AI-Based Multiclass Grading of Hepatic Steatosis From B-Mode Ultrasound: Generalization Across Modalities and Clinical Comparison With Radiologists

Non-alcoholic fatty liver disease (NAFLD) is a growing public health challenge, underscoring the need for scalable, non-invasive tools to grade hepatic steatosis. Although B-mode ultrasound is accessible and safe, its reliability is limited by operator and scanner variability. We present the Deep Domain Adaptation Neural Network (DDANN), a deep learning system for multiclass steatosis classification (Normal, Mild, Moderate, Severe) from ultrasound that emphasizes cross-device generalizability. To mitigate distribution shifts across scanners (LOGIQ, iU22, EPIQ), DDANN combines a MobileNetV2 backbone with triplet loss, entropy-based domain adaptation, and preprocessing that includes speckle suppression, percentile normalization, and LOGIQ-specific harmonization. Trained on a biopsy-confirmed, multi-institutional cohort (primarily LOGIQ and iU22), the model was externally validated on an unseen EPIQ test set of 1,083 images from 47 patients, achieving 98.71% accuracy, 0.9872 macro <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">F</i><sub xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">1</sub>-score, and 0.9998 AUC-ROC, outperforming baselines. In a separate radiologist–AI comparison on 224 biopsy-confirmed images not used for training or validation, the AI reached 91.96% accuracy, significantly exceeding radiologists’ 19.64%–31.70% (McNemar’s test, <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">p</i> < 0.001), with strong agreement to ground truth (κ = 0.893) versus radiologists’ poor-to-slight agreement (κ = 0.006–0.194). The AI maintained balanced class-wise <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">F</i><sub xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">1</sub>-scores (0.90–0.94), while radiologists struggled, particularly with Mild and Moderate cases, and exhibited substantial inter-reader variability (κ = 0.068–0.648). These results demonstrate robust cross-device performance and support integrating AI as a reliable second reader or primary screening tool to reduce subjectivity in steatosis assessment.