Clinical Validation of a Deep Learning-Based 2D Ultrasound Steatosis Algorithm: Cutoff Transferability, Scanner Generalizability, and Comparison with FibroScan

Background: Liver steatosis assessment by 2D ultrasound is widely used but remains subjective. We previously developed a deep learning (DL) algorithm for objective steatosis quantification. This study aimed to (1) establish histology-based cutoffs, (2) evaluate their transferability across different imaging views, and (3) validate performance on a new scanner not included in training. Methods: We retrospectively analyzed 588 ultrasound studies from 457 histology-proven cases and prospectively collected paired scans using a new scanner (Philips Affiniti 70). Images from right intercostal, left hepatic lobe, and subcostal views were processed with the DL algorithm, and mean values from 3–5 images per view were correlated with histology. Results: Across three views, the DL algorithm achieved AUROCs of 0.891–0.936 across steatosis grades, consistently outperforming FibroScan’s controlled attenuation parameter (0.840–0.905), especially in moderate-to-severe steatosis (p < 0.001). Cutoffs established from right intercostal images (N = 565) were applied to images from left hepatic lobe (N = 464) and subcostal views (N = 341), yielding accuracies of 0.792–0.850. On Affiniti 70 images, AUROCs remained high (0.838–0.896), supporting scanner generalizability. Conclusions: The DL algorithm provides accurate, view-independent steatosis grading across different ultrasound scanners and outperforms CAP, supporting its real-world use for objective, reproducible quantification.