Colorectal Cancer Survival Prediction Using Multimodal Fusion.

This paper presents a novel multimodal fusion-based framework for colorectal cancer (CRC) survival prediction, integrating whole-slide images (WSIs) with clinical and biomarker data. The framework combines the ProvGigaPath model for WSI feature extraction and the TabTransformer for tabular data representation. Two fusion strategies were explored: contrastive learning (CL) on limited data and cross-attention (CA) for intermediate-level integration. Experiments on the Molecular and Cellular Oncology (MCO) dataset showed that CL achieves AUROCs of 0.82 (tabular) and 0.73 (image), while CA-based multimodal fusion attains a C-index of 0.77, surpassing unimodal baselines.