CoReg-Net: contrastive registration for structurally consistent one-shot medical image segmentation

Area of Science:

• Medical Image Analysis
• Computational Anatomy
• Deep Learning

Background:

• One-shot segmentation relies on anatomical label transfer via registration.
• Current deep learning registration methods struggle with appearance variations, causing inconsistent deformation fields.
• Improving structural consistency in registration is crucial for reliable segmentation.

Purpose of the Study:

• To enhance structural consistency in deformable registration for one-shot segmentation.
• To develop a framework that aligns inter-subject anatomical representations for robust label propagation.
• To address the sensitivity of deep learning registration to appearance variations.

Main Methods:

• Proposed CoReg-Net, a contrastive registration framework.
• Integrated an inter-subject contrastive learning (ISCL) module for consistent feature representations.
• Introduced a cascaded skip connection (CSC) mechanism to bridge semantic gaps and refine deformation details.

Main Results:

• Demonstrated robust and consistent one-shot segmentation across diverse datasets (IXI, OASIS, BCV).
• CoReg-Net generated deformation fields with higher structural similarity compared to state-of-the-art methods.
• Showcased adaptive deformability, preserving topology in brain tissues and capturing non-rigid abdominal deformations.

Conclusions:

• Explicitly enforcing structural consistency via contrastive registration improves similarity and segmentation accuracy.
• CoReg-Net offers a reliable and generalizable solution for atlas-guided anatomical analysis.
• The framework effectively handles heterogeneous imaging modalities and anatomical variations.