Diagnostic accuracy of computed tomography (CT)-based radiomics and artificial intelligence (AI) models in...

Hepatocellular carcinoma (HCC) is among the leading causes of cancer-related deaths worldwide. The clinical utility of artificial intelligence (AI) and radiomics models based on contrast-enhanced computed tomography (CECT) for HCC detection remains underexplored. This systematic review and meta-analysis evaluated the diagnostic performance of machine learning and radiomics-based models.

A comprehensive literature search (2015-2025) across four major databases identified 29 eligible studies according to the Preferred Reporting Items for Systematic Reviews and Meta-Analyses guidelines. The risk of bias was assessed using the QUADAS-2 tool. Of these, 17 studies provided independent performance metrics and were included in the meta-analysis. Pooled estimates of sensitivity, specificity, accuracy, and the F1 score were calculated using a random effects model.

Models evaluated on independent test datasets, the pooled sensitivity and specificity were 0.562 and 0.570, respectively. A subgroup of radiomics-specific models (n=3) demonstrated comparatively higher sensitivity (0.660) and specificity (0.670), with minimal heterogeneity in sensitivity (I² = 0%, τ² ≈ 0), suggesting strong internal consistency. However, due to the limited number of studies, these results should be interpreted with caution. Combined models showed lower sensitivity (0.368) and specificity (0.297), with high heterogeneity (I² = 98.3%), likely due to the absence of independent test sets, which may have led to limited generalisability.

Radiomics-based models evaluated on independent datasets exhibit more consistent diagnostic accuracy for HCC detection using CECT. However, methodological variability underscores the need for standardised reporting, robust external validation, and incorporation of explainable AI techniques to enhance clinical adoption.