Fast MR elastography via deep learning-based phase interpolation: A technical feasibility study

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Biophysics

Background:

• Clinical MR elastography (MRE) requires four vibration phase images, increasing acquisition time and risk of slice misalignment.
• Inadequate breath-holding can compromise MRE data quality and accuracy.

Purpose of the Study:

• To evaluate the technical feasibility of a deep learning-based method for reducing MRE phase acquisitions.
• To assess the potential of interpolating missing vibration phase images using spatiotemporal wave periodicity.

Main Methods:

• Developed two deep learning models: a 3-to-1 (25% reduction) and a 2-to-2 (50% reduction) phase acquisition model.
• Validated models using phantom experiments and in vivo liver MRE in 13 healthy volunteers.
• Evaluated wave images using SSIM/PSNR and elastograms using Bland-Altman analysis/ICC.

Main Results:

• Both models generated high-quality wave images with excellent SSIM and PSNR.
• Shear stiffness measurements showed high agreement and excellent ICC (>0.90) in phantoms and in vivo.
• In vivo measurements showed no significant differences compared to the conventional 4-phase method (p > 0.05).

Conclusions:

• The proposed deep learning method is technically feasible for reducing MRE phase acquisitions.
• This approach can potentially reduce acquisitions by up to 50% while maintaining clinically acceptable accuracy.
• The method offers a promising alternative to conventional MRE, improving efficiency and potentially reducing motion artifacts.