CoRe: Joint Optimization with Contrastive Learning for Medical Image Registration

Area of Science:

• Medical image analysis
• Computer vision
• Machine learning

Background:

• Medical image registration aligns images from different sources, crucial for diagnosis and treatment planning.
• Challenges include intensity variations and complex tissue deformations, hindering registration accuracy.
• Self-supervised learning shows potential for robust feature extraction in medical imaging.

Purpose of the Study:

• To develop a novel framework integrating equivariant contrastive learning directly into medical image registration.
• To enhance the robustness and accuracy of image registration by learning deformation-invariant representations.
• To improve the suitability of learned features for the registration task through joint optimization.

Main Methods:

• Proposed a novel framework integrating equivariant contrastive learning within the registration model.
• Leveraged contrastive learning to acquire feature representations invariant to tissue deformations.
• Jointly optimized contrastive and registration objectives for informative and task-specific representations.

Main Results:

• Evaluated the method on abdominal and thoracic image registration tasks (intra- and inter-patient).
• Demonstrated significant performance improvements compared to strong baseline methods.
• Showcased the effectiveness of integrating contrastive learning directly into the registration framework.

Conclusions:

• The proposed framework significantly enhances medical image registration performance.
• Equivariant contrastive learning provides robust, deformation-invariant features beneficial for registration.
• This approach offers a promising direction for improving medical image analysis.