Effectiveness of Machine Learning in Detecting Vessels Encapsulating Tumor Clusters in Hepatocellular Carcinoma: Systematic Review and Meta-Analysis

TL;DR

This study evaluates how well machine learning can detect tumor-encapsulating vessels in liver cancer, finding moderate accuracy but room for improvement.

Contribution

The first quantitative meta-analysis of machine learning models for detecting vessels encapsulating tumor clusters in hepatocellular carcinoma.

Findings

Nonradiomic models showed moderate accuracy with an SROC AUC of 0.80.

Radiomic models achieved higher accuracy with an SROC AUC of 0.84.

Traditional machine learning outperformed deep learning in sensitivity but not in specificity.

Abstract

Vessels encapsulating tumor clusters (VETC) are significantly associated with poor prognosis in hepatocellular carcinoma (HCC). However, identifying VETC early remains challenging. Recently, machine learning has shown promise for VETC detection, but their diagnostic accuracy lacks systematic validation.

This meta-analysis aimed to systematically evaluate the diagnostic accuracy of machine learning models for detecting VETC in patients with HCC.

The Cochrane Library, Embase, Web of Science, and PubMed were searched up to June 21, 2025. Eligible studies focused on machine learning models for HCC VETC diagnosis. Studies that merely analyzed risk factors or lacked outcome measures were excluded. The Prediction Model Risk of Bias Assessment Tool was used to evaluate the risk of bias. A bivariate mixed-effects model was used for a meta-analysis based on 2×2 diagnostic tables. Subgroup analyses were performed according to modeling variables (nonradiomic vs radiomic features) and model types (traditional machine learning vs deep learning).

This meta-analysis included 31 studies comprising 6755 patients with HCC (2699 VETC-positive). Nineteen studies used machine learning models based on nonradiomic features, and 12 used radiomic features (including deep learning). In the validation set, the nonradiomic model demonstrated a pooled sensitivity of 0.72 (95% CI 0.66-0.78), specificity of 0.74 (95% CI 0.68-0.80), and an area under the summary receiver operating characteristic curve (SROC AUC) of 0.80 (95% CI 0.76-0.83). The radiomic model showed sensitivity of 0.81 (95% CI 0.73-0.87), specificity of 0.73 (95% CI 0.67-0.79), and SROC AUC of 0.84 (95% CI 0.80-0.87). Traditional machine learning achieved sensitivity of 0.84 (95% CI 0.71-0.92), specificity of 0.75 (95% CI 0.67-0.81), and SROC AUC of 0.83 (95% CI 0.80-0.86). Deep learning exhibited sensitivity of 0.77 (95% CI 0.69-0.84), specificity of 0.70 (95% CI 0.59-0.79), and SROC AUC of 0.81 (95% CI 0.77-0.85).

This meta-analysis is the first to quantitatively assess the efficacy of machine learning models in HCC VETC diagnosis, addressing an evidence gap in this field. Unlike previous descriptive reviews, this analysis provides the first quantitative evidence revealing the potential value of machine learning in detecting HCC VETC. The findings provide a foundation for developing and refining subsequent intelligent detection tools. Despite their promising prospects, machine learning models have not yet reached the maturity required for clinical translation, owing to methodological heterogeneity, limited validation, and a high risk of bias. Future research should focus on conducting multicenter, large-sample, standardized, prospective studies to advance clinical translation.

PROSPERO CRD420251084894; https://www.crd.york.ac.uk/PROSPERO/view/CRD420251084894

Figures