#8 · ML / AI methods

Learned models for task-fMRI, MRI retrieval, and multimodal Alzheimer's diagnosis

The common thread is learned modeling of brain imaging data built on structured inputs, including ROI-level connectivity, anatomical structure, and shared embedding spaces. CoHyFuse and MovieSTAGE encode task-fMRI functional connectivity as hypergraphs, fused with whole-session or whole-movie FC; MovieSTAGE classifies ADHD from movie-fMRI in 260 CMI-HBN. BRAID-fMRI aligns ROI-wise responses across 8 datasets into a shared CLIP space, while NeuroCBIR retrieves whole-brain and region-specific 3D T1w MRI via VAE and contrastive learning over 103 regions. CoPoE handles missing modalities in Alzheimer's diagnosis through a disease-coordinate product-of-experts evaluated on ADNI AUROC. What is new is the joint use of condition-specific graph structure, cross-dataset semantic alignment, and explicit handling of incomplete inputs within learned neuroimaging frameworks.