Confidence-aware self-supervised learning for dense monocular depth estimation in dynamic laparoscopic scene.

Area of Science:

• Medical imaging
• Computer vision
• Surgical technology

Background:

• Accurate depth estimation from monocular laparoscopic images is crucial for surgical navigation but is hindered by dynamic environments and unreliable ground truth.
• Factors such as bleeding, smoke, and instrument occlusion introduce noise and outliers, complicating machine learning-based depth estimation.

Purpose of the Study:

• To develop a robust model learning framework for precise monocular depth estimation in dynamic surgical settings.
• To address the challenges of noise and data inconsistencies in laparoscopic surgical videos.

Main Methods:

• A novel framework utilizing a generic laparoscopic surgery video dataset for training.
• Incorporation of binocular disparity confidence as a self-supervisory signal alongside stereo laparoscope disparity.
• A unique loss function that adaptively weights depth data based on confidence, mitigating outlier influence from tissue deformation, smoke, and instruments.

Main Results:

• The model demonstrated exceptional generalization performance on the Hamlyn Dataset and a static dataset.
• The proposed method proved effective across various scene dynamics, laparoscope types, and surgical sites.
• Robust learning was achieved despite the presence of significant noise and outliers common in surgical environments.

Conclusions:

• The developed framework offers a significant advancement in monocular depth estimation for laparoscopic surgery.
• The self-supervisory approach and adaptive loss function enable robust and accurate depth perception in challenging surgical conditions.