An automatic extraction method on medical feature points based on PointNet++ for robot-assisted knee arthroplasty

Area of Science:

• Medical Robotics
• Computer-Aided Surgery
• Deep Learning in Medical Imaging

Background:

• Image registration is vital for robot-assisted knee arthroplasty, enabling real-time surgical guidance.
• Current manual feature point identification is time-consuming and subjective, impacting surgical consistency.

Purpose of the Study:

• To develop an automated method for medical feature point extraction to improve efficiency and consistency in knee arthroplasty.
• To address the limitations of manual feature point identification in preoperative planning.

Main Methods:

• Proposed Point_RegNet, a deep learning model based on PointNet++, for automatic medical feature point extraction.
• Modified PointNet++ by replacing classification/segmentation layers with a regression layer to predict feature point locations.
• Conducted comparative experiments to optimize the PointNet++ abstraction layers for feature point extraction.

Main Results:

• The optimal network configuration utilized three set abstraction layers for effective feature point extraction.
• Achieved a mean error of less than 5 mm for feature point prediction, outperforming manual methods by 1 mm.
• Automated extraction completed in under 3 seconds, a substantial improvement over the >30 minutes required for manual marking.

Conclusions:

• The deep learning-based Point_RegNet method enhances surgical accuracy and significantly reduces preoperative preparation time.
• This automated approach offers potential applications in other surgical navigation systems, improving overall efficiency and consistency.