Quantitative sonazoid contrast-enhanced ultrasound features for predicting microvascular invasion of hepatocellular...

We developed and validated an interpretable machine learning (ML) model integrating quantitative Sonazoid contrast-enhanced ultrasound (CEUS) and clinical features to predict microvascular invasion (MVI) in hepatocellular carcinoma (HCC).

We retrospectively analyzed 556 histopathologically confirmed HCCs from three Chinese hospitals. Ten ML models were constructed using significant Sonazoid CEUS and clinical features. Model performance was evaluated via the area under the receiver operating characteristic curve (AUC). The combined model (CEUS + clinical) was compared to the clinical model to assess the added predictive value.

The tumor diameter, vitamin K Absence II (PIVKA-II) (≥40 ng/mL), alpha-fetoprotein (AFP) (≥400 ng/mL), ten-minute ratio, and standard deviation of peak intensity were identified as the significant variables for model development. XGBoost was selected as the optimal combined model for accurate MVI prediction in internal validation (AUC = 0.862) and external validation (AUC = 0.841) cohorts. Compared to the clinical model, the combined model resulted in higher AUC values in the internal validation (0.862 vs. 0.631, p = 0.012) and external validation (0.841 vs.0.653, p = 0.004) cohorts.

The XGBoost combined model represents a promising approach for predicting MVI in HCC patients. The superior performance of the combined model to the clinical model highlighted the significant added value of quantitative Sonazoid CEUS features in enhancing MVI predicting accuracy.