K-MIND: A Knowledge-guided Multi-modal Interaction Network for Brain Disorders

Published in Manuscript in Preparation, 2026

Abstract

Existing multimodal fusion methods coarsely concatenate fMRI and sMRI features without incorporating validated medical knowledge to guide cross-modality interaction. K-MIND encodes PubMedBERT-derived semantic priors for 246 brain subregions, and fuses fMRI and sMRI connectivity via structure-guided projection, orthonormal clustered attention, and QFormer cross-modal interaction — knowledge explicitly directs modality attention at brain-region granularity. Evaluated on 2,350 subjects across three private TLE datasets and two public benchmarks (ABIDE, ADHD-200).

Recommended citation: Fang, Z., Chen, H.*, Li, R.*. (In Preparation). K-MIND: A Knowledge-guided Multi-modal Interaction Network for Brain Disorders.
Download Paper