Abstract: Objective: Using hepatic MRI proton density fat fraction (PDFF) as the reference standard, we constructed a 3D-ResNet model based on dual-echo Dixon fat-only images to explore the value of deep learning in the grading of fatty liver disease. Materials and Methods: Overall, 132 participants were prospectively enrolled to undergo both dual-echo and multi-echo Dixon liver MRI. Among them, 82 participants were diagnosed with fatty liver disease by MRI-PDFF, and 42 participants were without fatty liver disease. The inter-sequence agreement of fat-only images between the two Dixon protocols was evaluated across hepatic lobes to determine the optimal lobe for region-of-interest (ROI) placement. Circular ROIs were drawn on dual-echo Dixon fat-only images, and the corresponding PDFF values from multi-echo Dixon served as the grading reference. In total, 3,104 ROIs were included: Grade 0 (normal), 1,026; Grade 1 (mild), 1,042; and Grade 2/3 (moderate-to-severe), 1,036. A 3D-ResNet model was developed using dual-echo Dixon fat-only images with hyperparameter optimization and overfitting prevention. The model performance was assessed on the test set via receiver operating characteristic (ROC) and precision-recall (P-R) curves. The values of area under the ROC curve (AUC), average precision (AP), accuracy, precision, recall, specificity, and F1-score were calculated. Results: The following were achieved by the 3D-ResNet model based on dual-echo Dixon fat-only images: a macro-average AUC of 0.971, mean AP of 0.943, accuracy of 88.60%, and weighted F1-score of 88.59% for distinguishing normal, mild, and moderate-to-severe fatty liver. Conclusion: The 3D-ResNet model built on dual-echo Dixon fat-only images demonstrates high accuracy and robustness for grading fatty liver disease and can serve as an effective screening tool for assessing hepatic steatosis.