Conventional magnetic resonance spectroscopic imaging requires long acquisition times. Echo planar spectroscopic imaging (EPSI) significantly reduces the scan time but is limited by conventional phase encoding. Non-uniform sampling and compressed sensing (CS) reconstruction can further accelerated 3D EPSI. We applied a Perona-Malik (PM) non-linear diffusion algorithm for CS reconstruction of 3D EPSI data in both retrospectively and prospectively undersampled phantom and in-vivo data sets, and compared results with those using Total Variation (TV). Our pilot findings demonstrate that PM produces improved reconstruction results compared to TV. Furthermore, PM eliminates the need for parameter tuning, giving it a great advantage over TV.