#31 · ML / AI methods

Learned representations and cross-cohort validation for MRI-based brain analysis

The four preprints apply machine learning to MRI-derived brain data but address different problems. Two focus on representation: a GNN framework learns a continuous precision matrix from fMRI, avoiding binarization and outperforming binary models, while Gyral-Sulcal-Net unifies gyral and sulcal folding landmarks into a single network tested on over 1,600 brains across ages and cohorts. A third performs systematic external validation of nnU-Net for stroke lesion segmentation across acute and chronic stages and DWI, FLAIR, and T1 modalities on public datasets. The fourth uses a 3D TransUNet to synthesize FA and MD diffusion maps from T1w MRI, improving multi-modal Alzheimer's classification and MCI detection without requiring dMRI scans. The dominant thread is movement beyond single-task, single-modality models toward representations and validation tested across cohorts, disease stages, and available modalities.