Liver Fat Assessment in Multiview Sonography Using Transfer Learning With Convolutional Neural Networks

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Hepatology

Background:

• Nonalcoholic fatty liver disease (NAFLD) diagnosis relies on accurate fat assessment.
• Ultrasound (US) imaging is a widely accessible modality for liver evaluation.
• Current US-based liver fat assessment methods have limitations.

Purpose of the Study:

• To develop and evaluate deep learning models for liver fat assessment using multiple US liver views.
• To compare the diagnostic performance of individual US views versus an ensemble model.
• To assess the accuracy of deep learning models in diagnosing fatty liver and advanced steatosis.

Main Methods:

• Deep convolutional neural networks (CNNs) were trained using transfer learning on US images from 135 participants.
• Four liver views (transverse and sagittal planes) were analyzed.
• Proton density fat fraction (PDFF) from MRI served as the ground truth for model training and validation.

Main Results:

• The right posterior portal vein view achieved an AUC of 0.90 for diagnosing fatty liver (PDFF ≥ 5%) and an SCC of 0.78 for PDFF quantification.
• An ensemble model integrating all four views demonstrated superior performance with AUCs of 0.91 (PDFF ≥ 5%) and 0.86 (PDFF ≥ 10%), and an SCC of 0.81.
• The ensemble model significantly improved the accuracy of liver fat assessment compared to single views.

Conclusions:

• Deep learning models applied to multi-view US images offer a promising approach for non-invasive liver fat assessment.
• The developed models can aid in the diagnosis and quantification of fatty liver disease.
• Integrating data from multiple US liver views enhances the diagnostic performance of AI-based systems.