Learning to Match 2D Keypoints Across Preoperative MR and Intraoperative Ultrasound.

Area of Science:

• Medical Imaging
• Computer Vision
• Image Registration

Background:

• Accurate matching of pre-operative Magnetic Resonance (MR) and intra-operative Ultrasound (US) images is crucial for image-guided surgery.
• Existing methods struggle with texture variations and modality differences between MR and US images.

Purpose of the Study:

• To develop a texture-invariant 2D keypoint descriptor for robustly matching MR and US images.
• To introduce a "matching-by-synthesis" strategy to bridge the gap between MR and US imaging modalities.

Main Methods:

• A novel texture-invariant 2D keypoint descriptor was developed.
• A "matching-by-synthesis" strategy was employed, synthesizing US images from MR images.
• A patient-specific descriptor network was trained using supervised contrastive learning on keypoints.

Main Results:

• The proposed approach achieved an average matching precision of 80.35% on real cases.
• The method demonstrated superior performance compared to state-of-the-art techniques.
• The developed descriptor proved effective in handling texture variations and modality differences.

Conclusions:

• The proposed texture-invariant keypoint descriptor and matching-by-synthesis strategy offer a robust solution for MR-US image registration.
• This advancement has the potential to improve the accuracy and reliability of image-guided interventions.