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

Gastric cancer has high incidence and mortality rates globally. Gastroscopy is widely considered the gold standard for screening gastric diseases. However, accurate and stable identification of gastric diseases under endoscopy remains challenging due to its heavy reliance on the experience and operational skills of the endoscopist. The development of robotic gastroscopes offers an effective solution for screening upper gastrointestinal diseases.

In this paper, we propose an innovative deep learning-based framework for autonomous robotic gastroscopy, integrating a combined anchor point (intervention keypoint) and polyp detection system (named KP-YOLO). This system enables the robotic gastroscope to operate autonomously and intelligently, facilitating both intervention navigation and diagnostic tasks. To support this framework, we have created and made publicly available two fully annotated datasets: a real gastroscopy dataset containing 3,454 images and a high-fidelity upper gastrointestinal (UGI) phantom dataset with 2,144 images.

The detection model trained on the real gastroscopy dataset achieved a precision (P) of 0.987 and a mean average precision (mAP) of 0.985 for intervention keypoints. For polyp detection, the model attained a P of 0.994 and an mAP of 0.964. Experimental validation was conducted utilizing a high-fidelity UGI phantom equipped with a robotic arm, motion control module, and Olympus endoscopic system.

Guided by the proposed detection system, the robotic gastroscope successfully performs autonomous intervention navigation and intelligent polyp detection on the UGI phantom. This work demonstrates the potential of leveraging artificial intelligence to enhance the autonomy and intelligence of robotic gastroscopy, paving the way for its future clinical application.