USLF-Net: A Task-Specific CNN for Liver Fibrosis Classification from Ultrasound with Mobile Clinical Integration

Abstract

Accurate liver fibrosis classification is crucial for making
treatment decisions in patients with chronic liver disease. Liver biopsy
and advanced noninvasive imaging modalities, including elastography,
computed tomography (CT), and magnetic resonance imaging (MRI),
remain challenges in resource-limited settings due to their high cost and
specialized equipment requirements. Consequently, conventional B-mode
ultrasound is widely available and cost-effective. Despite the considerable
promise of deep learning in medical imaging, prior works often have high
computational demands, which limit their practicality. In this study, we
introduce USLF-Net, a novel UltraSound-Based Liver Fibrosis Clas-
sification using Convolutional Neural Network (CNN). Our model was
validated using a comprehensive dataset comprising 6,323 ultrasound im-
ages, stratified across five fibrosis stages (F0-F4) according to the Meta-
analysis of Histological Data in Viral Hepatitis (METAVIR) scoring sys-
tem. USLF-Net achieved 97.64% classification accuracy, demonstrating
superior performance compared to recent transfer learning approaches.
We further developed a clinical application for automated liver fibrosis
diagnosis to enhance diagnostic capabilities in healthcare settings