Evaluation of the effectiveness of ProScan, an artificial intelligence based lesion detection software, in diagnosing small bowel polypoid lesions with NaviCam SB capsule endoscopy (SBCE) system

Aims The ProScan AI module developed by AnX Robotics was evaluated for its effectiveness indetecting polypoid lesions and tumors in the small bowel using the NaviCam SBCE system. The study aimed to assess detection rates and time efficiency of AI-assisted readings, comparing them to standard human readings and a subsequent consensus reading by experts,which served as the reference standard.

Methods From a database of 1840 NaviCam capsule endoscopy reports, we identified 39 patients with 43 documented polypoid lesions, including 2 small bowel malignant tumors and various benign findings (e.g., benign polyps, adenomas, SETs, lipomas, polypoid lymphangiectasias). Each video was reanalyzed in a two-step process: first by an expert using ProScan assistance and then through a consensus review by two independent experts without AI input. This consensus served as the reference standard for analysis. ProScan-assisted and standard readings were compared to the consensus reference for sensitivity and lesion detection accuracy. Additionally, the average interpretation times for standard and ProScan-assisted readings were recorded and compared. Statistical analyses assessed detection performance and time efficiency, with significance set at p>0.05.

Results The average small bowel transit time was 193 minutes. Using the consensus reading as the reference, ProScan correctly identified 44/47 lesions (94% sensitivity), and missed 3 sub-centimetric polypoid lesions (one lipoma, one small duodenal adenoma and one hyperplasiogen polyp). Standard readings identified 43/47 lesions (91% sensitivity) but missed 4 lesions identified as a real pathology during the consensus reading (all were sub-centimetric hyperplasiogen polypoid lesions). Average interpretation time decreased from 1910 seconds (standard reading) to 110 seconds (ProScan-assisted reading), representing a significant reduction (p<0.001).

Conclusions ProScan demonstrated high sensitivity (94%) in detecting small bowel lesions, with non-inferior diagnostic sensitivity compared to standard expert reading during SBCE. The average reading time was significantly less with ProScan compared to standard reading. The combined use of AI and expert readings yielded the most comprehensive diagnostic results. Incorporating AI-based lesion detection into capsule endoscopy workflows has the potential to improve clinical efficiency and diagnostic accuracy, but a combined approach with expert inaccuracy remains essential to minimize diagnostic gaps.

Conflicts of Interest

Authors do not have any conflict of interest to disclose.

Publication History

Article published online:
27 March 2025