Few-Shot Learning for Deformable Medical Image Registration With Perception-Correspondence Decoupling and Reverse...

Area of Science:

• Medical Imaging
• Computer Vision
• Machine Learning

Background:

• Unsupervised registration models lack perception, causing misalignment and distortion.
• Label-constrained models face issues with texture constraints and high labeling costs.

Purpose of the Study:

• To propose the first few-shot deformable medical image registration framework (PC-Reg).
• To embed perception ability into registration models using few labels, enhancing accuracy and reducing distortion.

Main Methods:

• Perception-Correspondence Decoupling: Separates perception and correspondence tasks into two CNNs for independent optimization.
• Reverse Teaching: Utilizes labeled and unlabeled images for few-shot learning, generating additional training data to improve generalization.

Main Results:

• PC-Reg demonstrates competitive registration accuracy and effective distortion reduction on three datasets with only five labels.
• Achieved significant improvements in registration accuracy (Reg-DSC) compared to LC-VoxelMorph.

Conclusions:

• PC-Reg offers a promising solution for accurate and distortion-minimized deformable medical image registration with minimal supervision.
• The framework shows substantial potential for clinical applications due to its efficiency and effectiveness.