DIMAFv2: Interpretable Multimodal Disentangled Representations for Breast Cancer Survival Prediction

Aniek Eijpe, Soufyan Lakbir, Melis Erdal Cesur, Sara P. Oliveira, Sanne Abeln, Wilson Silva

Keywords: Cancer survival prediction, Disentangled representation learning, Multimodal fusion, Interpretability in AI, Histopathology whole-slide images, Transcriptomics data

Abstract: Understanding tumor biology and predicting patient survival is challenging due to the complex and heterogeneous information across multiple modalities. We propose an interpretable framework that integrates histopathology whole-slide images and transcriptomics data by explicitly separating modality-shared and modality-specific representations through disentangled attention fusion. Our approach achieves state-of-the-art breast cancer survival prediction and enhanced disentanglement. DIMAFv2's inherent interpretability revealed important factors driving prediction, such as KRAS signaling and G2M checkpoint pathways interacting with tumor morphology.

Serve As Reviewer: ~Aniek_Eijpe1

Submission Number: 36