Risk Prediction Of Gastric Cancer Tumor Mutational Burden Based On Pathological Images And Deep Learning

Gastric cancer is one of the most common cancers in the world.In the latest cancer ranking,gastric cancer ranks fifth in incidence rate and fourth in mortality.In recent years,immunotherapy based on immune checkpoint inhibitors(ICI)has played an important role in the treatment of gastric cancer.Research has found that tumor mutational burden(TMB)characterized by the number of coding errors in somatic cells is beneficial for immunotherapy,and the gold standard for measuring TMB is whole exome sequencing(WES).Due to the high cost,complex operational procedures,and long waiting times of WES testing,it is difficult to apply it to clinical treatment.It is urgent to develop a simpler method for evaluating TMB.This article developed a multimodal deep learning model framework called MDLPGC,which combines hematoxylin and eosin(H&E)stained pathological images with clinical and omics features to predict the risk of TMB in gastric cancer.The framework is a deep learning classification model based on residual networks.The process mainly includes data collection,tumor tissue region labeling,image normalization,training,and testing of the model.Finally,the evaluation of the model performance.Specifically,this article first downloaded the required H&E digital pathological images of gastric cancer and corresponding clinical information and omics data from the Cancer Genome Atlas(TCGA)website.Secondly,pathological experts annotate the downloaded pathological images with tumor regions.Thirdly,use the Open slide software package to segment the labeled pathological images,use the Open CV software package to reduce noise,and use the Macenko algorithm to normalize the color of pathological images.Then,the normalized pathological images are input into the model for training and testing.Finally,the classification prediction model was evaluated using acceptable prediction area under curve(AUC),accuracy rate(ACC),accuracy,recall rate,and F1 score.In addition,we also conducted a differential analysis of the correlation between TMB and clinical,omics,and H&E pathological image features.In 5-fold cross validation,using H&E pathological images alone to predict TMB in gastric cancer,the acceptable predictive area under curve(AUC)was 0.749.Compared to using H&E pathological images alone,establishing a multimodal model that fused omics features of H&E pathological images showed better predictive performance,reaching a maximum AUC of 0.971 when combined with H&E pathological images and m RNA data.Therefore,combining H&E pathological images and deep learning can predict TMB in gastric cancer,and combining appropriate omics data can further improve its accuracy.