Self-Supervised Monocular Depth and Pose Estimation for Endoscopy With Latent Priors.

Area of Science:

• Medical Imaging
• Computer Vision
• Gastroenterology

Background:

• Accurate 3D reconstruction in gastrointestinal (GI) endoscopy is crucial for quantitative lesion characterization.
• Monocular endoscopy systems face challenges with depth and pose estimation due to limited data and generalizability issues.
• Existing methods often struggle with the complex visual conditions within the GI tract.

Purpose of the Study:

• To develop a robust self-supervised monocular depth and pose estimation framework for endoscopic imaging.
• To improve the accuracy and generalizability of 3D reconstruction in challenging endoscopic environments.
• To enhance lesion characterization capabilities in the GI tract.

Main Methods:

• Proposed a framework integrating a StyleGAN-based generator for realistic depth prediction and a Variational Autoencoder (VAE) for pose estimation.
• Leveraged natural image depth scenes to condition the depth network, enhancing prediction robustness via latent feature priors.
• Incorporated a prior transfer module to distill motion knowledge from natural scene SLAM systems into the endoscopic domain for improved pose stability.

Main Results:

• The framework demonstrated superior performance in endoscopic depth and pose estimation compared to existing self-supervised methods.
• Evaluations on SimCol, C3VD, and EndoSLAM datasets confirmed the method's effectiveness.
• Achieved accurate depth and pose predictions, addressing complex GI tract textures and lighting.

Conclusions:

• The developed dual refinement pipeline significantly enhances 3D reconstruction accuracy in monocular endoscopy.
• The self-supervised framework offers a robust solution for quantitative lesion characterization in the GI tract.
• This approach improves trajectory consistency and generalizability in challenging endoscopic conditions.