Robust Multi-Modal Cancer Survival Prediction

Cancer survival models are typically built and evaluated under an assumption that rarely holds in the clinic: that every modality a patient needs will be available. In practice PET is skipped for cost or access, CT protocols vary across sites, and models trained on complete data quietly fail when a modality goes missing. This poster traces a three-part progression toward survival prediction that is robust to this reality. Prior work first establishes that self-supervised multi-modal fusion of CT, PET and RNA-Seq,refined by within-patient (MPE) and cross-patient (CPM) structure which improves prognosis over single-modality baselines. RobSurv then pushes past a single cohort and clean acquisition, using dual vector-quantized and continuous paths with cross-modal attention to stay reliable across multiple cancer datasets under noise and motion artifacts. Finally, MIMOSA targets modalities absent at inference, by learning shared and modality-specific quantized embeddings so that available scans still map to clinically meaningful codebook prototypes, preserving risk stratification even as CT or PET is fully withheld. Together these works close the gap between "robust on paper" and "robust in practice," pointing toward survival systems deployable at sites without on-site PET, cohorts with shifting protocols, and patients for whom a scan is simply skipped.