Enhancing gastric cancer prognosis prediction via multi-step multi-modality ensemble survival modeling of HE-stained...

Gastric cancer's profound molecular and phenotypic heterogeneity hinders precision therapy. Delineating tumor microenvironment (TME) heterogeneity enables improved prognosis and tailored treatment strategies. In this study, we propose a Multi-step Multi-modality Ensemble Survival (MMES) model for gastric cancer prognosis based on HE-stained histological images and multiplex immunohistochemistry (mIHC) data.

The MMES model employs a multi-step approach to decompose the complex overall survival (OS) prediction task into smaller, manageable subtasks, enabling the use of lightweight models at each stage. To integrate multi-modal information, the model leverages a pathology foundation model to extract tissue texture features from HE-stained images and utilizes hypergraph neural networks (HGNN) to capture spatial distribution pattern of immune-related molecules from cell-hypergraphs constructed from mIHC data. The MMES model was trained and tested using data of 456 gastric cancer patients across five batches of tissue microarrays (TMAs) from multiple centers. We further expanded the input of the MMES model from TMAs to HE-stained whole slide images (WSIs) of 354 patients in The Cancer Genome Atlas (TCGA) gastric cancer cohort.

Experimental results on our in-house TMA dataset demonstrate that the proposed approach surpasses traditional end-to-end multi-modality modeling strategies, with a 9.9 % increase in concordance index (C-index) and an 11.3 % improvement in mean time-dependent AUC, underscoring the advantages of multi-step modeling. The integration of multi-modal data further enhances the robustness of survival prediction, achieving a 5.4 % higher C-index compared to the average performance of single-modality modeling strategies. Notably, the risk score generated by the MMES model serves as an independent prognostic indicator for gastric cancer patients. Beyond this, results on TCGA data show that the risk score extracted from WSIs is also a strong independent prognostic marker.

The MMES model, through multi-step modeling and the effective integration of HE-stained images and mIHC data, demonstrates significant potential as a reliable and independent prognostic tool for gastric cancer, with promising future applications in WSI-scale multimodal data analysis.