Survival Analysis of Histopathological Image Based on a Pretrained Hypergraph Model of Spatial Transcriptomics Data

Abstract

Survival analysis is essential in clinical decision-making and prognosis for breast cancer, a leading cause of mortality among women worldwide. Traditional methods often integrate histopathology images with bulk RNA-seq data to predict patient outcomes. While these multimodal approaches have enhanced the precision of survival predictions, they typically overlook the spatial distribution of cellular elements, a factor critical in understanding tumor behavior and progression. This study introduces a pioneering framework, the Multimodal Hypergraph Neural Network for survival analysis (MHNN-surv), which innovatively combines spatial transcriptomics with Whole-Slide Imaging (WSI) data. Our approach starts with the segmentation of WSI into image patches, followed by feature extraction and predictive modeling of gene expressions. We construct a novel dual hypergraph model where the image-based hypergraph is built using three-dimensional nearest-neighbor relationships, and the gene-based hypergraph leverages spatial transcriptional similarities. This dual-model integration allows for an advanced level of analysis, harnessing the rich morphological and genetic data to provide a more granular understanding of tumor environments. MHNN-surv employs the Cox proportional hazards model to perform robust survival analysis, demonstrating superior performance over existing state-of-the-art multimodal models through extensive validation on a comprehensive breast cancer dataset. Our findings not only affirm the benefit of integrating spatial genomic data into survival analysis but also pave the way for more precise and individualized cancer treatment strategies, potentially transforming patient care by providing deeper insights into the underlying mechanisms of tumor progression and resistance to therapies.