Shape Registration for Laparoscopic Images Using Offline Diffusion Learning.

Area of Science:

• Medical Imaging
• Computer-Aided Surgery
• Machine Learning

Background:

• Accurate shape registration between patient organ geometries and endoscopic images is vital for image-guided surgery.
• Challenges exist in collecting 3D training data due to limited intraoperative 3D imaging.
• Domain discrepancies between synthetic and real images hinder robust offline learning.

Purpose of the Study:

• To propose a diffusion-based offline learning strategy for liver mesh registration in laparoscopic camera images.
• To mitigate the domain gap between synthetic and real images using shared semantic organ labels.
• To enhance the accuracy and robustness of 2D/3D registration for surgical guidance.

Main Methods:

• A diffusion-based offline learning framework was developed for 2D/3D shape registration.
• Semantic organ labels were utilized as shared image features to bridge the domain gap.
• Gaussian noise was introduced into registration parameters during training, with visual changes in 2D organ labels guiding noise prediction.

Main Results:

• The proposed method demonstrated superior prediction accuracy compared to conventional approaches.
• The model successfully registered liver mesh shapes in laparoscopic camera images.
• Generated image overlays effectively visualized tumors and vascular structures for intraoperative guidance.

Conclusions:

• The diffusion-based offline learning strategy effectively addresses domain discrepancies in 2D/3D registration.
• The developed model provides accurate intraoperative guidance by visualizing critical anatomical structures.
• This approach has significant clinical relevance for enhancing precision in laparoscopic surgery.