Foundational model-based geometric consistency monocular depth estimation framework for colonoscopy.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Gastroenterology

Background:

• Colonoscopy is vital for detecting precancerous polyps but lacks real-time 3D visualization, leading to missed lesions.
• Current monocular depth estimation methods struggle with complex colon surface features like textures, reflections, and folds, degrading performance.

Purpose of the Study:

• To introduce ColonFMGC, a novel monocular depth estimation framework for improved 3D reconstruction in colonoscopy.
• To enhance the accuracy and generalizability of depth estimation in challenging colon environments.

Main Methods:

• Utilized a foundational model encoder pretrained on natural images for robust feature extraction.
• Developed an Adaptive Depth Adjustment (ADA) module to refine depth distributions using multi-scale features.
• Implemented a geometric consistency loss enforcing surface normal, gradient, and structural similarity.

Main Results:

• ColonFMGC demonstrated superior performance on SimCol and C3VD datasets.
• The model achieved strong zero-shot transfer to real colonoscopy frames when trained solely on synthetic data.
• The framework successfully integrates foundational models with geometry-aware refinement for accurate depth estimation.

Conclusions:

• ColonFMGC provides accurate and generalizable monocular depth estimates for colonoscopy.
• The approach facilitates real-time 3D reconstruction, potentially improving lesion detection and clinical guidance.
• Combining AI with geometric constraints offers a promising direction for advancing endoscopic imaging.