VolTA-3D: Self-Supervised Learning for Brain MRI using 3D Volumetric Token Alignment

A novel self-supervised learning framework named VolTA-3D has been developed for analyzing brain MRIs. This technique employs a 3D Vision Transformer to create transferable volumetric representations by aligning global class-style tokens with local patch tokens in a student-teacher framework, while also promoting detailed structural reconstruction. It tackles the challenges posed by the limited semantic diversity and intricate anatomical features of brain MRIs, which hinder existing SSL techniques. Designed for broad applicability across various datasets, imaging protocols, and downstream tasks, this model surpasses the limitations of current 3D models, which often focus solely on segmentation or classification. The findings are available in a preprint on arXiv (2605.16775).