Cascaded Deep Learning Neural Network for Automated Liver Steatosis Diagnosis Using Ultrasound Images

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Hepatology

Background:

• Diagnosing liver steatosis is crucial for early detection of hepatocirrhosis and liver cancer.
• Ultrasound (US) image quality issues like speckle noise and blurring challenge automated diagnosis.

Purpose of the Study:

• To develop a fully automated deep learning model for accurate liver steatosis prediction from US images.
• To overcome the limitations of current automated diagnostic methods for fatty liver disease.

Main Methods:

• Utilized three deep learning neural networks for automated liver steatosis grading.
• Employed transfer learning for semantic segmentation of liver-kidney regions and subsequent cropping.
• Developed a specialized network, SteatosisNet, for severity grading using cropped liver-kidney areas.

Main Results:

• Achieved high diagnostic performance with 99.78% sensitivity, 100% specificity, 100% positive predictive value (PPV), 99.83% negative predictive value (NPV), and 99.91% accuracy.
• The model's performance is comparable to that of medical experts in grading fatty liver disease.

Conclusions:

• The proposed automated deep learning model accurately predicts liver steatosis severity from ultrasound images.
• This approach offers a reliable tool for early detection and management of fatty liver disease, potentially improving patient outcomes.