Comparison of machine learning models and CEUS LI-RADS in differentiation of hepatic carcinoma and liver metastases in patients at risk of both hepatitis and extrahepatic malignancy

Table 3 Comparison of diagnostic performance of the different criteria for HCC/LM in the validation cohorts

Diagnostic criteria	Sensitivity	Specificity	PPV	NPV	Accuracy
LI-RADS	68.2%	88.6%	85.7%	26.4%	0.784
Gradient Boosting Model	75.0%*	86.4%	83.9%	28.1%	0.807
Random Forest	79.5%*	86.4%	85.4%	19.1%*	0.830*
General Linear Model	77.3%*	88.6%	87.2%	20.4%*	0.830*

Machine Learning (GBM, RF and GLM): based on Arterial phase enhancement, Homogeneity, Washout type II, Unclearly border and Rim enhancement
GBM: Gradient Boosting Model; RF: Random Forest; GLM: General Linear Model
^*There was statistical difference compared with LI-RADS (Two-sided P-values < 0.05)

ISSN: 1470-7330