The Role of Artificial Intelligence in Prospective Real-Time Histological Prediction of Colorectal Lesions during Colonoscopy: A Systematic Review and Meta-Analysis

Abstract

Artificial intelligence (AI) presents a novel platform for improving disease diagnosis. However, the clinical utility of AI remains limited to discovery studies, with poor translation to clinical practice. Current data suggests that 26% of diminutive pre-malignant lesions and 3.5% of colorectal cancers are missed during colonoscopies. The primary aim of this study was to explore the role of artificial intelligence in real-time histological prediction of colorectal lesions during colonoscopy. A systematic search using MeSH headings relating to "AI", "machine learning", "computer-aided", "colonoscopy", and "colon/rectum/colorectal" identified 2290 studies. Thirteen studies reporting real-time analysis were included. A total of 2958 patients with 5908 colorectal lesions were included. A meta-analysis of six studies reporting sensitivities (95% CI) demonstrated that endoscopist diagnosis was superior to a computer-assisted detection platform, although no statistical significance was reached (p = 0.43). AI applications have shown encouraging results in differentiating neoplastic and non-neoplastic lesions using narrow-band imaging, white light imaging, and blue light imaging. Other modalities include autofluorescence imaging and elastic scattering microscopy. The current literature demonstrates that despite the promise of new endoscopic AI models, they remain inferior to expert endoscopist diagnosis. There is a need to focus developments on real-time histological predictions prior to clinical translation to demonstrate improved diagnostic capabilities and time efficiency.

Keywords: artificial intelligence; colonoscopy; colorectal cancer; colorectal polyps; machine learning.

Figure 1

PRISMA flowchart demonstrating article selection.

Figure 2

Forest plots demonstrating the sensitivity (95%) and specificity (95%) across seven studies assessing the diagnostic performance of AI platforms in colonoscopic real-time histological prediction of colorectal lesions [18,19,22,24,25,26,27].

Figure 3

Forest plot demonstrating no significant difference in distinguishing colorectal neoplastic and non-neoplastic lesions when comparing endoscopist diagnosis to CAD output (p = 0.43) [18,19,22,25,26,27]. Green square = sensitivity with 95% CI.