Leveraging Near-Field Lighting for Monocular Depth Estimation from Endoscopy Videos

Area of Science:

• Medical Imaging
• Computer Vision
• Surgical Technology

Background:

• Monocular depth estimation in endoscopy is crucial for surgical assistance and robotics.
• Existing methods struggle with endoscopy images due to poor geometric features and lighting.
• Photometric cues offer a potential solution for improved depth perception.

Purpose of the Study:

• To enhance monocular depth estimation in endoscopy videos.
• To address limitations of current depth estimation techniques in endoscopic imaging.
• To leverage photometric information for more accurate depth mapping.

Main Methods:

• Development of novel supervised and self-supervised loss functions using per-pixel shading.
• Introduction of a new depth refinement network (PPSNet) based on per-pixel shading.
• Application of teacher-student transfer learning for depth map generation.

Main Results:

• Achieved state-of-the-art performance on the C3VD dataset.
• Successfully generated high-quality depth maps from clinical endoscopy data.
• Demonstrated the effectiveness of photometric cues and the PPSNet architecture.

Conclusions:

• The proposed methods significantly improve monocular depth estimation in endoscopy.
• PPSNet and photometric cues enhance depth accuracy for surgical applications.
• The approach is effective for both synthetic and real-world clinical data.