Automatic x-ray to CT registration using embedding reconstruction and lite cross-attention.

Area of Science:

• Medical Imaging
• Computer-Aided Surgery
• Image Registration

Background:

• Intraoperative x-ray image registration with preoperative CT is crucial for image-guided surgery.
• Existing methods struggle with interpretable and stable fusion of multi-modal image data.
• Current feature extraction and fusion techniques limit the accuracy of pose regression.

Purpose of the Study:

• To develop a rapid, robust method for accurate initial registration using both x-ray and CT images.
• To integrate learning-based and traditional registration approaches for accelerated outcomes.
• To achieve accurate registration across a broad search space.

Main Methods:

• A regression-based framework with constrained feature fusion via embedding reconstruction.
• Introduction of a lightweight cross-attention (LCAT) mechanism for effective feature fusion.
• Utilizing CMA-ES for intensity-based refinement of pose regression results.

Main Results:

• Achieved median rotation error of 1.9° and translation error of 5.6 mm on simulated data with a large search range.
• On real x-ray images, achieved median rotation error of 1.6° and translation error of 3.8 mm with a smaller search range.
• Demonstrated that LCAT and embedding reconstruction modules are vital for performance, with their absence leading to random initialization results.

Conclusions:

• The proposed method offers superior accuracy and a larger capture range compared to existing techniques for x-ray and CT image registration.
• Experimental results on both simulated and real images highlight the method's potential for clinical application in image-guided surgery.