A Unified Deep Learning Framework for Liver MR Elastography Postprocessing: Proof-of-Concept Study.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Hepatology

Background:

• Magnetic resonance elastography (MRE) is crucial for non-invasive liver stiffness measurement and hepatic fibrosis assessment.
• Current MRE analysis requires manual region of interest (ROI) delineation, which is time-consuming, expert-dependent, and prone to variability.

Purpose of the Study:

• To assess the feasibility of a single deep learning (DL) pipeline for fully automating liver MRE analysis, from data acquisition to stiffness quantification.
• To develop and validate a DL framework for reconstructing stiffness maps and segmenting the liver directly from MRE images.

Main Methods:

• A convolutional neural network (CNN)-based framework was developed and trained using MRE data from 83 adult volunteers.
• Multiple neural network architectures (U-Net, ResNet, CycleGAN hybrids) were evaluated for performance.
• Automated liver segmentation and stiffness quantification were performed directly from MRE magnitude and phase images.

Main Results:

• The DL pipeline achieved liver stiffness estimates with less than 11% deviation from reference values, and under 3% in optimal configurations.
• Automated analysis showed agreement comparable to interreader agreement, with intraclass correlation coefficients (ICC) of 0.86 and 0.89.
• The total inference time per examination was rapid, averaging 23.7 ± 4.4 seconds.

Conclusions:

• A fully automated, AI-driven postprocessing pipeline for liver MRE is technically feasible in a controlled proof-of-concept setting.
• This AI approach shows promise for reducing analysis time and operator dependence in liver stiffness assessment.
• Further validation in diverse clinical populations is necessary for broader generalization and potential clinical application.