3D reconstruction from endoscopy images: A survey.

Area of Science:

• Medical Imaging
• Computer Vision
• Surgical Technology

Background:

• Three-dimensional (3D) reconstruction from endoscopic images is crucial for numerous medical applications.
• Endoscopes are classified as monocular or binocular, influencing depth estimation strategies.
• Traditional depth estimation methods struggle with endoscopic challenges like poor illumination and sparse textures.

Purpose of the Study:

• To comprehensively review and classify depth estimation methods for endoscopic imaging.
• To analyze the evolution and performance of traditional and deep learning-based approaches.
• To provide guidance for selecting appropriate methods in clinical settings.

Main Methods:

• Systematic literature review of over 170 papers from 2013-2023.
• Classification of methods based on endoscope type (monocular/binocular).
• Analysis of traditional (feature matching, multi-view geometry) and deep learning techniques.
• Summary of common datasets, performance metrics, and scene representations.

Main Results:

• Deep learning methods show promise in overcoming limitations of traditional techniques in endoscopic environments.
• A taxonomy of methods, their advantages, and drawbacks are presented.
• Comparative analysis of qualitative and quantitative performance, robustness, and processing time is provided.
• Commonly used scene representation methods in endoscopy are summarized.

Conclusions:

• Deep learning-based depth estimation is advancing rapidly for endoscopic applications.
• The review facilitates informed selection of 3D reconstruction methods for medical use.
• Future prospects of depth estimation in medical imaging are discussed.