Scale-consistent 3D reconstruction in monocular colonoscopy via camera-intrinsics-guided learning.

Area of Science:

• Medical Imaging
• Computer Vision
• Gastroenterology

Background:

• Accurate 3D spatial understanding of the colon lumen is vital for colonoscopy quality assessment.
• Monocular colonoscopes lack depth sensors, hindering precise scale-consistent depth and camera motion recovery.
• Existing methods face challenges in achieving stable and accurate 3D reconstruction from monocular endoscopic video.

Purpose of the Study:

• To develop a novel framework for scale-consistent 3D reconstruction in monocular colonoscopy.
• To recover accurate depth and camera motion using a sensor-free approach.
• To improve geometry-aware analysis and quality assessment in colonoscopy procedures.

Main Methods:

• A two-step training strategy involving pretraining a depth estimation network on synthetic data with an Adaptive Depth Adjustment module.
• Utilizing an intrinsics-guided scale alignment module to convert synthetic-camera depth predictions to pseudo-depth labels.
• Employing joint depth-pose learning with a geometry-aware 3D loss and a globally consistent optimization strategy for inference.

Main Results:

• Achieved state-of-the-art depth and pose accuracy on the C3VD dataset, significantly outperforming existing baselines.
• Demonstrated anatomically plausible colon lumen geometry reconstruction from clinical videos, indicating stable and interpretable scale recovery.
• Reduced AbsRel by 16.1% and trajectory APE by 86.7% compared to the strongest baseline.

Conclusions:

• The proposed framework offers a sensor-free solution for scale-consistent 3D reconstruction in monocular colonoscopy.
• Enables reliable geometry-aware analysis and quality assessment by providing accurate depth and camera motion.
• Facilitates improved downstream analysis and understanding of colon anatomy from endoscopic procedures.