A systematic review and meta-analysis of artificial intelligence–assisted colon capsule endoscopy for colorectal polyp detection.

Abstract

e15520 Background: Colorectal cancer remains a leading cause of cancer-related mortality worldwide, with early polyp detection crucial for prevention. Colon capsule endoscopy (CCE) offers a non-invasive alternative to traditional colonoscopy, but diagnostic accuracy is limited by the time-consuming image review and human error. Artificial intelligence (AI) has emerged as a promising tool to enhance CCE performance. This systematic review and meta-analysis aimed to evaluate the diagnostic accuracy of AI-assisted CCE for detecting colorectal polyps. Methods: We conducted a systematic review and diagnostic meta-analysis of studies evaluating AI-assisted CCE for colorectal polyp detection. Five studies published between 2020 and 2025 were included (Yamada 2020, Saraiva 2021, Gilabert 2022, Mascarenhas 2022, and Nadimi 2025). All employed retrospective designs using deep learning algorithms, primarily convolutional neural networks (CNNs), trained on CCE video frames. Primary outcomes were pooled sensitivity and specificity. Secondary outcomes included diagnostic accuracy, positive predictive value (PPV), negative predictive value (NPV), and area under the receiver operating characteristic curve (AUC). Random-effects models with the DerSimonian-Laird method were applied, and heterogeneity was assessed using the I² statistic. Results: Five studies comprising 3,897 polyp-positive frames were analyzed. The pooled sensitivity for AI-assisted CCE in detecting colorectal polyps was 93.44% (95% CI: 76.41%-98.43%), with high heterogeneity (I² = 98.1%). The pooled specificity was 94.63% (95% CI: 84.87%-98.22%, I² = 96.3%) based on four studies reporting this metric. Overall diagnostic accuracy was 94.97% (95% CI: 82.44%-98.70%, I² = 97.3%). Individual study sensitivities ranged from 79.0% (Yamada 2020) to 99.9% (Nadimi 2025), while specificities ranged from 79.2% (Saraiva 2021) to 99.4% (Nadimi 2025). Among studies reporting AUC values, performance ranged from 0.80 to 0.99. The high heterogeneity observed across all metrics suggests substantial variability in study populations, AI algorithms, and methodological approaches. Conclusions: AI-assisted CCE demonstrates excellent diagnostic performance for colorectal polyp detection, with pooled sensitivity and specificity both exceeding 93%. These findings suggest that AI integration can significantly enhance the diagnostic capabilities of CCE, potentially making it a viable alternative to conventional colonoscopy for colorectal cancer screening. High negative predictive values (86.8%-99.8%) indicate that AI-assisted CCE is effective at ruling out polyps when the test is negative, which is crucial for screening applications. Future research should focus on prospective validation studies and standardization of AI reporting to facilitate clinical implementation.