Contrastive Learning of Relative Position Regression for One-Shot Object Localization in 3D Medical Images

Deep learning networks have shown promising performance for object localization in medical images, but require large amount of annotated data supervised training. To address this problem, we propose: 1) A novel contrastive method which embeds the anatomical structure by predicting Relative Position Regression (RPR) between any two patches from same volume; 2) An one-shot framework organ and landmark volumetric images. Our main idea comes that tissues organs different human bodies own similar relative position context. Therefore, could predict positions their non-local patches, thus locate target organ. is composed three parts: deep network trained to project input patch into a 3D latent vector, representing its position; coarse-to-fine contains projection networks, providing more accurate target; 3) Based on model, transfer bounding-box (B-box) detection locating six extreme points along x, y z directions query volume. Experiments multi-organ head-and-neck (HaN) abdominal CT volumes showed our acquired competitive real time, \(10^5\) times faster than template matching methods with setting Code available at https://github.com/HiLab-git/RPR-Loc.