Learning dual-scale context with overlap awareness for keypoint-driven partial-overlap medical image registration

Area of Science:

• Medical Imaging
• Computer Vision
• Computational Anatomy

Background:

• Medical image registration is crucial for clinical applications but challenged by partial anatomical overlap.
• Existing methods often fail with non-overlapping regions due to reliance on local similarity measures.

Purpose of the Study:

• To develop a robust registration framework for medical images with partial overlap.
• To improve cross-image correspondence matching by incorporating anatomical context.

Main Methods:

• A keypoint-driven registration framework using learned descriptors embedded with anatomical context.
• A Dual-scale Context Aggregation Module (DualCAM) to enhance descriptor discriminability by modeling global and local structures.
• An overlap-aware guidance mechanism to focus on reliable overlapping regions.

Main Results:

• The framework successfully builds robust cross-image correspondences even with partial overlap.
• Enhanced descriptors with anatomical context lead to more reliable matching.
• The overlap-aware mechanism mitigates interference from non-overlapping regions, boosting accuracy.

Conclusions:

• The proposed method significantly outperforms state-of-the-art approaches for partial-overlap medical image registration.
• Demonstrated robustness and generalization capabilities on multi-organ abdominal CT datasets.
• Offers a promising solution for clinical applications requiring comparison of images with varying fields of view.