Introduction

Magnetic Resonance parametric mapping can provide quantitative information to characterize tissue properties.^1,2 However, slow data acquisition speeds usually hinder real-time magnetic resonance (MR) parameter mapping. In this work, a novel single-shot T₂ mapping method, which can be acquired in hundreds of milliseconds, was proposed based on multiple overlapping-echo detachment (MOLED) with a planar imaging acquisition scheme, termed the multi-train MOLED (MMOLED) method. Multiple overlapping echo signals from different echo trains with the different T₂-weighting were obtained simultaneously.³ A convolution neural network (CNN) was proposed to reconstruct T₂ mapping. The robustness and efficiency of the MMOLED sequence were demonstrated using numerical brains, phantom experiments and human brains.

Methods

The novel single-shot MMOLED T₂ mapping sequence designed is shown in Fig.1. The current strategy uses sequences that have three echo trains, and for each echo train, four echo signals with different evolution time are generated in turn by four excitation pulses that hold a same small flip angle of α. Four echo-shifting gradients (G₁,G₂,G₃,G₄) were used to shift the echoes from the k-space center along the phase-encoding and frequency-encoding directions.
To compare the reconstruction results of different echo train numbers during post-processing, we used first echo train (dubbed MMOLED-1 train), the first two echo trains (dubbed MMOLED-2 trains) and the first three echo trains (dubbed MMOLED-3 trains) as the input of the convolution neural network (CNN). The pure synthetic training samples were utilized, the detail can be seen in the previous paper.^4-7 The phantom and human brain validation experiments were performed on a whole-body 3T scanner ( MAGNETOM Prisma, Siemens Healthcare, Erlangen, Germany). For the single-shot MMOLED-1, MMOLED-2, and MMOLED-3 trains sequence, the scan time was 118 ms, 185 ms and 252 ms respectively with acquisition matrix 128×128 and FOV 22×22 cm². The conventional SE sequence was used for comparison.

Results and Discussion

A numerical Brain pattern (including T₂ , B₁ and proton density templates) was constructed. The numerical simulation results are shown in Fig.2. The images before reconstruction (Fig. 2(a)) had visible cross stripes, since they were composed of multiple images with different linear phase ramps. A reference T₂ map (Fig. 2(b)) was used as a templet for the simulation. We enlarge a region in numerical brain for comparison. The reconstruction results of the three input types mentioned are also shown in Fig. 2(c-e). The T₂ mapping textures from the MMOLED-2 (Fig. 2(d)) and MMOLED-3 train3 (Fig. 2(e)) were most consistent with the T₂ mapping reference, especially in regions with relatively large T₂ values, as indicated by the red arrows. The red circle shown in Fig. 2(a) denotes a trace for the comparison of the T₂ values from MMOLED-1, MMOLED-2, MMOLED-3 trains and reference T₂ mapping. The T₂ profiles along the red trace from the different T₂ mapping schemes were drawn in Fig. 2(f), which showed that in the cerebrospinal fluid (CSF) regions where high T₂ values usually exist, the results from multi-train OLED including MMOLED-2 trains and MMOLED-3 trains had better agreement with the reference than MMOLED-1 train. The Structural Similarity Index (SSIM) for the entire image results of MMOLED-1, MMOLED-2, and MMOLED-3 trains were 0.775, 0.991 and 0.988, respectively.
For the phantom validation, the mean T₂ values ranged from 15 ms to 180 ms (Fig. 3). We found that T₂ mapping reconstructed from MMOLED-2 and MMOLED-3 trains performed better when T₂ value was a relatively large value (>75ms). Moreover, the maximum compartment 16 deviation for MMOLED-1 train was about 11.1 ms, while that of the MMOLED-2 and MMOLED-3 trains was about 6.2 ms and 5.7 ms, respectively.
In Fig.4, the mean T₂ values and standard deviations (SDs) for 17 regions of interest (ROIs) marked in Fig. 4(a) were calculated and the results were given in Fig. 4(b). The 17 ROIs were chosen manually to represent different regions of the brain, including white matter, gray matter, and deep gray matter with different T₂ values. All of the reconstructed T₂ mapping results were compared with the T₂-mapping reference obtained from SE sequence using the FNIRT tool (FMRIB’s Nonlinear Image Registration Tool) of FSL (Oxford, UK). Most results agreed well with SE result, but in ROI10 where T₂ value was relatively large (>75ms), the reconstructed results from MMOLED-2 and MMOLED-3 trains performed better than those of the MMOLED-1 train. We found that the reconstructed T₂ mappings from MMOLED-2 and MMOLED-3 trains (Fig. 4(a)) were closer in texture details to the T₂ reference compared with the mapping from MMOLED-1 train. The B₁ and M₀ maps could also be reconstructed through the convolution neural network output from three different methods as shown in Fig. 5.⁸ As one can see, The reconstructed B₁ mapping from MMOLED-2 and MMOLED-3 trains was also enhanced than from the MMOLED-1 train.

Conclusion

The MMOLED-2 and MMOLED-3 trains provided better single-shot T₂ mapping and B₁ mapping than did the MMOLED-1 train, and all could be acquired within milliseconds while maintained high image quality, which is useful for dynamic imaging. The successful application of deep learning method and synthetic training dataset proved their great potential in making complex MRI sequences available.

Acknowledgements

No acknowledgement found.