Liver fat quantification at 0.55 T enabled by locally low-rank enforced deep learning reconstruction

Abstract

Purpose: The emergence of new medications for fatty liver conditions has increased the need for reliable and widely available assessment of MRI proton density fat fraction (MRI-PDFF). Whereas low-field MRI presents a promising solution, its utilization is challenging due to the low SNR. This work aims to enhance SNR and enable precise PDFF quantification at low-field MRI using a novel locally low-rank deep learning-based (LLR-DL) reconstruction.

Methods: LLR-DL alternates between regularized SENSE and a neural network (U-Net) throughout several iterations, operating on complex-valued data. The network processes the spectral projection onto singular value bases, which are computed on local patches across the echoes dimension. The output of the network is recast into the basis of the original echoes and used as a prior for the following iteration. The final echoes are processed by a multi-echo Dixon algorithm. Two different protocols were proposed for imaging at 0.55 T. An iron-and-fat phantom and 10 volunteers were scanned on both 0.55 and 1.5 T systems. Linear regression, t-statistics, and Bland-Altman analyses were conducted.

Results: LLR-DL achieved significantly improved image quality compared to the conventional reconstruction technique, with a 32.7% increase in peak SNR and a 25% improvement in structural similarity index. PDFF repeatability was 2.33% in phantoms (0% to 100%) and 0.79% in vivo (3% to 18%), with narrow cross-field strength limits of agreement below 1.67% in phantoms and 1.75% in vivo.

Conclusion: An LLR-DL reconstruction was developed and investigated to enable precise PDFF quantification at 0.55 T and improve consistency with 1.5 T results.

Keywords: deep learning; liver PDFF; locally low‐rank; low‐field MRI; reconstruction.

FIGURE 1

Visualization of the proposed LLR–DL reconstruction method. The echo images are reconstructed using regularized SENSE with patch‐wise local SVD being performed along the echo dimension. The echoes are projected on their singular values and processed using U‐Net architecture. The low‐rank output of the U‐Net is projected on the original images and used as prior for subsequent iterations. DL, deep learning; LLR, locally low‐rank; SVD, singular value decomposition.

FIGURE 2

A visualization of the magnitude of the fat dephasing coefficients computed using the nine peaks fat model. The acquired TEs at the different field strengths are depicted. At 0.55 T, the echoes are acquired at minimum TE (top x‐axis) and can only sample one fat dephasing cycle, whereas it is possible to sample up to three fat dephasing cycles at 1.5 T (bottom x‐axis)

FIGURE 3

Echo image of the fat‐and‐iron phantom acquired at 0.55 T and TE = 6.8 ms (left). Nominal PDFF values of eight vials from the datasheet and $R_2^{*}$ values at 1.5 T (right). PDFF, proton density fat fraction.

FIGURE 4

Reconstruction results of opposed‐phase and in‐phase images at low‐field corresponding to echoes 3 and 6 in both protocols. The upper two rows are the reconstruction results of the data acquired using protocol 1. The proposed LLR–DL shows superior image quality to the conventional CAIPIRINHA reconstruction. The lower two rows show the reconstruction results of protocol 2. Here, the conventional CAIPIRINHA method provides images with good SNR; however, the LLR–DL enhanced the SNR. For all reconstruction results, the SNR level is visible in the fitted quantitative PDFF and $R_2^{*}$ Maps. An example ROI from the liver segmentation is overlaid on the quantitative maps, and the corresponding mean and standard deviation are reported. The corresponding high‐field dataset is shown in Figure 5, enabling cross‐field comparison of reconstruction quality. CAIPIRINHA, controlled aliasing in parallel imaging results in higher acceleration; ROI, region of interest.

FIGURE 5

Reconstruction results of opposed‐phase and in‐phase images at 1.5 T, corresponding to echoes 2 and 4 (protocol 1) and echoes 1 and 2 (protocol 2). The upper two rows are the reconstruction results of the data acquired using protocol 1. The lower two rows show the reconstruction results of protocol 2. Both LLR–DL and conventional CAIPIRIHINIA provided images with good SNR. This dataset is acquired on the same volunteer as in Figure 4. An example ROI from the liver segmentation is overlaid on the quantitative maps, and the corresponding mean and standard deviation are reported.

FIGURE 6

PDFF assessment of phantom. The upper row shows the measurements at 0.55 T and the lower row shows the measurements at 1.5 T. The mean of the scan repetitions was used to compute these Bland–Altman plots. The MD values of the different protocols ranged between 1.09 and 1.25. Similar LoA were achieved using the different protocols. MD, mean difference. LoA, limits of agreement.

FIGURE 7

Bland–Altman plot showing the MD, LoA, and reproducibility of liver PDFF across the field strengths in vivo. Each measurement was repeated once. Volumetric analysis with liver segmentation and vessel exclusion was performed. The mean of the scan repetitions is used for Bland–Altman analysis. For LLR–DL, the repeatability coefficients in PDFF unit at 0.55 T were 0.46% for protocol 1 and 0.79% for protocol 2. At 1.5 T, the coefficients were 1.54% for protocol 1 and 0.48% for protocol 2. MD, mean difference.