A robust and interpretable deep learning framework for multi-modal registration via keypoints.

Area of Science:

• Medical Imaging
• Computer Vision
• Machine Learning

Background:

• Current deep learning image registration methods lack robustness to large misalignments and interpretability.
• Existing models often fail to leverage problem symmetries and provide only single predictions.

Purpose of the Study:

• To introduce KeyMorph, a deep learning framework for robust and interpretable image registration using automatically detected keypoints.
• To address limitations of current methods by incorporating keypoint-based transformations and symmetries.

Main Methods:

• KeyMorph employs a differentiable closed-form expression for optimal transformation derived from corresponding keypoints.
• Keypoints are learned end-to-end for registration without requiring ground-truth keypoint data.
• The framework is designed for equivariance to translations and symmetry with respect to input image ordering.

Main Results:

• KeyMorph achieves substantially more robust registration and enhanced interpretability by visualizing keypoint-driven alignment.
• The method efficiently computes multiple deformation fields at test time for various transformation variants.
• Demonstrated superior registration accuracy over state-of-the-art methods, particularly for large displacements in 3D multi-modal brain MRI scans.

Conclusions:

• KeyMorph offers a robust, interpretable, and efficient solution for medical image registration, especially in challenging scenarios with large misalignments.
• The keypoint-centric approach provides novel insights into the registration process and outperforms existing techniques.