Dependence of MRI’s prognostic contribution on co-modalities and fusion strategy

Characterize how the prognostic contribution of FLAIR MRI features to glioma survival prediction depends on which other modalities are included (haematoxylin and eosin-stained whole-slide histopathology and/or RNA-seq gene expression) and on the integration strategy used (early feature concatenation, late score combination, or joint end-to-end fusion) within a multimodal deep learning Cox proportional hazards framework.

Background

Multiple fusion strategies—early, late, and joint—are commonly used to integrate modalities in survival models. Prior work reported that early fusion performs best for WSI+RNA-seq, but it was unclear how MRI’s contribution would vary with different modality combinations and fusion mechanisms.

The paper explicitly raises this as an unresolved question to guide a systematic evaluation across unimodal, bimodal, and trimodal configurations, aiming to understand when and how MRI adds value depending on context and fusion design.