NeuroAlign improves MCI detection with novel hierarchical fusion of fMRI and DTI

Multimodal neuroimaging fusion of functional MRI (fMRI) and diffusion tensor imaging (DTI) has long promised better cognitive impairment analysis, but clinicians and researchers have struggled with misaligned representations and heterogeneous feature spaces. NeuroAlign, a new hierarchical framework from a team led by Xiongri Shen, tackles this head-on. The architecture introduces Dual-Modal Hierarchical Alignment (DMHA), which models multi-scale dynamic connectivity and explicitly aligns dynamic-static and functional-structural embeddings. It pairs this with Dual-Domain Hierarchical Interaction (DDHI), enabling fine-grained modulation and global interaction between connectivity- and region-level features. The result is a system that can simultaneously leverage the temporal dynamics of fMRI and the structural integrity of DTI.

To make the model inspectable, the team also designed Synergistic Activation Mapping (SAM), a gradient-free attribution method that works on features like dynamic/static functional connectivity, ALFF, and fractional anisotropy. NeuroAlign was tested on three major datasets—GUTCM, ADNI, and OASIS—using five-fold cross-validation. It achieved competitive detection of mild cognitive impairment (MCI) and subjective cognitive decline (SCD), and showed promising cross-dataset transferability. The attribution analyses revealed both modality-specific and partially consistent brain patterns, offering evidence that the model is learning clinically relevant representations. This work bridges the gap between advanced AI and interpretable neuroimaging, potentially accelerating early diagnosis of Alzheimer's and related dementias.