A diffusion-stimulated CT-US registration model with self-supervised learning and synthetic-to-real domain adaptation

Area of Science:

• Medical Imaging
• Computer Vision
• Surgical Technology

Background:

• Accurate 2D ultrasound (US) to 3D computed tomography (CT) registration is crucial for abdominal interventional procedures but faces challenges with cost, real-time performance, and domain adaptation.
• Existing deep learning methods often require extensive manual annotations or struggle to bridge the gap between different imaging modalities.

Purpose of the Study:

• To develop a novel, robust, and cost-effective method for precise CT-US registration in abdominal interventions.
• To overcome limitations of traditional tracking sensors and current deep learning approaches for cross-modality image alignment.

Main Methods:

• A diffusion-stimulated CT-US registration model was proposed, generating synthetic US images from CT data using physical diffusion properties.
• A synthetic-to-real domain adaptation strategy with a diffusion model was employed to reduce discrepancies between real and synthetic US images.
• A dual-stream self-supervised regression neural network was trained on synthetic data for pose estimation within CT space.

Main Results:

• The proposed method accurately initialized US image pose within an acceptable error range and refined it for precise alignment.
• Validation using a dual-modality human abdominal phantom demonstrated the effectiveness of the approach.
• The method achieved real-time, tracker-independent, and robust rigid registration of CT and US images.

Conclusions:

• The novel diffusion-stimulated model offers a significant advancement in CT-US image registration for abdominal interventions.
• This approach provides a cost-effective, accurate, and robust solution for real-time image guidance, enhancing procedural safety and efficiency.