AI-assisted capsule endoscopy reading in suspected small bowel bleeding: a multicentre prospective study

Area of Science:

• Gastroenterology
• Medical Imaging
• Artificial Intelligence

Background:

• Capsule endoscopy reading is time-intensive and requires sustained reader attention.
• Deep convolutional neural networks show potential to exceed human performance in identifying capsule endoscopy findings.
• Reducing reading time and improving accuracy are key goals in capsule endoscopy analysis.

Purpose of the Study:

• To assess the non-inferiority of artificial intelligence (AI)-assisted capsule endoscopy reading versus standard reading for small bowel bleeding lesions.
• To evaluate the diagnostic yield of AI-assisted reading for P1 and P2 small bowel bleeding lesions.
• To compare the mean reading time between AI-assisted and standard capsule endoscopy reading modalities.

Main Methods:

• Prospective enrollment of 133 patients (aged ≥18) with suspected small bowel bleeding across 14 European centers.
• Patients underwent capsule endoscopy using the Navicam SB system with an integrated AI (ProScan) for lesion detection.
• Blinded reading was performed in standard mode, followed by AI-assisted reading where AI performed the initial automated read.

Main Results:

• AI-assisted reading demonstrated non-inferiority and superiority (73.7% vs 62.4%) in the per-patient diagnostic yield of P1 and P2 lesions.
• The mean reading time was significantly reduced with AI-assisted reading (3.8 min) compared to standard reading (33.7 min).
• 133 patients were analyzed, with 112 completing the capsule endoscopy procedure.

Conclusions:

• AI-assisted capsule endoscopy reading offers a more accurate detection of small bowel bleeding lesions.
• AI significantly reduces the time required for capsule endoscopy interpretation.
• AI-assisted reading presents a promising advancement for diagnosing small bowel bleeding.