network architecture, adopting a graph node feature dimension as a preset value, establishing an adjacent relation according to space coordinates, setting a global position of the adjacent relation, adopting a global position scaling factor, calculating a full-scale threshold, adopting a global position scaling factor, and a full-scale channel scaling threshold, calculating a full-scale channel feature map, and a full-scale feature pool, and using a global position scaling factor, and a full-scale feature pool, the method comprises the steps of taking a context embedded vector as input, predicting an infiltration intensity value of regulatory T cells through a fully connected regression head, wherein the infiltration intensity value is used for representing the infiltration degree of the regulatory T cells in a tumor area of a unit area, when adjacent slice immunohistochemical quantification is adopted as a supervision signal, the infiltration intensity value is defined as an infiltration density value, the infiltration density value is defined as the equivalent number of FOXP3 positive cells in the tumor area of the unit area, when a transcriptome immunodeconvolution result is adopted as the supervision signal, the infiltration intensity value is the infiltration proportion or score of the regulatory T cells obtained by immunodeconvolution, model training adopts a mean square error loss function, an anti-regularization term is introduced to improve the robustness of HE staining variation, the anti-regularization term is realized by introducing a discriminator network, the discriminator is used for discriminating HE image batches from different hospitals or different scanners, the prediction model of the regulatory T cells needs to minimize the prediction loss in the training process, meanwhile, the degree of the discriminator is maximized, the infiltration level of the regulatory T cells is output, the infiltration intensity value is divided according to the infiltration intensity value, and the infiltration intensity value is superimposed on the original slice images in a mixed-up mode.
2. The method of claim 1, wherein the U-shaped convolutional neural network comprises a plurality of downsampling stages and a corresponding number of upsampling stages, each stage uses convolution kernels of a predetermined size, the activation function is a modified linear unit, and the loss function uses weighted cross entropy, wherein the weighting coefficients of the tumor region are higher than the weighting coefficients of the background region and the normal tissue region.
3. The method for predicting the infiltration of regulatory T cells based on digitized pathological sections of claim 1, wherein the cascade convolution module is composed of a plurality of parallel branches, each branch corresponds to a low scale, a middle scale and a high scale, each branch comprises at least two layers of convolution layers, the number of convolution kernels is sequentially increased, and outputs of the branches are spliced and fused through channels to form a multi-scale feature vector with a preset dimension.
4. The method of claim 1, wherein the model of T-cell infiltration prediction uses a migration learning strategy during the training phase, and the initial weights are derived from a visual transducer backbone pre-trained on a large-scale public pathology image dataset.
5. The method for predicting the infiltration of the regulatory T cells based on the digitized pathological section according to claim 1 is characterized in that the thermodynamic diagram is generated by mapping the infiltration intensity value back to a full-section coordinate system by bilinear interpolation, the color mapping is of blue-white-red gradient color, the thermodynamic diagram transparency is set to be a preset transparency value, the underlying tissue structure can still be clearly recognized, meanwhile, the standard deviation of the infiltration intensity value is calculated by forward propagation of a Monte Carlo Dropout method for a plurality of times in an inference stage, and when the standard deviation is larger than a preset standard deviation threshold value, the system automatically marks the area as a low confidence level area.
6. The method for predicting the infiltration of the regulatory T cells based on the digitized pathological section of claim 1, wherein the level of infiltration of the regulatory T cells, the index of spatial distribution heterogeneity and the index of uncertainty are packed into structured data, and pushed to an electronic medical record system of a hospital through a standard medical communication protocol, so as to provide decision support reference information related to the treatment of the immune checkpoint inhibitor for oncologists.

Description

Regulatory T cell infiltration prediction method based on digital pathological section Technical Field The application belongs to the technical field of artificial intelligence and digital pathology intersection, and particularly relates to a regulatory T cell infiltration prediction method based on a digital pathological section. Background With the continuous advancement of tumor immunotherapy, the infiltration level of immunosuppressive cells such as regulatory T cells (tregs) in tumor microenvironments has become a key biomarker for evaluating therapeutic response, judging prognosis and formulating personalized immune intervention strategies. The traditional detection means mainly depend on multiple immunohistochemistry or flow cytometry, and has the problems of high cost, limited flux, complex operation and the like although the detection means have higher specificity, and the detection means are difficult to integrate into the conventional pathological diagnosis flow. In recent years, an artificial intelligence analysis method based on digital pathological sections provides a new path for label-free prediction of specific immune cell infiltration from widely available hematoxylin-eosin (HE) staining images, and the accessibility of technology and clinical transformation potential are remarkably improved. However, existing deep learning-based pathology image analysis methods still face serious challenges in modeling a specific immune subpopulation of regulatory T cells. On the one hand, regulatory T cells lack definite morphological identification under conventional HE staining, their nuclei are highly similar to common lymphocytes, resulting in models that are difficult to distinguish efficiently by visual features of a single scale, on the other hand, most current methods focus on the global recognition of generalized tumor infiltrating lymphocytes, not aimed atThe Treg and other sub-populations with immunosuppressive functions are specially optimized, and the modeling capability of the interaction relation between the spatial distribution mode and the microenvironment is lacking. In addition, although the accurate quantification of cell subtypes is realized by means of multiple fluorescence immunohistochemistry according to a part of high-precision schemes, the workflow of a conventional pathology department mainly comprising HE (human-animal-derived) staining cannot be adapted due to the dependence on a special staining technology, and the popularization and the application of the workflow in basic medical institutions are severely restricted. Therefore, a calculation method for predicting the infiltration level of regulatory T cells only by relying on conventional HE (human-animal) staining full-section images in an inference stage is needed, and in the method, supervision signals can be constructed by utilizing adjacent section immunohistochemical quantitative or transcriptome immune deconvolution results in a training stage, and a deep learning characterization system for Treg is constructed by fusing multi-scale morphological clues and local-global spatial context information, so that a technical support capable of being popularized and low in cost is provided for quantitative analysis of tumor immune microenvironment while clinical compatibility is maintained. Disclosure of Invention The invention aims to provide a method for predicting the infiltration of regulatory T cells based on a digital pathological section, which can effectively solve the problems in the background technology. In order to achieve the above purpose, the technical scheme adopted by the invention is as follows: The method for predicting the infiltration of the regulatory T cells based on the digitized pathological section comprises the following specific steps that step (1) a conventional hematoxylin-eosin staining full-section digitized image is obtained, wherein the full-section image digitized by a scanner is called from a pathology department information system, the pixel size of the image is not more than a preset pixel size threshold, the color space is RGB three channels, and the image size is within a preset pixel range; the method comprises the steps of (1) carrying out tissue region segmentation and tumor region positioning on a full-slice image, namely carrying out pixel-level classification on a tissue region, a background region and a tumor parenchymal region in the image by adopting a semantic segmentation model based on a U-shaped convolutional neural network, outputting a tumor region mask map, extracting a multiscale morphological feature map in the tumor region, namely dividing the tumor region into a plurality of non-overlapping visual field units, each visual field unit has a preset size, respectively extracting morphological characterization features at a low scale, a middle scale and a high scale by a cascade convolution module to generate a multiscale morphological feature tensor, constructing a loc