Depth-Regularized 3D Gaussian Splatting for Robust Endoscopic Reconstruction in Feature-Scarce Environments.

Area of Science:

• Medical Imaging
• Computer Vision
• Geometric Modeling

Background:

• 3D reconstruction in endoscopy is crucial for diagnostics, surgical planning, and training.
• Existing methods like Neural Radiance Fields and 3D Gaussian Splatting (3DGS) face limitations in gastrointestinal (GI) environments due to narrow viewpoints and uniform textures.
• Reliable 3D modeling in GI endoscopy remains a significant challenge.

Purpose of the Study:

• To develop an advanced 3DGS-based framework for robust 3D reconstruction in endoscopic GI imaging.
• To overcome the limitations of existing methods in challenging GI environments.
• To improve diagnostic accuracy, surgical planning, and training through enhanced 3D modeling.

Main Methods:

• Integration of a deep learning-based Structure-from-Motion (SfM) technique with 3DGS.
• Utilizing Super-point and Superglue for robust feature extraction and matching in GI tract scenes for accurate camera pose estimation.
• Refining reconstruction via alignment of monocular depth predictions (Depth-Anything-V2) with SfM-derived depth using L1 loss.
• Implementing hard depth regularization (Huber loss) and global-local depth normalization for precise Gaussian placement and structural consistency.

Main Results:

• The proposed framework demonstrates significantly enhanced 3D reconstruction quality in endoscopic imaging.
• Reduced artifacts were observed compared to standard methods.
• The method proved effective in challenging GI environments, improving camera pose estimation and 3D Gaussian initialization.
• Depth alignment and regularization techniques ensured precise reconstruction and preserved fine details.

Conclusions:

• The advanced 3DGS framework effectively addresses the challenges of 3D reconstruction in GI endoscopy.
• The integration of deep learning-based SfM and sophisticated depth regularization leads to superior reconstruction quality.
• This approach holds promise for improving clinical applications reliant on accurate endoscopic 3D models.