4708
Ultrafast 3D Partial Fourier Reconstruction with Well-Preserved Phase using DNN1Neusoft Medical Systems, Shenyang, China, 2Neusoft Medical Systems, Shanghai, China
Synopsis
Partial Fourier (PF) is a widely used fast imaging scheme. Since phase information is crucial in many applications, such as SWI, it is necessary that PF can preserve phase well. Many PF methods cannot preserve phase well especially at locations with rapid phase change. DPA is a method can recover both magnitude and phase well, but suffers from low speed for two-directional PF acquisition. Considering recent advances in deep learning, we proposed a DNN-based framework for two-directional PF reconstruction. Preliminary experiments demonstrate that the proposed method is almost 50 times faster while restores magnitude and SWI even better than DPA.
INTRODUCTION
As an important fast imaging method, Partial Fourier (PF) acquisition has been widely used in many clinical MRI scans. However, problems still exist in conventional PF reconstruction techniques. First, most algorithms that apply phase estimation in image space 1-2 usually produce artifacts in the regions of fast local phase change. Second, Double Phase Approximation (DPA) algorithm 3 and its variations can reduce the above-mentioned artifacts to a certain degree, but often suffer from long reconstruction time for two-directional PF acquisition data. To address the above problems, in this work, we proposed a deep neural network (DNN) based method for two-directional Partial Fourier 3D MRI reconstruction, which can restore both magnitude and phase accurately using close to 50 times less reconstruction time.METHODS
Considering that DNNs possess strong capability in learning high-level features and complex relationships, we designed a DNN to learn the nonlinear mapping between PF acquired data and fully-sampled image. The proposed network is composed of two parts: a 3D U-net with seven-layer convolutional network and a data-consistency (DC) unit. Here, we trained the 3D U-net by taking advantage of the local correlation between adjacent slices, and also exploited a complex-valued network for accurate phase recovery. The DC layer substitutes the K-space values of the preceding U-net output with the acquired data to ensure the data fidelity. The overall architecture is presented in Fig.1. As can be seen that, the proposed network takes the complex images derived from multi-channel undersampled K-space data 4 as input, and uses the coil-combined complex-valued images derived from the fully-sampled K-space as ground truth. To demonstrate the performance of our proposed network, multi-echo acquisition for STAGE 5-7 was used as an example in this work. 9 sets of 6-echo 3D STAGE data with 0.67x0.67x2.7 mm3 resolution were acquired on a 1.5T MR scanner (NSM S15P, Neusoft Medical Systems, China) using an 8-channel head coil for network training, and 3 additional sets were acquired for testing. Undersampled data were retrospectively obtained using two-directional Partial Fourier sampling mask with PF fraction of 76% along ky and 86% along kz (totally 66% of the fully-sampled data). For evaluation, we compared the proposed approach with the conventional DPA method.RESULTS
Figure 2 shows the comparisons of magnitude results from two echoes of STAGE images. The corresponding error maps and RMSE values between ground truth and the reconstructed images are also presented. As can be clearly seen that, the magnitude images derived from the proposed network have lower reconstruction error than DPA method. The corresponding phase images reconstructed from DPA and the proposed network are also compared in Figure 3. It can be seen that the proposed method achieves well-preserved details of phase change. In addition, the SWI images based on the reconstruction of all 6-echo STAGE dataset were evaluated. Figure 4 shows the SWI images and the corresponding error maps obtained from DPA method and the proposed network. As can be seen that, the vein structure derived from the proposed method is closer to the ground truth than DPA. Table 1 summarizes the comparisons between the results from DPA method and the proposed network in terms of RMSE, SSIM and PSNR. The quantitative results further demonstrate that the proposed method yields comparable and even higher accuracy than DPA. Moreover, DPA took 96 seconds to reconstruct all 6-echo STAGE images, while the proposed method only cost 2 seconds.DISCUSSION
According to the above results, we have empirically verified that the proposed DNN approach can obtain more accurate reconstructions than the conventional DPA method from 66% two-directional partially acquired data. The better performance of the proposed method is due to the fact that DNN is able to establish the complex non-linear mapping between partially acquired data and ground truth via fully data-driven learning. Another advantage of the proposed method is that it provides much faster reconstruction speed than DPA which is based on online kernel learning.CONCLUSION
In this work, a deep-learning-based two-directional Partial Fourier 3D reconstruction method is proposed. Preliminary results have demonstrated that the proposed method can restore more accurate magnitude and SWI than conventional DPA method, and accelerate the reconstruction speed by almost 50 times.Acknowledgements
No acknowledgement found.References
1. Cuppen J, van Est A. Reducing MR imaging time by one-sided reconstruction. Magn Reson Imaging 1987;5:526–527.
2. Haacke EM, Lindskog ED, Lin W. A fast, iterative, partial-Fourier technique capable of local phase recovery. J Magn Reson 1991;92:126–145.
3. Huang F, Lin W, Li Y. Partial Fourier reconstruction through data fitting and convolution in k-space. Magn Reson Med, 2009; 62:1261-1269.
4. Pruessmann KP, Weiger M, Scheidegger MB, et al. SENSE: sensitivity encoding for fast MRI. Magn Reson Med, 1999; 42(5): 952-962.
5. Chen Y, Liu S, Wang Y, et al. STrategically acquired gradient Echo (STAGE) imaging, part I: Creating enhanced T1 contrast and standardized susceptibility weighted imaging and quantitative susceptibility mapping. Magn Reson Imag, 2017; 46:130-139.
6. Wang Y, Chen Y, Wu D, et al. Strategically acquired gradient Echo (STAGE) imaging, part II: Correcting for RF inhomogeneities in estimating T1 and proton density. Magn Reson Imag, 2017;46:140-150.
7. Chen Y, Liu S, Kang Y, et. al. An interleaved sequence for simultaneous MRA, SWI and QSM. ISMRM 2017; p1215.
Figures
