Image Intrinsic-Based Unsupervised Monocular Depth Estimation in Endoscopy.

Area of Science:

• Computer Vision
• Medical Imaging
• Surgical Technology

Background:

• Unsupervised monocular depth estimation is crucial for endoscopy in minimally invasive surgery (MIS).
• Endoscopic imaging challenges traditional photometric consistency assumptions, hindering depth estimation.
• Existing methods often rely on image pre-processing, which can be suboptimal.

Purpose of the Study:

• To develop a novel framework for unsupervised monocular depth estimation in endoscopy.
• To integrate intrinsic image decomposition (IID) with deep learning for improved depth accuracy.
• To overcome the limitations posed by endoscopic imaging characteristics without pre-processing.

Main Methods:

• A novel end-to-end framework combining an image intrinsic decomposition module and a synthesis reconstruction module.
• Leveraging the albedo map from IID to address challenging endoscopic image properties.
• Designing dedicated loss functions for robust network training based on IID integration.

Main Results:

• The proposed intrinsic-based method outperforms state-of-the-art techniques on SCARED and Hamlyn datasets.
• Validation confirms the method's generalization ability and the effectiveness of its components.
• Achieved superior results in unsupervised monocular depth estimation for endoscopic images.

Conclusions:

• The intrinsic-based unsupervised monocular depth learning framework offers a pioneering solution for MIS.
• This method enhances 3D reconstruction quality and robustness for augmented reality navigation.
• The approach effectively circumvents challenging endoscopic imaging characteristics through IID.