3DINO: A Generalizable Self-Supervised Learning Framework and Foundation Model for 3D Medical Imaging
This work introduces 3DINO, a self-supervised learning (SSL) framework for 3D medical imaging, and 3DINO-ViT, a general-purpose model pretrained on a large, diverse dataset of ~100,000 scans. Unlike prior methods limited by simple pretext tasks and domain-specific data, 3DINO-ViT exhibits strong generalizability across modalities and organs, outperforming SOTA on various segmentation and classification tasks, including out-of-distribution evaluations. The framework and model are being released to foster research in 3D medical imaging foundation models.