Machine Learning-Based Quantitative Prediction of Hepatic Steatosis Using Ultrasound Signal Attenuation: A Validation Study with MRI-PDFF

Background/Objectives: This study aimed to develop and evaluate a prediction model for fatty liver (hepatic steatosis) using ultrasound-derived quantitative variables, with MRI DIXON-based fat fraction (MRI-FF) as the reference standard. Methods: Twenty-seven participants with above-normal BMI underwent ultrasound and MRI examinations con-currently. Ultrasound images of the liver, kidney, and spleen were acquired at dynamic range (DR) settings of 100, 150, and 200. Quantitative variables (signal intensity, linear slope, exponential attenuation coefficient, and R²) were extracted using ImageJ. Key varia-bles were selected via principal component analysis (PCA) and orthogonal partial least squares discriminant analysis (OPLS-DA) with VIP scores ≥1.25. A support vector ma-chine (SVM) model was constructed using training (n=21) and validation (n=6) datasets. Results: PCA and OPLS-DA revealed that the liver-to-kidney attenuation ratio, liver at-tenuation R² (DR200), and linear slope R² (DR200) correlated most strongly with MRI-FF (r=0.814, 0.753, 0.724; all p < 0.001). Attenuation variables were significantly higher in fatty liver groups across all MRI-FF thresholds. The SVM model demonstrated excellent predic-tive performance (RMSE=2.1997, r=0.82, p < 0.001). Conclusions: Ultrasound-derived sig-nal attenuation characteristics correlate strongly with MRI-FF, enabling accurate quantita-tive assessment of hepatic steatosis through machine learning. This noninvasive, cost-effective approach shows significant potential for screening and longitudinal moni-toring of fatty liver disease.