Large-Scale 3D Medical Image Pre-Training With Geometric Context Priors

Area of Science:

• Medical Image Analysis
• Artificial Intelligence
• Computer Vision

Background:

• Medical image analysis faces annotation scarcity, especially for 3D data.
• Large-scale pre-training is a promising label-efficient solution but underexplored in medical imaging.
• Learning high-level semantics from unlabeled medical data is challenging.

Purpose of the Study:

• To introduce a self-supervised framework, Volume Contrast (VoCo), for medical image pre-training.
• To leverage inherent geometric context in 3D medical images for representation learning.
• To address the challenge of learning semantics without annotations.

Main Methods:

• Developed the Volume Contrast (VoCo) framework using contrastive learning on 3D medical volumes.
• Extracted base crops to create positive and negative pairs for self-supervised learning.
• Predicted the contextual position of random crops based on similarity to base crops, encoding geometric context.

Main Results:

• Introduced PreCT-160K, the largest medical image pre-training dataset (160K CT volumes).
• Demonstrated VoCo's superiority across 51 diverse medical tasks (segmentation, classification, registration, vision-language).
• Showcased VoCo's strong transferability to unseen modalities and datasets, enhancing performance on limited-data scenarios.

Conclusions:

• VoCo effectively leverages geometric context for self-supervised learning in medical imaging.
• The framework significantly improves performance and accelerates fine-tuning, especially with limited annotations.
• VoCo offers a scalable and versatile approach for medical image pre-training and downstream tasks.