Self-supervised monocular depth estimation for high field of view colonoscopy cameras

Area of Science:

• Medical imaging
• Robotics
• Artificial intelligence

Background:

• Optical colonoscopy is the standard for colorectal cancer detection but misses 22%-28% of polyps, leading to interval cancers.
• Improving polyp detection rates and patient comfort is crucial for effective colorectal cancer screening.
• Accurate 3D environmental understanding is vital for autonomous colonoscope navigation and adenoma detection.

Purpose of the Study:

• To develop a self-supervised monocular depth estimation model for colonoscopy.
• To enable accurate 3D environmental perception for autonomous colonoscope navigation.
• To enhance adenoma detection rates and reduce missed polyps during colonoscopy.

Main Methods:

• Utilized a self-supervised deep learning approach for monocular depth estimation directly from colonoscopy video sequences.
• Employed view synthesis techniques to train the model without requiring ground-truth depth maps.
• Adapted the model to handle challenges specific to colonoscopy, including wide field-of-view cameras, deformable surfaces, specular lighting, non-Lambertian surfaces, and high occlusion.

Main Results:

• The model successfully estimates depth from colonoscopy videos using self-supervision.
• Demonstrated accommodation of wide field-of-view cameras and challenging visual conditions.
• Achieved near real-time performance for depth estimation on synthetic datasets, a colonoscopy training model, and real colonoscopy videos.

Conclusions:

• Self-supervised monocular depth estimation is a viable solution for 3D scene understanding in colonoscopy.
• This technology can significantly improve the accuracy and systematic inspection capabilities of autonomous colonoscopes.
• The developed model offers a pathway to reduce missed polyps and interval cancers, enhancing colorectal cancer screening outcomes.