Toward autonomous robotic gastroscopy with a novel interventional keypoint and polyp detection system

Area of Science:

• Medical Robotics
• Artificial Intelligence
• Gastroenterology

Background:

• Gastric cancer presents significant global health challenges.
• Traditional gastroscopy relies heavily on endoscopist expertise, limiting accuracy and stability.
• Robotic gastroscopes offer a promising solution for improved upper gastrointestinal disease screening.

Purpose of the Study:

• To develop an AI-driven framework for autonomous robotic gastroscopy.
• To enhance autonomous navigation and diagnostic capabilities in gastroscopy.
• To improve the accuracy and consistency of gastric disease identification.

Main Methods:

• Proposed a deep learning framework integrating keypoint and polyp detection (KP-YOLO).
• Developed and utilized two annotated datasets: real gastroscopy images (3,454) and UGI phantom images (2,144).
• Validated the system using a robotic arm, motion control, and an Olympus endoscopic system on a UGI phantom.

Main Results:

• The detection model achieved high precision (P) and mean average precision (mAP) for intervention keypoints (P=0.987, mAP=0.985).
• Polyp detection performance was also excellent (P=0.994, mAP=0.964).
• Experimental validation confirmed the system's effectiveness on a high-fidelity UGI phantom.

Conclusions:

• The AI-guided robotic gastroscope successfully demonstrated autonomous intervention navigation and intelligent polyp detection.
• This AI framework significantly enhances the autonomy and intelligence of robotic gastroscopy.
• The findings support the potential for future clinical applications of intelligent robotic gastroscopy.