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

Clinical MR elastography (MRE) typically requires acquisition of four vibration phase images synchronized with external vibrations, prolonging acquisition time and increasing slice misalignment risk due to inadequate breath-holding. To address this limitation, we developed a deep learning-based method that reduces required phase acquisitions by interpolating missing vibration phase images through exploitation of spatiotemporal wave periodicity. The purpose of the present study is to evaluate the technical feasibility of this novel approach. Two deep learning models were developed: a 3-to-1 model requiring three phases (simulating 25% reduction) and a 2-to-2 model requiring two phases (simulating 50% reduction). Validation was performed using phantom experiments and in vivo liver MRE in 13 healthy volunteers. To eliminate acquisition-related variations, single complete 4-phase datasets were acquired and specific phases were computationally selected during post-processing to simulate reduced-phase acquisitions. Generated wave images were evaluated using structural similarity index (SSIM) and peak signal-to-noise ratio (PSNR), while elastograms were evaluated using Bland-Altman analysis and intraclass correlation coefficient (ICC). Statistical significance was tested using the Wilcoxon signed-rank test (p < 0.05). Both models generated wave images with high similarity to the reference standard, demonstrating excellent SSIM and PSNR in both phantom and in vivo studies. Shear stiffness measurements demonstrated high agreement with 95% limits of agreement in the range of 0.021-0.143 kPa (phantoms) and 0.209-0.311 kPa (liver). ICC analysis showed excellent agreement (>0.90) in both studies. In vivo measurements showed no significant differences from the reference standard (p > 0.05). These results demonstrate the technical feasibility of the proposed method, suggesting its potential to reduce required phase acquisitions by up to 50% while maintaining clinically acceptable measurement accuracy comparable to the conventional method.