Implementation of linear filters for iterative penalized maximum likelihood SPECT reconstruction

Six low-pass linear filters applied in frequency space were implemented for iterative penalized maximum-likelihood (ML) single photon emission computed tomography (SPECT) image reconstruction. The filters implemented were the Shepp-Logan filter, the Butterworth filter, the Gaussian filter, the Hann filter, the Parzen filter, and the Lagrange filter. The low-pass filtering was applied in frequency space to projection data for the initial estimate and to the difference of projection data and reprojected data for higher-order approximations. The projection data were acquired experimentally from a chest phantom consisting of nonuniform attenuating media. All the filters could effectively remove the noise and edge artifacts associated with the ML approach if the frequency cutoff was properly chosen. The improved performance of the Parzen and Lagrange filters relative to the others was observed. The best image. by viewing its profiles in terms of noise-smoothing, edge-sharpening, and contrast, was obtained by Parzen filter. However, the Lagrange filter has the potential to consider the characteristics of the detector response function