Learning-Based Depth and Pose Estimation for Monocular Endoscope with Loss Generalization.

Area of Science:

• Medical Imaging
• Computer Vision
• Gastroenterology

Background:

• Gastroendoscopy is crucial for diagnosing and treating digestive system conditions but lacks 3D perception, hindering navigation and lesion localization.
• Current deep learning methods for monocular gastroendoscopy aim to provide depth and pose information but face generalization challenges.

Purpose of the Study:

• To develop a novel supervised approach for training depth and pose estimation networks using consecutive endoscopy images.
• To improve endoscope navigation and lesion localization within the stomach.

Main Methods:

• Generated real depth and pose training data using a 3D reconstruction pipeline for the stomach.
• Proposed a generalized photometric loss function to balance depth and pose estimation without manual weight tuning.
• Trained networks using consecutive endoscopy images for supervised learning.

Main Results:

• The proposed method successfully generates real depth and pose data, overcoming generalization issues with computer-generated models.
• The novel generalized photometric loss function outperformed existing direct supervision approaches in experiments.
• Demonstrated improved performance in depth and pose estimation for gastroendoscopy.

Conclusions:

• The developed deep learning approach effectively enhances monocular gastroendoscopy with crucial depth and pose information.
• The generalized photometric loss function offers a more robust and simpler alternative for training depth and pose estimation networks.
• This advancement promises to improve the safety and efficacy of endoscopic procedures in the stomach.