Preoperative Diagnosis of Dual‐Phenotype Hepatocellular Carcinoma Using Enhanced MRI Radiomics Models.

Background: Dual‐phenotype hepatocellular carcinoma (DPHCC) is highly aggressive and difficult to distinguish from hepatocellular carcinoma (HCC). Purpose: To develop and validate clinical and radiomics models based on contrast‐enhanced MRI for the preoperative diagnosis of DPHCC. Study type: Retrospective. Population: A total of 87 patients with DPHCC and 92 patients with non‐DPHCC randomly divided into a training cohort (n = 125: 64 non‐DPHCC; 61 DPHCC) and a validation cohort (n = 54: 28 non‐DPHCC; 26 DPHCC). Field Strength/Sequence: A 3.0 T; dynamic contrast‐enhanced MRI with time‐resolved T1‐weighted imaging sequence. Assessment: In the clinical model, the maximum tumor diameter and hepatitis B virus (HBV) were independent risk factors of DPHCC. In the radiomics model, a total of 1781 radiomics features were extracted from tumor volumes of interest (VOIs) in the arterial phase (AP) and portal venous phase (PP) images. For feature reduction and selection, Pearson correlation coefficient (PCC) and recursive feature elimination (RFE) were used. Clinical, AP, PP, and combined radiomics models were established using machine learning algorithms (support vector machine [SVM], logistic regression [LR], and logistic regression‐least absolute shrinkage and selection operator [LR‐LASSO]) and their discriminatory efficacy assessed and compared. Statistical Tests: The independent sample t test, Mann–Whitney U test, Chi‐square test, regression analysis, receiver operating characteristic curve (ROC) analysis, Pearson correlation analysis, the Delong test. A P value < 0.05 was considered statistically significant. Results: In the validation cohort, the combined radiomics model (area under the curve [AUC] = 0.908, 95% confidence interval [CI]: 0.831–0.985) showed the highest diagnostic performance. The AUCs of the PP (AUC = 0.879, 95% CI: 0.779–0.979) and combined radiomics models were significantly higher than that of clinical model (AUC = 0.685, 95% CI: 0.526–0.844). There were no significant differences in AUC between AP or PP radiomics model and combined radiomics model (P = 0.286, 0.180 and 0.543). Conclusion: MRI radiomics models may be useful for discriminating DPHCC from non‐DPHCC before surgery. Evidence Level: 4 Technical Efficacy: Stage 2 [ABSTRACT FROM AUTHOR]