A 3-D Cross-Modal Keypoint Descriptor for MR-US Matching and Registration
IEEE Transactions on Medical Imaging
|April 3, 2026
Summary
This study introduces a new 3D cross-modal keypoint descriptor for aligning real-time ultrasound (iUS) with MRI. The method improves intraoperative registration accuracy despite differences in imaging modalities.
Area of Science:
- Medical Imaging
- Computer Vision
- Surgical Navigation
Background:
- Intraoperative registration of real-time ultrasound (iUS) to preoperative Magnetic Resonance Imaging (MRI) is challenging due to significant differences in appearance, resolution, and field-of-view.
- Existing methods struggle to bridge these modality-specific gaps, hindering accurate surgical guidance.
Purpose of the Study:
- To develop a novel 3D cross-modal keypoint descriptor for robust MRI-iUS matching and registration.
- To enable accurate intraoperative alignment of real-time ultrasound with preoperative MRI data.
Main Methods:
- A patient-specific matching-by-synthesis approach generates synthetic iUS from MRI for supervised contrastive training.
- A probabilistic keypoint detection strategy identifies salient, modality-consistent locations.
- Curriculum-based triplet loss with hard negative mining trains rotation-invariant descriptors robust to iUS artifacts.
Main Results:
- The proposed descriptor achieved 69.8% average precision in matching across 11 patients, outperforming state-of-the-art methods.
- The registration approach yielded a competitive mean Target Registration Error of 2.39 mm on the ReMIND2Reg benchmark.
- The framework demonstrated robustness to iUS field-of-view variations and required no manual initialization.
Conclusions:
- The novel 3D cross-modal keypoint descriptor effectively addresses the challenges of MRI-iUS registration.
- This approach offers an interpretable, accurate, and robust solution for intraoperative image guidance, enhancing surgical navigation.
