Development of a convolutional neural network-based AI-assisted multi-task colonoscopy withdrawal quality control...

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Gastroenterology

Background:

• Colonoscopy is vital for colorectal cancer screening and diagnosis.
• The withdrawal phase is critical for mucosal inspection and lesion detection.
• Current methods lack standardized quality control for the withdrawal phase.

Purpose of the Study:

• To develop an AI system for multitask withdrawal quality control in colonoscopy.
• To monitor withdrawal speed, total withdrawal time, and effective withdrawal time.
• To improve the adequacy of mucosal inspection and lesion detection rates.

Main Methods:

• Utilized a convolutional neural network (YOLOv11) with transfer learning on colonoscopy data.
• Annotated images/videos for ileocecal intubation, instrument operation, and normal mucosa.
• Integrated Laplacian operator for blur detection and perceptual hash algorithm for speed monitoring.

Main Results:

• YOLOv11 m achieved 96.58% weighted average precision and 0.9975 AUC on the test set.
• The AI system demonstrated high consistency with expert endoscopists in measuring effective withdrawal time (ICC = 0.969).
• EWT-SpeedNet provides real-time speed visualization and automatic calculation of withdrawal times.

Conclusions:

• The developed EWT-SpeedNet system effectively supports standardized and efficient colonoscopy withdrawal monitoring.
• AI-driven quality control can enhance the reliability of colonoscopy procedures.
• This system has the potential to improve colorectal cancer detection rates through optimized colonoscopy quality.