SENSE AND GRAPPA RECONSTRUCTION OF MULTI-SHOT MULTI-ECHO EPI DATA

本子项目是利用由NIH/NCRR资助的中心赠款提供的资源的众多研究子项目之一。子项目和研究者（PI）可能已经从另一个NIH来源获得了主要资金，因此可以在其他CRISP条目中表示。列出的机构是中心的，不一定是研究者的机构。介绍。并行成像最有希望的好处之一是，它可以通过缩短快速采集技术（如EPI）的读出时间来减少畸变（1,2）。这种改进的代价是由于接收器几何形状不完美（g因子）和较少的测量而导致信噪比下降；由于更复杂的重建算法依赖于对接收线圈附加信息的估计，因此也有可能产生误差。这些附加信息通常是通过单独的低分辨率扫描获得的，或者通过将以奈奎斯特采样率获得的k空间段直接合并到感兴趣的序列中获得的。这两种技术都有局限性和缺点。我们最近开发了一种多镜头多回波EPI脉冲序列来测量灌注与多回波和时间增强（PERMEATE），在本卷的其他地方描述。该序列允许更大的重建灵活性：每个镜头都可以被视为单独的低采样k空间采集，并使用GRAPPA(3)或SENSE(4)进行重建，其有效缩减因子R等于交错的数量Ni。通过借鉴其他交错的信息，可以生成GRAPPA所需的自动校准信号（ACS），或者SENSE所需的全视场灵敏度图(4)。如果Ni足够大，则可以获得额外的灵活性，以便将各种拍摄组合整理成R < Ni的k空间集。在这项工作中，我们探讨了渗透数据的基本GRAPPA和SENSE重建的实现和性能之间的差异。材料与方法。使用带有8通道头部阵列（MRI Devices）的1.5 T扫描仪（GE Signa）获取健康志愿者的图像。采用96 ~ 96分辨率的PERMEATE脉冲序列，15个切片，4个交错，4个回波（TE = 12.4, 27.4, 42.4和57.4 ms）， TR = 1.2 s， 20个时间帧。在进行GRAPPA或SENSE重建之前，将所有4个连续交织片组装成一个全采样（R = 1） k空间，并进行EPI校正：使用熵最小化算法（在本卷的其他地方描述）进行相位校正，并重新划分以考虑斜坡采样。然后可以为每个切片和回波提取两种类型的相位编码子采样：将奇数和偶数镜头合并以使有效R = 2，或者将每个镜头单独处理以使R = 4。GRAPPA是在整个kx范围内使用5¿4内核进行的；使用的ACS线路数为2。使用典型的展开方法执行SENSE，其中找到R像素的最小二乘解，这些像素在每个获得的线圈分量图像中混叠到相同的位置。每个展开操作得到的矩阵是由由全采样数据导出的线圈灵敏度映射的数据组成的。这里使用的方法是通常用于从单独的校准扫描生成地图的方法的变体。通过对整个R = 1 k空间数据或中心子集（用于低分辨率估计）进行傅里叶变换，可以获得全视场图像。在后一种情况下，在图像域中执行三次样条插值到所需的分辨率。然后，每个线圈图像通过与平方核的卷积来平滑，并除以平方和图像。另一个改进是包含一个二进制掩模，它可以用来限制已知的像素的数量，以在混叠图像中贡献一个给定的像素，因此，减少来自感兴趣的解剖结构以外的区域的噪声贡献。结果和讨论。图1显示了各种R = 4重构（底部4行）的示例，顶部一行显示了R = 1重构，以便进行比较。GRAPPA重建表现良好，在大多数情况下，可以将其视为黑箱操作。SENSE重建的初步结果似乎不太乐观。这些是将该算法直接应用于临床扫描的结果，其中敏感性图来源于FGRE校准扫描。在这种情况下，使用10¿10内核对全采样图像进行平滑以生成地图。这种方法在头部边缘留下了相当多的残留混叠伪影，因为线圈灵敏度估计差；此外，在一些高信号堆积的区域，地图在估计接收器相位方面做得很差。为了改进这种重建，下一步是通过提取R = 1数据的中心32¿32区域来仅对地图使用低分辨率估计。这大大减少了高强度的区域，但仍然留下了不希望的残余混叠量。在这里，地图是由完整的k空间数据制作的，并用3¿3内核进行平滑。虽然这似乎是SENSE重建中最好的，特别是基本上没有残余混叠，但与GRAPPA相比，在高敏感性变化区域仍然存在更多的信号堆积。对此的一种可能解释是，用于制作线圈图的EPI数据中固有的敏感性缺失和扭曲对这些地区有很强的影响。引用。(1) Bammer R等。MRM 2001; 46:548。(2)杨秋霞，等。MRM 2004; 52:1418。(3) Griswold MA等。MRM 2002; 47:1202。(3) Pruessmann等。MRM 1999; 42:952。(4) Skare等。人脑定量扩散核磁共振成像方法研讨会2005:17。确认。美国国立卫生研究院（1R01EB002771），斯坦福大学高级磁共振技术中心（P41RR09784），卢卡斯基金会，橡树基金会。感谢MA Griswold分享opengrappa Matlab代码。

This subproject is one of many research subprojects utilizing the resources provided by a Center grant funded by NIH/NCRR. The subproject and investigator (PI) may have received primary funding from another NIH source, and thus could be represented in other CRISP entries. The institution listed is for the Center, which is not necessarily the institution for the investigator. Introduction. One of the most promising benefits of parallel imaging is that it can reduce distortions by shortening the read-out times of fast acquisition techniques such as EPI (1,2). The trade-off for this improvement is a drop in SNR due to imperfect receiver geometry (g-factor) and fewer measurements; there is also the potential for errors due to the more complicated reconstruction algorithms which rely on estimations of additional information about the receiver coils. This additional information is typically acquired either by a separate low-resolution scan or by incorporating a segment of k-space acquired at the Nyquist sampling rate directly into the sequence of interest; both of these techniques have limitations and disadvantages. We have recently developed a multi-shot multi-echo EPI pulse sequence to measure PERfusion with Multiple Echoes and Temporal Enhancement (PERMEATE), described elsewhere in this volume. This sequence allows for greater reconstruction flexibility: each shot can be treated as a separate under-sampled k-space acquisition and reconstructed using either GRAPPA (3) or SENSE (4) with an effective reduction factor, R, equal to the number of interleaves, Ni. By borrowing information from the other interleaves, it is possible to generate the auto-calibration signals (ACS) required for GRAPPA, or the full-FOV sensitivity maps required for SENSE (4). Additional flexibility is possible for acquisitions in which Ni is large enough so that various combinations of shots can be collated into k-space sets with R < Ni. In this work, we explore the differences between the implementation and performance of basic GRAPPA and