Self-supervised neural network-based endoscopic monocular 3D reconstruction method.

Area of Science:

• Medical Imaging
• Computer Vision
• Deep Learning

Background:

• Monocular visual 3D reconstruction using deep learning is established in conventional fields.
• Self-supervised deep learning is crucial for medical endoscopic imaging due to data acquisition challenges.
• Existing endoscopic 3D reconstruction research is primarily lab-based, lacking clinical environmental adaptability.

Purpose of the Study:

• To develop a robust monocular endoscopic 3D reconstruction framework for complex clinical surgical environments.
• To improve the accuracy and reliability of 3D reconstruction from endoscopic video data.
• To address challenges like inconsistent brightness and intricate scene details in clinical endoscopic imaging.

Main Methods:

• An optical flow-based neural network was employed to manage frame-to-frame brightness variations.
• Attention modules were integrated to enhance focus on pixel texture and depth differentials.
• Inter-layer losses were utilized for multi-scale supervision within the reconstruction network.

Main Results:

• A complete monocular endoscopic 3D reconstruction framework was established and validated on a clinical dataset.
• The proposed framework demonstrated superior simulation of frame mapping during endoscope motion compared to other self-supervised methods.
• Quantitative experiments using the cross-correlation coefficient metric confirmed the framework's effectiveness.

Conclusions:

• The developed framework successfully addresses the complexities of clinical endoscopic scenes for 3D reconstruction.
• The method shows excellent generalization capabilities, performing well on external datasets like SCARED.
• This work advances self-supervised endoscopic 3D reconstruction for practical clinical applications.