Self-supervised monocular depth estimation for colonoscopy using normal-guided cross-attention

Reliable 3D reconstruction of colonoscopic scenes can facilitate navigation and enhance lesion assessment by indicating the regions that have been inspected and revealing the geometric structure of polyps. However, accurate depth estimation in colonoscopy remains highly challenging due to specular highlights and low-texture areas. This study aims to improve monocular depth estimation in colonoscopy by incorporating surface normal information.

We propose a self-supervised deep learning method for relative depth estimation that integrates surface normal information through a cross-attention mechanism. Surface normals are first predicted by an existing normal estimation model and then used as auxiliary geometric priors. The depth estimation network employs cross-attention to adaptively fuse normal features with image features, enabling spatially selective geometric refinement.

Quantitative and qualitative comparisons show that the proposed method outperforms state-of-the-art self-supervised methods on SimCol3D, C3VD, and real colonoscopy data. The proposed method generates more realistic depth maps, preserving mucosal folds and lumen geometry. Ablation studies verify that the cross-attention module is crucial for effectively exploiting normal information for accurate depth estimation.

By leveraging cross-attention to fuse predicted surface normals with image features, the proposed method enhances monocular depth estimation accuracy and robustness in colonoscopy. This approach provides a general and lightweight strategy for incorporating geometric priors into self-supervised frameworks, offering potential benefits for downstream tasks such as 3D reconstruction and endoscopic navigation.