SENSE reconstructions of PERMEATE data. Materials and Methods. Images from a healthy volunteer were acquired using a 1.5 T scanner (GE Signa) with an 8-channel head array (MRI Devices). The PERMEATE pulse sequence was used with 96¿96 resolution, 15 slices, 4 interleaves, 4 echoes (TE = 12.4, 27.4, 42.4, and 57.4 ms), TR = 1.2 s, and 20 time frames. Prior to GRAPPA or SENSE reconstruction, all 4 consecutive interleaves were assembled into one fully-sampled (R = 1) k-space and EPI correction was performed: phase correction using an entropy-minimization algorithm (described elsewhere in this volume) and regridding to account for ramp sampling. Two types of phase-encoding sub-samplings could then be extracted for each slice and echo: odd and even shots were combined for an effective R = 2, or each shot was treated separately for R = 4. GRAPPA was performed using a 5¿4 kernel applied over the entire kx range; the number of ACS lines used was 2. SENSE was performed using the typical unfolding method in which a least-squares solution is found for the R pixels that are aliased into the same location in each of acquired coil component images. The matrices that get inverted for each unfolding operation are comprised of data from coil sensitivity maps derived from the fully-sampled data. The methods used here are variations on those typically used to generate maps from a separate calibration scan. The full-FOV image can be obtained by taking the Fourier transform of the entire R = 1 k-space data or of a central subset (for a low-resolution estimate). In the latter case, cubic spline interpolation to the desired resolution is performed in the image domain. Each coil image is then smoothed by convolution with a square kernel and divided by the sum-of-squares image. One more refinement is the inclusion of a binary mask that can be used to limit the number of pixels known to contribute a given pixel in the aliased image and, hence, reduce noise contribution from areas outside the anatomy of interest. Results and Discussion. Figure 1 shows examples of various R = 4 reconstructions (bottom 4 rows) with the R = 1 reconstruction show in the top row for comparison. The GRAPPA reconstruction performs well and, for the most part, can be treated as if it were a black-box operation. The initial results from the SENSE reconstruction seemed less promising. These were the results of a direct port of the algorithm being used for clinical scans in which the sensitivity maps were derived from a FGRE calibration scan. In this case, the fully-sampled images were smoothed with a 10¿10 kernel to generate the maps. This method leaves considerable residual aliasing artifacts from poor coil sensitivity estimation at the edges of the head; also, there are regions of high signal pile-up in locations where the maps do a poor job of estimating the receiver phase. To improve this reconstruction, the next step was to use only a low-resolution estimate for the maps by extracting the center 32¿32 region of the R = 1 data. This significantly reduced the regions of hyper-intensity but still left an undesirable amount of residual aliasing. Here, maps were made from the full k-space data and smoothed with a 3¿3 kernel. While this appears to be the best of the SENSE reconstructions, especially in that there is essentially no residual aliasing, there is still more signal pile-up in regions of high susceptibility variation as compared to GRAPPA. One possible explanation for this is that the susceptibility drop-outs and distortions inherent in the EPI data used to make the coil maps have a strong influence in those regions. References. (1) Bammer R, et al. MRM 2001;46:548. (2) Yang QX, et al. MRM 2004;52:1418. (3) Griswold MA, et al. MRM 2002;47:1202. (3) Pruessmann, et al. MRM 1999;42:952. (4) Skare et al. Workshop on Methods for Quantitative Diffusion MRI of Human Brain 2005:17. Acknowledgements. NIH (1R01EB002771), The Center for Advanced MR Technology at Stanford (P41RR09784), The Lucas Foundation, The Oak Foundation. Thanks to MA Griswold for sharing the opengrappa Matlab code.