Development of a Deep Learning Model for Classification of Hepatic Steatosis from Clinical Standard Ultrasound

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Hepatology

Background:

• Early detection of hepatic steatosis is crucial for preventing disease advancement.
• Standard grayscale ultrasound (US) is a common diagnostic tool.
• Deep learning (DL) offers potential for automated image analysis.

Purpose of the Study:

• To develop a DL program for classifying hepatic steatosis from US images.
• To differentiate normal liver from steatotic liver.
• To categorize steatosis severity (mild vs. moderate/severe).

Main Methods:

• Retrospective study using 403 grayscale US exams (2010-2022).
• Dataset labeled with magnetic resonance imaging proton density fat fraction (MRI-PDFF).
• Developed DL multi-instance models for binary and multi-class classification.

Main Results:

• The S0/1 vs. S2/3 model achieved 95.9% AUC, demonstrating high accuracy in distinguishing moderate/severe steatosis.
• The S0 vs. S1/2/3 model showed 81.3% AUC for detecting any steatosis.
• Multi-class model provided accurate classification across different steatosis grades.

Conclusions:

• The developed DL program demonstrates high sensitivity and accuracy for hepatic steatosis detection and categorization.
• This AI tool can enhance the utility of standard ultrasound for liver fat assessment.
• Potential for improved early diagnosis and management of fatty liver disease.