Magnetic resonance spectroscopic imaging (MRSI) is a useful technique for assessing the in vivo spatial profiles of various key metabolites like N-acetylaspartate (NAA), creatine, choline, and lactate. Changes in metabolite concentrations can help distinguish between healthy and diseased tissues, thus, providing diagnostic and prognostic information to the clinician, potentially leading to improved treatment strategies. However, MRSI scans are not accommodated in the current clinical acquisition window due to the lengthy scan time involved – a 16x16 spatial grid would require 256 phase encoding gradients to be played out. This results in a scan time of ~9-12 min depending on the TR, with a further increase in time for larger volumes. Such long scan times are particularly not desirable when imaging pediatric cases, particularly those under anesthesia. Acceleration techniques like compressed sensing (CS)^1-2 allow a faithful reconstruction of the data even when the kspace is undersampled well below the established Nyquist limit, and CS has been adopted in MRSI reconstructions to reduce the scan time³. The objective of this study was to evaluate the fidelity of the CS reconstructions when applied to the MR spectroscopic imaging of pediatric patients.All MRSI data was acquired on a Philips 3T Ingenia MRI scanner. Fully sampled MRSI data was collected on patients in the age range of 1 - 10 years, with the following acquisition parameters – 16x16x2048 grid, TE/TR = 46/1500 ms, 1 average, 10 mm slice thickness, total scan time for the fully sampled reference 1X dataset = 9 min. The kspace was pseudo-randomly undersampled to generate datasets that represented 50% (2X), 33% (3X), 25% (4X), and 20% (5X) acceleration. Retrospectively undersampled MRSI datasets were reconstructed in Matlab by employing the conjugate gradient algorithm to solve the convex optimization problem: argmin∥F_um−y∥₂+λ_L1∥Wm∥₁+λ_TVTV(m). The undersampled reconstructions were quantitatively compared with the 1X dataset in terms of the SNR and peak amplitudes of various brain metabolites. The error in reconstruction was quantified using the normalized root mean square error (nRMSE) metric. Minimal post processing was applied to the reconstructed datasets in jMRUI⁴ namely, apodization, phase correction, and removal of residual water, and the corresponding metabolite maps were generated.Figure 1 shows the metabolite maps for NAA, creatine, and choline for various acceleration factors 1X-5X in a 9 year old female patient with no abnormalities detected. Figure 2 depicts the metabolite maps following quantification in a 10 year male with a brain tumor. In both cases, the CS reconstructions for various acceleration factors demonstrate good fidelity with the fully-sampled reference dataset, with negligible reconstruction errors. The error in reconstruction, as quantified by the normalized RMSE, is illustrated for both cases in Figure 3 and is below 2% for reconstructions up to 5X. All twelve patient datasets retrospectively tested using the CS-MRSI reconstruction algorithm showed very low reconstruction errors till 5X, indicating the robustness of the reconstruction.The feasibility of CS-MRSI in pediatric patients has been demonstrated, with the accelerated reconstructions showing negligible errors. Further studies currently underway focus on prospective CS acquisitions on the scanner and on improvements in the reconstruction algorithm.