Deep Learning-Enabled Automated Quality Control for Liver MR Elastography: Initial Results.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Hepatology

Background:

• Liver magnetic resonance elastography (MRE) image quality is crucial for reliable stiffness measurements.
• Factors like driver positioning and patient characteristics can impair MRE quality.

Purpose of the Study:

• To evaluate deep learning (DL) architectures for automated classification of liver MRE image quality.
• To assess the performance of DL models in distinguishing diagnostic from non-diagnostic MRE slices.

Main Methods:

• Retrospective study of 90 patients undergoing liver MRE.
• Two observers scored 914 MRE slices for quality.
• Evaluated DL architectures (ResNet, SqueezeNet, MobileNetV2) for binary quality classification.
• An ensemble model combining predictions from the best architecture was developed.

Main Results:

• DL models achieved an average accuracy of 0.851 for MRE quality classification.
• The SqueezeNet ensemble model reached an accuracy of 0.921.
• Excellent inter-observer agreement (Kappa 0.896) was observed.

Conclusions:

• Automated DL-based classification of liver 2D MRE quality is feasible.
• DL models, particularly SqueezeNet ensembles, show high accuracy in quality assessment.
• This approach can enhance the reliability of MRE for liver stiffness assessment.