1527 Low Rank SENSE : A Robust Reconstruction Method for ms-EPI based high spatial resolution DWI
F. Huang,Zhe Zhang,R. Duensing,Bida Zhang,Hua Guo
2013-01-01
Abstract:Purpose To improve the spatial resolution of diffusion weighted imaging (DWI), multi-shot EPI (ms-EPI) has been proposed. Due to the high sensitivity to motion, different shots of ms-EPI may have inconsistent phase, which result in aliasing artifacts in the reconstructed image [1, 2]. To remove the impact of phase differences, parallel imaging, such as SENSE [3], is conventionally used to generate individual images for each shot. The magnitude average of these images is used as the final reconstruction. This method may result in clinically unacceptable image when the number of shots is high. The image will be corrupted by the g-factor and aliasing artifacts due to high acceleration factors, which is equal to the number of shots. However, more number of shots is expected for higher spatial resolution and/or lower geometry distortion, especially for high field system [2]. The goal of this work is to provide a reconstruction algorithm for msEPI, which can robustly deliver high image quality even when the number of shots is as high as 8 with an 8-channel RF coil. Methods Unlike previous methods for ms-EPI phase correction [1,2], the ms-EPI is treated as accelerated dynamic imaging in this work. Each single shot corresponds to one static image. Hence the acceleration factor of the dynamic image set is the number of shots. A unique property of this dynamic image set is that these images have the same magnitude [1], but possibly different phase. Therefore, the rank of the image set is low. Low-rank property has been successfully used for the reconstruction of highly accelerated dynamic imaging [4,5]. Therefore, we propose to use low rank property to solve this accelerated dynamic imaging problem for ms-EPI. The mathematical model of the proposed method can be written as ∑ ‖ ‖ (1). s is the count of shot; is the image of shot s; F is the encoding operator; is the acquired data for shot s; ∗ is the rank calculation operator; is an non-negative parameter to balance these 2 terms. Since number of signal averages (NSA) for the acquisition of DWI is often larger than 1 to improve the SNR, s can include both number of shots and number of average. When multi-channel data is used, the first data fidelity term is the SENSE term. The second term is the low-rank regularization term, which forces the low rank property of the images corresponding to multi-shots of the same subject. Hence, the model is actually low rank regularized SENSE. If there is no motion at all, then all of these images should be exactly same and the rank is one. If there is only occasional relatively large scale motion, the rank is still low. In these scenarios, the model is theoretically sound. The model will fail when there is continuous random motion. Ideally, the final reconstruction is the registered average of . This method is called Low-rank SENSE. To evaluate the performance of low-rank SENSE, 10 sets of high resolution DTI images were acquired on a Philips 3T Achieva system with an 8-ch head coil (Invivo Corp., Gainesville, U.S.A). A multi-shot spin echo EPI sequence (b = 800 s/mm, number of average = 3 or 4, number of directions (nDir) = 6, 10 or 15, FOV = 230mm, in-plane spatial resolution = 0.8 mm, slice thickness 4mm, partial Fourier ratio = 0.6, FA = 90, TR = 2.8s, and TE = 70ms) was used for data acquisition. Number of shots 4, 6 and 8 were used in the experiment. In the results, a simplified numerical implementation using variable splitting [6] for model (1) was used for the reconstruction: rank = 1 was enforced for the magnitude of . The simplified version is theoretically correct when there is no large scale motion. Results Fig. 1 shows the comparison of the reconstruction with and without using the low rank regularization. A 6-shot data set (NSA = 3, nDir = 10) was used in this example. The first row of Fig. 1 shows that ms-EPI resulted in serious artifacts without taking care of the phase variation. The conventional SENSE average scheme resulted in serious artifacts due to the high acceleration factor 6 with a 8-channel coil (second row). With low-rank regularization, the image quality was dramatically improved (third row). Fig. 2 shows the FA map comparison for data sets with number of shots 4, 6, and 8. With the increase of number of shots, the advantge of using low-rank regularization becomes more obvious. Even for number of shots 8, low-rank SENSE can still result in clinically acceptible FA map.