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

Area of Science:

• Medical Imaging
• Computer Vision
• Deep Learning

Background:

• Accurate 3D reconstruction in colonoscopy aids navigation and polyp assessment.
• Challenges in depth estimation include specular highlights and low-texture areas.

Purpose of the Study:

• To improve monocular depth estimation in colonoscopy by incorporating surface normal information.
• To enhance 3D reconstruction and lesion assessment capabilities.

Main Methods:

• A self-supervised deep learning method using cross-attention to integrate surface normal information.
• Surface normals are predicted and used as geometric priors.
• Cross-attention adaptively fuses normal and image features for geometric refinement.

Main Results:

• Outperforms state-of-the-art self-supervised methods on benchmark datasets and real colonoscopy data.
• Generates realistic depth maps, preserving colon geometry and mucosal folds.
• Ablation studies confirm the effectiveness of the cross-attention module.

Conclusions:

• The proposed method enhances monocular depth estimation accuracy and robustness in colonoscopy.
• Offers a lightweight strategy for incorporating geometric priors into self-supervised frameworks.
• Potential benefits for 3D reconstruction, endoscopic navigation, and lesion assessment.