Contrastive learning based method for X-ray and CT registration under surgical equipment occlusion

Area of Science:

• Medical Imaging
• Computer Vision
• Machine Learning

Background:

• Deep learning-based 3D/2D surgical navigation registration shows promise.
• Occlusion by surgical equipment limits accuracy in current methods.

Purpose of the Study:

• To develop a robust 3D/2D surgical navigation registration method that overcomes occlusion challenges.
• To enhance registration accuracy and reliability in complex surgical scenarios.

Main Methods:

• A contrastive learning strategy was employed, treating occluded and unoccluded X-rays as positive samples.
• A Transformer-based residual connection (ResTrans) model was designed to improve long-sequence mapping and feature retrieval.
• A learning-based region of interest (RoI) fine-tuning method was incorporated to refine misalignment.

Main Results:

• The ResTrans model achieved a mean target registration error (mTRE) of 3.25 mm.
• The method demonstrated a fast running time of 1.59 s.
• Experimental results showed superior performance compared to state-of-the-art methods on occluded X-rays.

Conclusions:

• The proposed ResTrans method effectively addresses occlusion issues in 3D/2D surgical navigation registration.
• This approach offers improved accuracy and efficiency for real-time surgical guidance.
• The contrastive learning and RoI fine-tuning strategies contribute to robust feature adaptation and precise alignment.