Dense Depth Estimation in Monocular Endoscopy With Self-Supervised Learning Methods.

Area of Science:

• Medical Imaging
• Computer Vision
• Machine Learning

Background:

• Accurate 3D reconstruction from monocular endoscopy is crucial for surgical guidance.
• Existing methods often require manual annotations or patient-specific anatomical models, limiting their applicability.
• Self-supervised learning offers a promising avenue to overcome these limitations.

Purpose of the Study:

• To develop a self-supervised deep learning method for dense depth estimation from monocular endoscopic videos.
• To eliminate the need for manual labeling or computed tomography (CT) scans during training and application.
• To evaluate the method's accuracy and compare its performance against existing approaches.

Main Methods:

• Utilized convolutional neural networks trained using a self-supervised approach.
• Employed monocular endoscopic videos and sparse supervision from multi-view stereo (MVS) methods like structure from motion (SfM).
• Avoided a priori modeling of anatomy or shading, and did not require manual labels or CT scans.

Main Results:

• Achieved submillimeter mean residual error in cross-patient experiments using CT scans as ground truth.
• Demonstrated superior performance compared to recent self-supervised depth estimation methods for natural videos on in vivo sinus endoscopy data.
• The source code is publicly available.

Conclusions:

• The proposed self-supervised method enables accurate dense depth estimation from monocular endoscopic data without manual annotations or CT scans.
• This approach significantly advances the capabilities of computer-assisted interventions and surgical navigation.
• The method shows strong potential for real-world clinical applications.