Integrating Plain 2D Ultrasound and Clinical Data for Identification of At-Risk MASH: A Multicenter Proof of Concept Study

BACKGROUND: Identifying people with at-risk metabolic dysfunction-associated steatohepatitis (MASH) is a major challenge. We aimed to use deep learning to develop a convolutional neural network (CNN) model to facilitate the detection of at-risk MASH. METHODS: A multi-center cohort study was conducted across five tertiary hospitals in Japan and China. Clinical data and conventional ultrasound images were retrospectively collected. Organ segmentation using YOLO version 11 with quality control was performed, followed by creation of an EfficientNet-B6 CNN regression model, with image-level predictions averaged per patient, to estimate FAST and AGILE 3+ risk categories. Model performance was assessed using agreement across clinically relevant rule-out and rule-in thresholds, R², root mean squared error (RMSE), and area under the receiver operating curve (AUROC). RESULTS: The final dataset comprised 416 patients for training (7643 images), 139 for internal validation (2463 images), and 154 for external testing (3539 images). The CNN model demonstrated good agreement with FAST (R² = 0.684; RMSE = 0.131; category agreement = 0.71) and moderate agreement with AGILE 3+ (R² = 0.445; RMSE = 0.1964; category agreement = 0.63). For FAST, AUROC was 0.974 (95% CI 0.947-0.993) for 0.35 and 0.932 (95% CI 0.884-0.969) for the 0.67 cut-off. For AGILE 3+, AUROC was 0.908 (95% CI 0.854-0.950) for 0.451 and 0.902 (95% CI 0.839-0.953) for the 0.679 cut-off. CONCLUSION: The CNN model applied to clinical ultrasound images can approximate the accuracy of VCTE-based composite risk scores for detecting at-risk MASH. This may support the detection of at-risk MASH in settings with limited access to transient elastography.