Implicit domain adaptation with conditional generative adversarial networks for depth prediction in endoscopy.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Gastroenterology

Background:

• Colorectal cancer is a leading cause of death globally, necessitating effective colonoscopy for early detection and treatment.
• Successful colonoscopy relies on accurate navigation and tissue inspection, which are dependent on endoscopist skill.
• Computer-assisted tools can enhance colonoscopy by improving mapping and abnormal tissue detection.

Purpose of the Study:

• To develop a computer-assisted tool for colonoscopy using AI to predict depth from monocular endoscopic images.
• To create a system that aids in colon navigation and polyp size measurement.
• To address the challenge of limited labeled endoscopic data by leveraging simulation and real-world video frames.

Main Methods:

• A conditional generative adversarial network (pix2pix) was trained to transform monocular endoscopic images into depth maps.
• A simulation environment was used to generate synthetic training data.
• The generative model was trained on both synthetic data and unlabeled real colonoscopy video frames for domain adaptation.

Main Results:

• The AI model demonstrated promising performance on synthetic, phantom, and real colonoscopy datasets.
• Generative models showed superior accuracy and robustness in depth prediction compared to discriminative models.
• The model successfully adapted to real colonoscopy environments without explicit domain mapping techniques.

Conclusions:

• The study successfully trained a single AI model for depth prediction from both synthetic and real colonoscopy images.
• Implicit domain adaptation was achieved by training on real unlabeled data, bridging the gap between simulation and reality.
• The developed model shows feasibility for enhancing colonoscopy navigation and tissue analysis in clinical settings.