Clinical Validation of a Deep Learning-Based 2D Ultrasound Steatosis Algorithm: Cutoff Transferability, Scanner...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Hepatology

Background:

• 2D ultrasound assessment of liver steatosis is subjective.
• A deep learning (DL) algorithm was previously developed for objective steatosis quantification.
• This study aimed to establish histology-based cutoffs, evaluate view transferability, and validate performance on a new scanner.

Purpose of the Study:

• To establish histology-based cutoffs for a DL algorithm in liver steatosis assessment.
• To evaluate the transferability of DL algorithm performance across different ultrasound imaging views.
• To validate the DL algorithm's performance on a novel ultrasound scanner not used in training.

Main Methods:

• Retrospective analysis of 588 ultrasound studies from 457 histology-proven cases.
• Prospective collection of paired scans using a Philips Affiniti 70 scanner.
• Processing images from right intercostal, left hepatic lobe, and subcostal views with the DL algorithm and correlating with histology.

Main Results:

• The DL algorithm achieved high AUROCs (0.891-0.936) across steatosis grades, outperforming FibroScan's controlled attenuation parameter (CAP).
• View-independent transferability showed accuracies of 0.792-0.850 when applying cutoffs from one view to others.
• The algorithm maintained high performance (AUROCs 0.838-0.896) on the new scanner, demonstrating generalizability.

Conclusions:

• The DL algorithm provides accurate, view-independent liver steatosis grading.
• It demonstrates superior performance compared to CAP, especially in moderate-to-severe steatosis.
• The algorithm supports objective, reproducible quantification for real-world clinical application.