Clinically important characteristics of maximum-likelihood reconstruction

SPECT images of a Jaszczak rod phantom, a single-slice Hoffman brain phantom and a uniform water-bath were acquired. Simulated noisy bar phantoms incorporating depth-dependent attenuation and blur were produced and compared to simulations with depth-independent attenuation and blur, as is the case in PET. Following iterative maximum-likelihood reconstruction, regularization was performed with use of Gaussian filters. While correction for attenuation is achieved in approximately 10 iterations, spatial resolution in the SPECT reconstructions, quantified by contrast in the bar simulations and by visual inspection of the real data, was highly nonuniform, being poorest at the center and improving toward the periphery. Image resolution continued to improve well beyond 50 iterations when regularization was applied that maintained a constant signal-to-noise ratio. Contrast in the simulated PET data also improved with increasing iterations, but the PET data showed uniform contrast throughout the transaxial slices at all numbers of iterations.