Answers are grounded in this paper and cite their sources.
Multi-modal cancer survival prediction in computational pathology aims to integrate diverse data streams, such as histological images and molecular profiles, to enhance prognostic accuracy across various cancer types. However, current fusion strategies often suffer from modality dominance, where high-density pathological features from whole-slide images (WSI) overwhelm sparse genomic signals. Moreover, they frequently overlook the critical high-level evidence provided by structured clinical data. To address these limitations, we propose MPOTSurv , a multi-modal framework that leverages M odality-Balancing P rototype Learning (MBPL) with O ptimal T ransport (OT) to map disparate modalities onto a symmetrical, unified latent manifold with more balanced importance for accurate cancer Surv ival prediction. Within the MBPL module, we introduce Sinkhorn-Knopp OT to achieve principled feature aggregation, enforcing strict marginal constraints that suppress morphological noise in WSI and tabular data. Our method further incorporates a Cross-modal Symmetrical Interaction stage to robustly fuse the features of the two branches, complemented by a Prototypical Regularization to ensure diverse and non-redundant prototype representations. Extensive experiments across five datasets demonstrate that MPOTSurv consistently outperforms state-of-the-art single-modal and multi-modal methods. Code will be available on GitHub .
No takes yet. Share an insight, caveat, or question.
Zeng et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: