Robust Self-Supervised Monocular Depth Estimation for Endoscopic Soft Tissue Deformation Scenes With Biomechanical...

Area of Science:

• Computer Vision
• Medical Imaging
• Robotics

Background:

• Self-supervised learning (SSL) excels in monocular depth and ego-motion estimation.
• Endoscopic scenes present unique challenges: specular reflections and soft tissue deformation hinder SSL accuracy.

Purpose of the Study:

• To develop a robust SSL method for monocular depth estimation in challenging endoscopic environments.
• To enhance accuracy by addressing specular reflections and soft tissue deformations.

Main Methods:

• Introduced an image distance transform strategy for handling specular reflections.
• Proposed a biomechanical soft tissue deformation constraint to improve pixel-level accuracy.
• Employed a lightweight architecture for efficient computation and faster inference.

Main Results:

• Achieved robust monocular depth estimation on public (SCARED, SERV-CT) and custom datasets.
• Demonstrated comparable accuracy and robustness to state-of-the-art (SOTA) methods.
• Attained an RMSE of 4.96 mm with only 2.25M parameters on the SCARED dataset.

Conclusions:

• The proposed method effectively overcomes limitations of existing SSL techniques in endoscopic vision.
• The approach shows significant potential for clinical applications in computer-assisted surgical navigation.
• Offers a balance of accuracy, robustness, and computational efficiency for real-time performance.