Ruiyi Yang

China Medical University

Student

PLCO (Learn more about this study)

PLCOI-1679

Sep 26, 2024

To improve the prognosis of cancer through a deep learning system based on TME described by pathological image

We aim to predict the tumor microenvironment (TME) by analyzing histological images and use this information to improve the prognostic prediction accuracy of cancer patients. The tumor microenvironment refers to the complex population of cells and non-cellular components surrounding tumor cells, which together influence tumor growth, invasion, and response to therapy. Although spatial transcriptomics (ST) technology can reveal the spatial gene expression patterns of TME, its high cost and experimental cycle limit its application in large-scale clinical studies. To overcome these limitations, we developed a deep learning model that combines the advantages of image processing and graph neural networks to be able to predict ST data from routine histological images. The IGI-DL model predicts the expression pattern of genes in the tumor region by analyzing the pixel intensity characteristics in hematoxylin-eosin (H&E) stained images and the geometric information of the tissue structure. This approach not only improves the accuracy of prediction but also reduces the reliance on expensive and time-consuming ST experiments. We validated the performance of the model on NSCLC tissue samples. By comparing with the five existing prediction models, it shows a higher average correlation, which means that it is able to predict gene expression levels in tumor tissues more accurately. In addition, the model also shows good cross-platform generalization ability, showing stable performance even on data generated by different technology platforms. In addition to predicting gene expression, we also constructed a super-patch graph survival model based on predicted spatial gene expression for cancer prognosis prediction. This model utilizes the spatial structure information of the entire tumor tissue to capture tumor heterogeneity through graph neural networks and predict the survival risk of patients.