JMIR Publications

ABSTRACT

Background:

Early clinical identification of Vessels Encapsulating Tumor Clusters (VETC) in hepatocellular carcinoma (HCC) remains challenging. Although machine learning (ML) has shown promise for VETC detection, its diagnostic accuracy lacks systematic validation.

Objective:

Thus, this meta-analysis aimed to systematically assess the performance of machine learning models (MLMs) in detecting VETC.

Methods:

The Cochrane Library, Embase, Web of Science, and PubMed were systematically searched up to June 21, 2025. Included studies were assessed using the Prediction Model ROB Assessment Tool for risk of bias. Subgroup analyses were performed according to modeling variables (non-radiomic vs. radiomic features) and model types (traditional ML vs. deep learning [DL]) during meta-analysis.

Results:

This meta-analysis encompassed 31 studies with 6,755 HCC patients, including 2,699 VETC-positive cases. Nineteen studies utilized ML based on non-radiomic features, while 12 employed radiomic features (including DL). In the validation cohorts, the pooled estimates (95% confidence interval [95% CI]) of non-radiomic MLMs were: sensitivity 0.72 (0.66-0.78), specificity 0.74 (0.68-0.80). For radiomic MLMs, the pooled estimates were: sensitivity 0.81 (0.73-0.87), specificity 0.73 (0.67-0.79). For traditional MLMs, the pooled estimates were: sensitivity 0.84 (0.71-0.92), specificity 0.75 (0.67-0.81). For deep learning models, the pooled estimates were: sensitivity 0.77 (0.69-0.84), specificity 0.70 (0.59-0.79).

Conclusions:

ML demonstrates feasible performance in detecting VETC status in HCC. Radiomic MLMs significantly outperformed non-radiomic approaches. Within radiomic MLMs, although DL enables efficient processing of imaging data, no substantial performance advantage over traditional ML was observed.