Hybrid Deep Learning and Machine Learning for Detecting Hepatocyte Ballooning in Liver Ultrasound Images

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Hepatology

Background:

• Hepatocyte ballooning (HB) is a key indicator of non-alcoholic fatty liver disease (NAFLD) and non-alcoholic steatohepatitis (NASH) progression.
• Liver biopsy, the current gold standard for HB detection, is invasive and carries risks, necessitating non-invasive diagnostic tools.

Purpose of the Study:

• To develop and validate a novel methodology integrating deep learning and machine learning for accurate identification and quantification of hepatobiliary abnormalities in liver ultrasound images.
• To establish a non-invasive diagnostic approach for hepatocyte ballooning, reducing reliance on liver biopsies.

Main Methods:

• Training deep convolutional neural networks (CNNs) including InceptionV3, ResNet50, DenseNet121, and EfficientNetB0 on an expanded dataset of liver ultrasound images.
• Implementing a hybrid approach combining InceptionV3 for feature extraction with a Random Forest classifier for enhanced accuracy and stability.
• Utilizing a dual dichotomy classification strategy to categorize images into healthy vs. diseased and subsequently mild vs. severe ballooning stages.

Main Results:

• The hybrid InceptionV3-Random Forest model achieved high performance, with 97.40% accuracy and a 0.99 area under the curve (AUC).
• The model demonstrated excellent sensitivity (99%) for detecting the 'Many' class (severe HB) in the third evaluation phase.
• Dual dichotomy classification significantly improved sensitivity for identifying severe hepatocyte ballooning, with cross-validation confirming model robustness.

Conclusions:

• The developed AI-driven methodology offers a non-invasive, accurate alternative for the early detection and monitoring of NAFLD and NASH, potentially reducing the need for invasive liver biopsies.
• Future research will focus on validating these models with larger, multi-center datasets to assess generalizability and facilitate clinical integration.