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

Area of Science:

• Artificial Intelligence
• Medical Imaging
• Computer Vision

Background:

• Current self-supervised learning (SSL) methods for 3D medical imaging are limited by simple pretext tasks and specialized datasets.
• This restricts their ability to generalize and scale across diverse medical imaging applications.

Purpose of the Study:

• To develop a more generalizable and scalable self-supervised learning method for 3D medical imaging.
• To pretrain a versatile model on a large, multimodal dataset for improved downstream task performance.

Main Methods:

• Introduction of 3DINO, a novel self-supervised learning approach tailored for 3D datasets.
• Pretraining of 3DINO-ViT, a general-purpose medical imaging model, on an extensive multimodal dataset comprising approximately 100,000 3D scans from over 10 organs.

Main Results:

• The 3DINO-ViT model significantly outperforms existing state-of-the-art pretrained models.
• Demonstrated superior performance across a wide range of downstream medical imaging tasks.

Conclusions:

• 3DINO represents a significant advancement in self-supervised learning for 3D medical imaging.
• The pretrained 3DINO-ViT model offers enhanced generalizability and scalability for diverse medical imaging applications.