A real-time deep learning-based system for colorectal polyp size estimation by white-light endoscopy: development and...

Area of Science:

• Gastroenterology
• Artificial Intelligence
• Medical Imaging

Background:

• Colorectal polyp size is crucial for determining polypectomy device and surveillance intervals.
• Accurate real-time polyp size estimation is needed to guide clinical decisions.

Purpose of the Study:

• To develop and evaluate a deep learning-based system (ENDOANGEL-CPS) for real-time colorectal polyp size estimation.
• To compare the system's performance against endoscopists in size measurement and surveillance recommendations.

Main Methods:

• The ENDOANGEL-CPS system estimates polyp size by calculating the distance from the endoscope lens.
• The depth estimator network was trained on virtual colon videos and tested on simulated and real-world clinical data.
• Performance was evaluated by comparing system estimations with ground truth and endoscopist assessments.

Main Results:

• ENDOANGEL-CPS demonstrated high accuracy in depth estimation (11.3% relative error) and strong concordance with ground truth (CCCs of 0.89-0.93).
• The system significantly outperformed endoscopists in polyp size estimation accuracy (89.9% vs. 54.7%) and concordance (0.89 vs. 0.41).
• ENDOANGEL-CPS showed a lower error rate for inappropriate surveillance recommendations compared to endoscopists (1.5% vs. 16.6%).

Conclusions:

• ENDOANGEL-CPS shows significant potential to enhance the accuracy of colorectal polyp size measurements.
• The system can improve the precision of size-based surveillance interval recommendations.
• AI-driven tools like ENDOANGEL-CPS may optimize colorectal cancer screening and management.