A bronchoscopic navigation method based on neural radiation fields.

Area of Science:

• Medical Imaging
• Computer Vision
• Robotics

Background:

• Current bronchoscopic navigation tools often require external infrastructure or active bronchoscope manipulation, posing limitations.
• Challenges in bronchoscopic navigation include low texture environments and difficult lighting conditions.

Purpose of the Study:

• To introduce a novel Neural Radiance Fields (NeRF)-based approach for passive bronchoscopic navigation.
• To enable passive endoscope localization solely from bronchoscopic images, overcoming limitations of existing methods.

Main Methods:

• Leveraged NeRF for endoscope localization in bronchoscopy.
• Developed a two-stage pipeline: offline training with preoperative data and online passive pose estimation.
• Utilized Anderson acceleration and semantic appearance transfer to bridge the sim-to-real gap.

Main Results:

• Achieved an average rotation error of 3.18° and translation error of 4.95 mm on a virtual dataset.
• Demonstrated an average rotation error of 5.14° and translation error of 13.12 mm on a physical phantom.
• Validated performance against SLAM-based methods.

Conclusions:

• The NeRF-based method provides effective bronchoscopic navigation without external infrastructure or active adjustments.
• Experimental results confirm the approach's efficacy in challenging conditions, including low texture and poor lighting.
• This technique offers a promising advancement for safer and more accurate bronchoscopic procedures.