Title Computer-aided detection-assisted colonoscopy: classification and relevance of false positives
First author Cesare Hassan (Digestive Endoscopy, Nuovo Regina Margherita Hospital, Rome, Italy)
Journal Gastrointestinal Endoscopy
Pub date June 2020 (online), October 2020 (print)
DOI 10.1016/j.gie.2020.06.021

Abstract

Objective

Evaluate the false positive (FP) images detected by computer-aided detection (CADe) to identify the cause and clinical relevance of the false positives.

Study design

Post-hoc analysis of FPs from images collected in the Artificial Intelligence for Colorectal Adenoma (AID) Study (Repici 2020). The study was a multicenter, randomized study where patients underwent colonoscopy, with and without CADe using GI Genius™ intelligent endoscopy module, for colorectal cancer screening, surveillance, or GI symptoms. 1092 FPs were evaluated in 40 consecutive videos from the CADe arm at one AID Study center.

False positives were adjudicated by a panel of four endoscopists, where each endoscopist described the cause of FP activation by GI Genius™ intelligent endoscopy module and the clinical relevance. >80% agreement on image assessment was required.

Key findings

The mean number of FPs per colonoscopy was 27.3 ±13.1, of which 88% and 12% were due to artifacts in the bowel wall and bowel content, respectively. Of the 27.3 FPs per colonoscopy, 5.7% required additional exploration time of 4.8 ± 6.2 seconds per FP. 

  • The two main domains for the causes of CADe FPs were identified as artifacts due to either the mucosal wall or bowel content, and clinical relevance was defined as the time spent on FPs and the FP rate per minute

Clinical relevance considered time spent by the endoscopists reviewing a FP and the rate of FP per minute (withdrawal time only).

  • Moderate correlation between the number of FPs per colonoscopy and the mean duration of the procedure (p=0.0003, r=0.5, 95% CI 0.2-0.7)
  • Average of 5.7% of FPs per colonoscopy impacted the time of colonoscopy
    • 15 (24.2%) had a mild impact on the time spent by the endoscopist (<1 second)
    • 33 (53.2%) had a moderate impact (between 1 and 3 seconds)
    • 14 (22.6%) had a severe impact (>3 seconds)
  • 2.4 + 1.2 FPs per minute of withdrawal time; statistically higher for FPs due to artifacts of the bowel wall (2.4 +0.6 vs. 0.3 +0.2, p<0.001)

Conclusion

This analysis demonstrated that most CADe activations (false positives) are explained by folds (bowel wall artifacts), and that the additional review time caused by FPs is minimal. 

More ways to learn about AI and connect

Photo of a 3D illustration of a network abstract background.

Artificial intelligence in healthcare

Learn about the world of artificial intelligence (AI). Discover Genius Academy.

Photo of a 3D illustration of a network abstract background.

The future is now — AI in gastroenterology

The benefits of GI Genius™ system introduces new possibilities in the field of endoscopy.

Photo of a 3D illustration of a network abstract background.

Sign up

Subscribe to receive updates on the latest advancements and physician perspectives of AI in healthcare.