Unsupervised Test-Time Adaptation for Hepatic Steatosis Grading Using Ultrasound B-Mode Images.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Hepatology

Background:

• Ultrasound (US) is crucial for assessing hepatic steatosis (fatty liver) due to its noninvasiveness.
• Deep learning models show promise for accurate steatosis grading but suffer from domain shift across clinical sites.
• Unsupervised domain adaptation is vital for generalizing AI models in diverse healthcare environments.

Purpose of the Study:

• To develop and evaluate a novel unsupervised domain adaptation technique for hepatic steatosis classification using ultrasound.
• To address the challenge of domain shift caused by variations in patient data, equipment, and protocols.
• To improve the generalizability and robustness of deep learning models in real-world clinical settings.

Main Methods:

• Proposed a test-time batch normalization (TTN) technique to adapt trained convolutional neural network models.
• TTN adapts selected batch normalization layer features in an unsupervised manner at test time.
• Evaluated the method on two distinct abdominal ultrasound datasets from different institutions.

Main Results:

• The TTN method significantly reduced the mean absolute error in steatosis grading by 37%.
• Improved the area under the receiver operating characteristic curves (AUC) for steatosis detection from 0.78 to 0.97.
• Demonstrated effective mitigation of domain shift without requiring labeled data from the target domain.

Conclusions:

• The proposed TTN technique offers a robust solution for domain shift in ultrasound-based hepatic steatosis diagnosis.
• This method enhances the reliability of AI models across diverse clinical settings, minimizing deployment risks.
• TTN shows potential for improving the clinical utility of deep learning in medical imaging applications.