Deep learning in negative small-bowel capsule endoscopy improves small-bowel lesion detection and diagnostic yield

Area of Science:

• Gastroenterology
• Medical Imaging
• Artificial Intelligence

Background:

• Small-bowel capsule endoscopy (SBCE) is crucial for diagnosing obscure gastrointestinal bleeding.
• Previous studies highlight AI's potential in detecting SBCE abnormalities, but clinical utility remains under-explored.
• Human interpretation of SBCE videos can be time-consuming and prone to errors, potentially missing subtle findings.

Purpose of the Study:

• To evaluate the clinical usefulness of a deep convolutional neural network (CNN) model in reanalyzing negative SBCE videos.
• To determine if meaningful findings, such as angioectasias and ulcers, can be detected in SBCE videos initially deemed negative by human readers.
• To assess the impact of CNN-assisted reanalysis on patient diagnosis and subsequent clinical outcomes.

Main Methods:

• Retrospective collection of SBCE videos from patients with suspected small-bowel bleeding at two academic hospitals.
• Reanalysis of SBCE videos initially classified as negative using a previously developed CNN algorithm.
• Independent review of CNN-identified images by two gastroenterologists to confirm meaningful findings like angioectasias and ulcers.

Main Results:

• Out of 202 SBCE videos, 103 (51.0%) were initially negative. The CNN model detected meaningful findings in 63 (61.2%) of these negative videos.
• Significant findings included 79 angioectasias in 40 videos and 66 ulcers in 35 videos.
• Reanalysis led to a diagnostic change in 10.3% of patients, with no statistically significant difference in rebleeding rates between groups with and without CNN-detected findings.

Conclusions:

• The developed CNN algorithm successfully identified clinically relevant findings in SBCE videos that were overlooked during initial human interpretation.
• Deep CNN analysis of SBCE images can serve as a valuable tool to augment human review and potentially reduce diagnostic errors.
• This AI-driven approach holds promise for enhancing the diagnostic yield of SBCE, particularly in cases of suspected small-bowel bleeding.