A generalizable 3D framework and model for self-supervised learning in medical imaging

Abstract

Current self-supervised learning (SSL) methods for 3D medical imaging rely on simple pretext formulations and organ- or modality-specific datasets, limiting their generalizability and scalability. We present 3DINO, a cutting-edge SSL method adapted to 3D datasets, and pretrain 3DINO-ViT: a general-purpose model for medical imaging, on a ultra-large multimodal dataset of ~100,000 3D scans from over 10 organs. We show 3DINO-ViT outperforms state-of-the-art pretrained models on numerous downstream imaging tasks.