Optimizing the Net Acceleration of GRAPPA and PEAK-GRAPPA

Introduction: Parallel imaging methods require calibration to coil sensitivities. The autocalibration approach acquires the calibration data along with the normal data acquisition while sampling the full k-space center with a certain number autocalibration lines Nacs [1]. Thus, the net acceleration of such an experiment depends on the undersampling factor R, the number of autocalibration lines Nacs and the number of phase encoding steps Ny. Different combinations of R/Nacs can result in the same net acceleration (Fig.1). The aim of this work was to systematically investigate different autocalibrating methods (standard spatial acceleration GRAPPA [2], spatio-temporal acceleration PEAK-GRAPPA[3]) with respect to their reconstruction performance depending on different R/Nacs combinations for the same net acceleration. Materials and Methods: All measurements were performed on a 3T system (Trio, Siemens) with full k-space sampling. To obtain undersampled data, phase encoding lines were subsequently removed and set to zero according to the sampling scheme. Each measurement was repeated without rf-excitation (flip angel = 0) for noise analysis. Phantom measurements were performed with a 12 channel body coil using an rf-spoiled CINE gradient echo sequence (matrix size = 256 x 256, spatial resolution = 0.94 x 0.94 mm, temporal resolution 16.8 ms (68 time frames)). The phantom consisted of a moving part filled with agarose gel and a static water bottle. The moving phantom oscillated with a frequency of approximately 1Hz. Additionally, in-vivo short axis cardiac images were acquired during breathold with 15 coil elements. A 2D bSSFP sequence was used with matrix size = 202 x 198, spatial resolution = 1.4 x 1.4 mm and temporal resolution = 33 ms (26 time frames). Three different reconstruction algorithms were evaluated: PEAK-GRAPPA (kernel size: by = bt = R+d with d=2 for R=2,3 and d = 4 for R>3 and bx =3, [3]), GRAPPA (kernel size: bx x by = 5x2), and view sharing. The autocalibration lines were copied back into k-space after reconstruction. To estimate image quality root mean square error (RMSE) and Noise in different regions (red lines in Fig. 2 and Fig. 3) was calculated. Image noise was estimated from the 'noise only' data, which underwent the same reconstruction chain as the measurements with rf-excitation, using the GRAPPA-weights obtained from the acquisition with excitation. Noise quantification was based on histograms of regional pixel intensities in the noise images [4].