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

Abstract

Colonoscopy remains the gold standard for detecting and treating precancerous polyps, yet it lacks real-time three-dimensional feedback, leading to blind spots and missed lesions. Using 3D reconstruction algorithms can alert clinicians to potentially missed areas, thereby improving the quality of colonoscopy examinations. Monocular depth estimation is crucial for achieving 3D reconstruction in colonoscopy. However, the complex image features of the colon surface, such as textures and reflections, along with varying geometric characteristics like deep folds, significantly affect depth distribution and degrade performance. To address these challenges, we introduce ColonFMGC, a monocular depth estimation framework that leverages a foundational model encoder pretrained on diverse natural images for robust feature extraction and cross-domain generalization. To accommodate colon-specific depth distributions and geometric structures, we design an Adaptive Depth Adjustment (ADA) module that refines local depth bins using multi-scale decoder features, and we apply a geometric consistency loss that enforces surface normal, gradient, and structural similarity constraints. Our approach demonstrated superior performance in extensive experiments on SimCol, C3VD datasets. Exploiting geometric similarities between synthetic and clinical colonoscopy data, our model trained solely on synthetic data also achieves strong zero-shot transfer to real colonoscopy frames. These results demonstrate that combining foundational models with geometry-aware refinement yields accurate, generalizable monocular depth estimates, paving the way for real-time 3D reconstruction and improved clinical guidance in colonoscopy.