In this study we evaluated pre- (PR-R) and post-reconstruction (PO-R) count-dependent Metz filters for PET brain studies in order to increase signal-to-noise ratio. A set of studies using a 3D Hoffman brain phantom was performed at various count levels, and a reference image set was created from extremely high count images. Several combinations of PR-R and PO-R filtering were considered to find the optimal means of processing, including: Hann filter alone; PR-R Metz filter without or with a PO-R low pass filter; and PO-R Metz filter without or with a PR-R low pass filter. A formula was established to correlate the optimal PO-R Metz filter order with the net counts. Resolution [full width at half-maximum (FWHM) and fill width at tenth maximum (FWTM)], normalized residual mean square differences (NRMSD) between the ideal and the processed images, noise reduction and contrast were used as parameters for the evaluation of the different filter combinations. Resolution is decreased by all filter combinations that can effectively control noise; however, FWTM increases less than FWHM. NRMSD indicates that the use of Hann and (optimal) PO-R Metz filter is the most powerful combination from among those tested. A close correlation (r = 0.969) was found between the net counts and the optimal order of the PO-R Metz filter. At the count levels of clinical studies the PO-R Metz filter was found to control noise much more effectively and enhance the contrast when compared to the routinely used Hann filter alone, and produced images of better quality.