Radiomics-Based AI for the Diagnosis and Prognosis of Vessels Encapsulating Tumor Clusters in Hepatocellular Carcinoma: Systematic Review and Meta-Analysis.

Authors

Abstract

Vessels encapsulating tumor clusters (VETCs), a CD34-positive vascular pattern in hepatocellular carcinoma (HCC), are linked to aggressive biology, early recurrence, and poor survival. Because pathologic VETC assessment requires postoperative immunohistochemistry and may be affected by sampling, preoperative noninvasive prediction remains clinically important. Radiomics-based artificial intelligence (AI) applied to routine contrast-enhanced imaging may provide a surrogate marker, but evidence across has not been comprehensively appraised. This study aimed to evaluate the diagnostic accuracy and prognostic value of radiomics-based AI models for noninvasive VETC prediction in HCC using PICOTS (patient population, intervention, comparator, outcomes, timing, and setting) and PRISMA-DTA (Preferred Reporting Items for Systematic Reviews and Meta-Analyses for Diagnostic Test Accuracy) frameworks. We searched PubMed, Embase, Web of Science, and the Cochrane Library, gray literature, and citations (original search July 11, 2025; updated April 17, 2026). The original search was completed on July 11, 2025, and the reconstructed strategy was rerun on April 17, 2026. Eligible retrospective cohort studies developed or validated radiomics or deep-learning models using contrast-enhanced magnetic resonance imaging (CEMRI), contrast-enhanced computed tomography (CECT), contrast-enhanced ultrasound (CEUS), or 2-[¹⁸F]fluoro-2-deoxy-D-glucose positron emission tomography or computed tomography ([18F]FDG PET/CT) to predict CD34-confirmed VETC and reported 2×2 diagnostic data and/or hazard ratios (HRs) for early recurrence. Mutually exclusive cohorts were treated as separate datasets only when patient overlap was absent. Risk of bias was assessed with the Prediction model Risk Of Bias Assessment Tool+AI, and certainty with GRADE (Grading of Recommendations, Assessment, Development, and Evaluation). Diagnostic accuracy was synthesized with bivariate random-effects models; prognostic HRs were pooled with restricted maximum likelihood+Hartung-Knapp-Sidik-Jonkman random-effects models. In total, 15 studies (729 internal-validation and 613 external-validation patients) were included; 14 were from China and 1 from Japan. Moreover, 10 studies used CEMRI, 3 CECT, 1 CEUS, and 1 [18F]FDG PET/CT. CEMRI-based AI showed the best performance: sensitivity=0.84 (95% CI 0.73-0.93; 95% prediction interval [PI] 0.45-1.00), specificity=0.79 (95% CI 0.70-0.86; 95% PI 0.50-0.97), and area under the curve=0.87. Meta-regression suggested that center type, validation design, algorithm class, and magnetic resonance imaging field strength contributed to specificity heterogeneity. CECT, CEUS, and [18F]FDG PET/CT evidence was limited. AI-predicted VETC positivity was associated with early recurrence (HR 2.34, 95% CI 1.93-2.84). GRADE certainty ranged from low to moderate, mainly due to imprecision, risk of bias, and heterogeneity. This review is innovative because it integrates diagnostic accuracy, modality comparison, algorithm performance, and recurrence prognosis for AI-based VETC prediction. Unlike previous modality-specific reviews, it clarifies what AI brings to the field: a potential preoperative bridge between imaging phenotypes and biologically aggressive HCC. Real-world use should remain cautious and decision-supportive, given retrospective designs, geographically concentrated cohorts, limited external validation, heterogeneity, risk of bias, and low-to-moderate GRADE certainty, rather than replacing histopathology or multidisciplinary clinical judgment in practice.

Tags

Topics