MRI-Based T2 and Arterial Radiomics to Differentiate Focal Nodular Hyperplasia (FNH) from Hepatocellular Adenoma (HCA): A High-Accuracy Spleen-Referenced Approach

Background Focal Nodular Hyperplasia (FNH) and Hepatocellular Adenoma (HCA) often present overlapping imaging features on MRI. Recent radiomics and machine learning (ML) strategies may enhance differentiation accuracy. Methods This retrospective study involved 197 lesions (77 FNH; 120 HCA, including 87 non-steatotic HCAs) from 98 patients (mean age 33.95 ± 9.28 years, 87.7% female). Multiparametric MRI included T2-weighted, arterial-phase, and diffusion-weighted (b-50, b-400, b-800) sequences. Over 100 features, including mass-to-spleen T2 ratios and spleen-referenced diffusion metrics, were extracted. Recursive feature elimination identified the most discriminative variables. Eight ML models were trained and evaluated using accuracy, F1-score, and AUC-ROC on separate training/testing sets. Results FNH lesions were significantly larger than HCA (48.95 ± 26.64 mm vs. 36.35 ± 28.42 mm; p = 0.002). The T2-weighted mass-to-spleen signal ratio effectively distinguished FNH from HCA/NSHCA (p < 0.001). DWI revealed higher signal intensities in FNH at b-50 and b-400 (402 ± 275, 222 ± 139) compared to HCA (273 ± 265, 166 ± 166; p = 0.002), and consistently greater mean mass-to-liver SIDs for b-50, b-400, and b-800 (all p < 0.001). Gradient Boosting Machine (GBM) and Random Forest demonstrated the highest test-set performance (AUC-ROC = 0.915), with GBM achieving 86.0% accuracy; a standalone decision tree classifier achieved 87.5% accuracy on training and 81.1% on testing sets. Across models, T2 lesion intensity, arterial phase signal ratio, and spleen-referenced DWI metrics ranked as the top three predictive features. Conclusion Integrating T2 mass-to-spleen ratios, arterial-phase enhancement, and advanced DWI significantly improves the differentiation of FNH from HCA, including non-steatotic variants. Ensemble ML methods outperformed simpler decision tree classifiers and underscore the importance of biologically aligned feature selection for robust lesion characterization.