Article Text
PDF
Abstract
Background In recent years, several artificial intelligence (AI) based computer-aided diagnosis (CADx) tools have been developed to help endoscopists distinguish neoplastic polyps from nonneoplastic ones. These tools have attracted endoscopists’ attention, as their potentially high accuracy could be leveraged for supporting cost-saving strategies and guiding novice endoscopists. However, a summary of the effectiveness of such CADx tools has not yet been given, nor has it been compared with that of endoscopists in a clinical setting. This study aims to compare the overall clinical performances of AI-powered CADx tools in the optical diagnosis of diminutive colorectal lesions (≤5mm) with those of endoscopists.
Methods We did a systematic search in databases EMBASE, MEDLINE, PubMed and Cochrane until 19 March 2024, for AI-based optical diagnosis of neoplastic colorectal polyps, and limited search results to the prospective clinical trials. The sensitivity, specificity, and accuracy of AI and endoscopists in the diagnosis of diminutive colorectal polyps and a number of polyps diagnosed were extracted and/or transformed for further analysis. Univariate random-effects meta-analysis models were used to calculate the pooled diagnostic odds ratios, due to a small number of studies concerning AI-assisted endoscopists. Meta-regressions were conducted to compare different AI tools as well as the performances of AI and endoscopists.
Results Among 631 articles retrieved, only 11 studies were included (IDDF2024-ABS-0373 Table 1). AI alone, AI-assisted endoscopists, and endoscopists alone achieved overall diagnostic odds ratios of 29.22 (16.64-51.33), 29.44 (9.50-91.23) and 29.11 (16.15-52.46) respectively, indicating there was no significant difference among diagnostic performances of these groups (IDDF2024-ABS-0373 Figure 1. Pooled diagnostic odds ratio for AI alone in diagnosis of diminutive polyps, IDDF2024-ABS-0373 Figure 2. Pooled diagnostic odds ratio of AI-assisted endoscopists in diagnosis of diminutive polyps, IDDF2024-ABS-0373 Figure 3. Pooled diagnostic odds ratio of endoscopists alone in characterization of diminutive polyps). Results from meta-regression showed (i) the endocytoscope-based AI had significantly better diagnostic performances than the blue-light-based AI (p=0.030), while the white-light-based AI did not (p=0.193); (ii) AI and AI-assisted endoscopists did not significantly outperform endoscopists (p=0.973; p=0.995).
Baseline characteristics of included articles
Pooled diagnostic odds ratio for AI alone in diagnosis of diminutive polyps.
Pooled diagnostic odds ratio of AI-assisted endoscopists in diagnosis of diminutive polyps.
Pooled diagnostic odds ratio of endoscopists alone in characterization of diminutive polyps.
Conclusions The findings indicate, overall, the performances of AI-powered real-time optical diagnoses in characterisation of diminutive colorectal lesions are comparable to that of endoscopists. Future studies to evaluate the cost-effectiveness of using AI in the characterisation of diminutive polyps during colonoscopy are warranted.