EnTrust: Modeling Inter-Modal Conflict for Trustworthy Multimodal Medical Image Analysis

EnTrust is a novel framework for multimodal medical image analysis that addresses inter-modal conflict to enhance predictive reliability. It features an EnFuse module that disentangles multimodal features into anatomical consensus, modality-specific cues, and conflict signals, and employs a diffusion-based generative segmentation model called SegDiff. Achieving state-of-the-art segmentation accuracy across four medical benchmarks while reducing calibration error by 40% compared to leading methods, EnTrust offers a more efficient alternative to deep ensembles, operating with a single model at approximately half the memory usage, which is crucial for practitioners seeking reliable and interpretable AI solutions in clinical settings.