Anatomy-aware computed tomography-to-ultrasound spine registration.

Area of Science:

• Medical Imaging
• Deep Learning
• Computational Anatomy

Background:

• Ultrasound (US) is effective for guiding lumbar spine injections, but image artifacts pose interpretation challenges.
• Computed tomography (CT)-to-US registration aligns pre-operative CT scans with intra-operative US for precise landmark localization.

Purpose of the Study:

• Propose a deep learning (DL) pipeline for CT-to-US registration.
• Address the need for annotated medical data by developing a data generation method for paired CT-US data with physically consistent spine deformations.
• Train a point cloud (PC) registration network using anatomy-aware losses for anatomically consistent predictions.

Main Methods:

• Generate paired CT-US data by modeling vertebral joint and disk properties based on biomechanical measurements.
• Simulate spine deformations in supine and prone positions using forces applied to spine models from the VerSe dataset.
• Employ anatomy-aware losses, including rigidity and bio-mechanical losses, to enforce realistic spine physics during network training.

Main Results:

• The data generation pipeline produces realistic spinal deformations and plausible simulated ultrasound images.
• Anatomy-aware losses reduce target registration error (TRE) by 0.25 mm compared to mean squared error (MSE) loss.
• The proposed method achieves results close to state-of-the-art (SOTA) on simulated US data (TRE of 3.89 mm) and demonstrates improved robustness against initialization errors (TRE of 4.88 mm vs. 5.66 mm).

Conclusions:

• A pipeline for spine CT-to-US registration using anatomy-aware losses is presented.
• A fully automatic method for synthesizing paired CT-US data with physically consistent deformations is proposed, enabling extensive dataset generation for network training.
• The generated dataset and source code are publicly available.