Cortical Supervertices Improve Alzheimer's Detection with Vision Transformers

A new AI model called CSV-ViT uses variable-sized cortical supervertices (CSVs) to analyze brain cortical surfaces from structural MRI scans, aiming to detect Alzheimer's disease pathologies more accurately. The approach addresses limitations of existing surface-based deep learning models, which often include non-cortical regions like the medial wall and suffer from duplicate vertices at patch boundaries. By preserving regions of interest (ROIs) and partitioning the cortical surface into vertex-based patches of varying sizes, CSV-ViT improves upon uniform patch methods. The model is designed to work on non-Euclidean manifolds, specifically the spherical topology of cortical surfaces. This research, published on arXiv (2605.26514), proposes a tokenization strategy that enhances the Vision Transformer architecture for medical imaging. The goal is to enable prescreening for Alzheimer's disease using structural MRI, which is less costly and invasive than PET scans. The study focuses on computational methodology rather than clinical validation.