Noninvasive identification of proliferative hepatocellular carcinoma based on CEUS quantitative morphological feature

To investigate the diagnostic performance of the quantitative morphological feature of solidity in preoperative prediction of proliferative hepatocellular carcinoma (HCC), and to compare the diagnostic performance between quantitative morphological features and machine learning (ML) models that incorporate Sonazoid contrast-enhanced ultrasound (CEUS) and clinical features.

This retrospective two-center study included 395 patients with histopathologically confirmed HCC. The morphological feature of solidity, along with clinical and CEUS features, was analyzed to predict the proliferative status of HCC. Eight ML models were trained and validated using area under the curve (AUC), calibration, and decision curve analysis (DCA). SHAP interpretability tools were used to elucidate feature contributions.

Solidity emerged as the strongest independent predictor of proliferative HCC with AUC of 0.836 and 0.768 in the two-center cohorts, respectively. The LightGBM model, which integratedsolidity, AFP ≥ 400 ng/mL, ratio of neutrophils to lymphocytes (N/L), ten-min ratio, and standard deviation (StdDev) of lesion, achieved superior performance, with AUCs of 0.887 (95% CI: 0.803-0.971) and 0.871 (0.791-0.948) in internal and external validation, respectively, significantly surpassingsolidity(P = 0.0026 and 0.0067) and LightGBM-4F (P = 0.0004 and 0.033). Robustness was confirmed via 10-fold cross-validation (mean AUC = 0.897). Calibration curves and DCA confirmed the clinical utility across cohorts. SHAP analysis highlightedsolidity(mean impact = 1.38) as the dominant predictor, followed by AFP (mean = 0.7).

Our interpretable ML model leverages quantitative CEUS features, spearheaded by the morphological biomarker solidity, to preoperatively predict HCC proliferative status. This enables noninvasive risk stratification, facilitating precision treatment planning.