A Simple and Effective Regional Contrastive Learning Method for 3D Medical Images

3D medical images can visualize the internal structure of human organs, which have significant advantages over 2D medical images. However, the annotations of 3D medical images are more difficult to obtain compared to 2D medical images, which makes it challenging to capture complete spatial information in 3D medical images. To address this challenge, we propose a simple but effective self-supervised learning method that utilizes the similarity between medical images for region comparison and learns the organ correspondence information from different images but between the same regions. We pre-trained our method on the UK Biobank's cardiac short-axis nuclear magnetic resonance dataset. Our method performs well on three downstream task-related datasets.