arXiv 2026 · New
VoxForge: A Unified Foundation Model for Dense 3D Features Across MRI, CT and PET
A single frozen 3D encoder, trained with self-distillation and an MRI-first-then-CT curriculum on 235K permissively licensed volumes. Evaluated strictly frozen, it has the best or tied-best average rank in every task group across segmentation, classification and registration, generalizes to PET it never sees, and matches top models on TotalSegmentator CT with a tenth of the labels.