Diagnostic accuracy of CADe-assisted reading versus clinician reading for polyp detection in colon capsule endoscopy: a multicentre prospective study
Background: Colon capsule endoscopy (CCE) offers a non-invasive alternative to colonoscopy for lower gastrointestinal (GI) investigations. However, its adoption is limited by lengthy reading times, reader fatigue and variable diagnostic accuracy. These factors are often influenced by bowel preparati...
Saved in:
| Main Authors: | , , , , , , , , , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
Elsevier
2025-07-01
|
| Series: | Clinical Medicine |
| Online Access: | http://www.sciencedirect.com/science/article/pii/S1470211825000995 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | Background: Colon capsule endoscopy (CCE) offers a non-invasive alternative to colonoscopy for lower gastrointestinal (GI) investigations. However, its adoption is limited by lengthy reading times, reader fatigue and variable diagnostic accuracy. These factors are often influenced by bowel preparation quality and completion rates. Artificial intelligence (AI) has shown the potential to improve the diagnostic accuracy of polyp detection and reduce reading times in capsule endoscopy.1–3 Objectives: The primary aim was to assess the diagnostic accuracy of Computer-Aided Detection (CADe) system (AiSPEED™)-assisted CCE reading using full-video analysis in a real-world clinical setting. Secondary outcome was to assess non-inferiority compared with standard clinician reading in detecting polypoid lesions and mean reading time, as well as per-polyp diagnostic performance, including assessment by polyp size, morphology and location. Methods: The CESCAIL study was a multi-centre study of 14 UK centres that recruited 720 participants between November 2021 and September 2024. Participants (prospectively and retrospectively enrolled) included those ≥18 years referred for CCE under the urgent cancer pathway or for post-polypectomy surveillance. CCEs were performed using PillCam™ COLON 2 and analysed using AiSPEED™. CCE videos were evaluated by clinicians, who provided standard readings, and AI-assisted reading, which included an AI-assisted pre-reader review and subsequent clinician validation.Per-patient and per-polyp analyses were performed. The per-polyp sub-analysis evaluated 1,803 polyps, examining sensitivity, positive predictive value and the impact of different factors on AI-assisted accuracy performance. Results: 720 patients were enrolled, with 673 included in the final analysis. The capsule excretion rate before battery depletion was 77.1%, with adequate bowel preparation achieved at 78.1% (with standard) and 74.0% (with AI-assisted reading) (McNemar p=0.108) (Fig 1). In the per-patient analysis, the polyp diagnostic yield was 63.6% (428/673) for standard reading and 70.6% (475/673) for AI-assisted reading (Fig 2), using the expert panel as the reference standard. These findings confirmed both non-inferiority (p<0.001) and superiority (p<0.001) of the AI-assisted reading. The diagnostic accuracy for detecting polyps requiring follow-up procedure was 0.96 (95% CI: 0.95–0.98) for AI-assisted reading, compared with 0.91 (95% CI: 0.89–0.93) for standard reading (p<0.0001). The mean reading time per video was 8.7 min for AI-assisted reading compared with 47.3 min for standard reading, representing a 5.4-fold reduction.Per-polyp analysis revealed that AI-assisted reading was significantly more sensitive than standard reading for polyps <10 mm (OR 2.27 vs 0.88, p<0.001), although no difference was seen for polyps ≥10 mm, where non-inferiority was maintained. The most marked improvement was in the detection of hyperplastic polyps (OR 5.4, p<0.001). No significant differences were observed for pedunculated, sub-pedunculated, lateral spreading tumours, or sessile serrated lesions. AI-assisted readings were notably more accurate for left-sided polyps compared with right-sided ones (OR 2.36 vs 1.66, p<0.0001). Conclusion: AI-assisted reading with AiSPEED™ offers the potential for greater accuracy in polyp detection and significantly reduced clinician reading times during CCE examination compared with standard reading methods. Performance gains were most evident for smaller adenomas, hyperplastic polyps and left-sided lesions. These findings support AI integration into clinical CCE workflows and highlight the need for next-generation AI tools to refine polyp characterisation, particularly for right-sided and diminutive lesions. |
|---|---|
| ISSN: | 1470-2118 |