AUTO-CALIBRATED PARALLEL IMAGING USING A DISTORTION-OPTIMAL FILTER-BANK

INTRODUCTION Most auto-calibrated parallel MRI (pMRI) methods, especially the popular GRAPPA method [1], perform image reconstruction by solving for missing k-space samples. Such methods, also known as k-space-based pMRI, utilize the redundancy in the multi-channel MR measurements to compensate f or the R-fold subsampling (R≥2) relative to the Nyquist rate [2]. In essence, k-space-based methods solve the inverse problem of interpolating the acquired k-space data to the full Nyqusit grid. As in any interpolation problem, constraints are typically introduced to ensure certain “consistency” conditions. In GRAPPA the consistency condition is matching of the reconstruction to the acquis ition over a fully sampled region, typically located at the center of k-space, called the auto-calibrating scan (ACS) region. In this work, we focus on non-iterative Cartesian k-space-based auto-calibrated pMRI with minimal amount of ACS data. We present a new method that achieves artifact-free image reconstruction for high acceleration factors (R=3 1D acceleration in the presented results) with few ACS lines (≈5% of phase-encodes). In-vivo results comparing the method to two alternative implementations of GRAPPA are provided. THEORY Fig. 1 is a diagram for general k-space-based pMRI acquisition and reconstruction scheme. The pMR measurements consist of: (i) R-fold uniformly subsampled data from the N channel outputs; (ii) calibration data produced by the data truncation operator ΠACS that removes all but the ACS phase encode lines. In the most general setting all pMR measurements are utilized in both the calibration process and the interpolation process (Fig. 1). The final outcome of the interpolation process is the fully encoded (interpolated) N-channel k -space data set, which is then combined after inverse FFT (e.g., using sum of squares) to form the final reconstructed image. The calibration scheme, indicated in all Figs. by the dashed lines, adjusts the coefficients in the interpolation kernel to optimize the output image and its consistency with the measured data. Fig. 2 depicts the detailed calibration diagram for GRAPPA, where the interpolation process involves the application of the GRAPPA weights to the R-fold subsampled data. The autocalibration process (Fig. 2) measures the error between the measured ACS data and the corresponding k-space samples in the interpolated output; and adjusts (estimates) the coefficients of the interpolation kernel to minimize this error in the least-squares (LS) sense. Therefore, the only error measure considered in GRAPPA is consistency over the ACS region. Here, we propose an improved scheme (Fig. 3) that uses not only the data consistency (over the entire measured k-space) but also integrates a predicted measure of the aliasing distortion into the calibration process—hence employing a “total distortion” measure rather than just data consistency. The method uses a multi-input multi-output (MIMO) filter bank as its kernel and is dubbed ACSIOM (auto-calibrated sensitivity-encoded imaging using an interpolation-optimal MIMO filter bank). METHODS All reconstruction experiments used MR data acquired on a 3T whole-body GE scanner (GE Healthcare, Waukesha, WI) from a healthy volunteer using an 8-channel head coil array (2D fast spoiled