Evaluation of Net Withdrawal Time and Colonoscopy Video Summarization Using Deep Learning Based Automated Temporal...

Area of Science:

• Gastroenterology
• Medical Imaging
• Artificial Intelligence

Background:

• Adequate colonoscopy withdrawal time is vital for polyp detection but traditional measurements are often inaccurate.
• Non-observation activities during colonoscopy can bias withdrawal time assessments, compromising procedural quality evaluation.

Purpose of the Study:

• To develop a deep learning (DL) model for accurate net withdrawal time measurement in colonoscopy.
• To create a DL model that excludes non-observation phases and provides quantitative visual summaries of procedural events.

Main Methods:

• A DL-based automated temporal video segmentation model was developed and trained on colonoscopy videos.
• The model classifies key events: cecum, intervention, outside, and narrow-band imaging (NBI) mode.
• Net withdrawal time was calculated, and representative images were extracted for video summarization.

Main Results:

• The DL model achieved over 93% F1 score for temporal video segmentation in both internal and external tests.
• Net withdrawal time demonstrated a strong correlation with endoscopist-recorded times (r > 0.97, p < 0.000).
• Generated representative images accurately summarized key procedural events, confirmed by endoscopist assessment.

Conclusions:

• The DL model provides an efficient, standardized, and objective method for assessing colonoscopy procedural quality.
• This AI tool has the potential to significantly enhance clinical practice and quality assurance in colonoscopy.
• Accurate net withdrawal time measurement can improve polyp detection rates and patient care.