Diagnostic Accuracy of Artificial Intelligence-Based Systems for Optical Diagnosis and Characterization of Colorectal Lesions: A Systematic Review and Meta-Analysis

This study aims to comprehensively evaluate the performance of artificial intelligence (AI) systems in the detection and characterization of colorectal cancer (CRC) and its precursor lesions, with a specific focus on the management of sessile serrated lesions (SSLs). To achieve this, an integrated, dual-track methodology was employed. Track A involves a cliometric and network bibliometric analysis using VOSviewer across 34 broad records to accurately map the research landscape and identify knowledge gaps. Track B executes a focused systematic review and Diagnostic Test Accuracy (DTA) meta-analysis on 11 core studies, strictly adhering to PRISMA-DTA guidelines. Risk of bias was assessed by independent reviewers utilizing the standard QUADAS-2 tool, complemented by ROBVIS for visual output. Quantitative analyses, including pooled estimates, meta-regression, heterogeneity evaluation, and subgroup analyses, were conducted in RStudio using the mada and metafor packages. Ultimately, data from all 45 included studies were integrated during the final data triangulation phase. The meta-analysis findings demonstrate the outstanding diagnostic performance of AI, yielding a Summary Receiver Operating Characteristic Area Under the Curve (SROC AUC) of 0.925. Beyond confirming the significant enhancement of diagnostic accuracy by AI algorithms over traditional methods, these results critically address current AI limitations and clinical biases, underscoring the necessity for the development of more generalizable, multicenter models.