Self-Supervised T2WI-Bridged Framework for Liver Segmentation and PDFF Prediction From US Images

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Hepatology

Background:

• Proton Density Fat Fraction (PDFF) via MRI is the gold standard for fatty liver diagnosis but is limited by MRI's accessibility.
• B-mode Ultrasound (US) is widely accessible for liver assessment, presenting an opportunity for non-invasive fatty liver diagnosis.

Purpose of the Study:

• To develop a novel framework for liver segmentation and PDFF prediction using only US images.
• To overcome limitations of limited paired US-PDFF data and US image noise/artifacts for accurate PDFF estimation.

Main Methods:

• A cross-task self-supervised pretext task was integrated to extract semantic features, enhancing generalization.
• T2-weighted imaging (T2WI) was used during training to bridge US and PDFF features, improving prediction accuracy.
• An uncertainty-augmented adversarial loss function refined liver boundary delineation for improved segmentation and PDFF prediction.

Main Results:

• The proposed framework achieved superior performance in liver segmentation and PDFF prediction compared to state-of-the-art methods.
• Predicted PDFF values demonstrated accuracy comparable to real PDFF for hepatic steatosis classification.
• The model relies solely on US for inference, offering a practical and cost-effective solution.

Conclusions:

• The developed US-based framework provides a clinically viable and cost-effective alternative for non-invasive fatty liver diagnosis.
• This approach has significant potential to broaden the clinical applicability of PDFF estimation.
• Publicly available code facilitates further research and development in this area.