Artifact suppression and quantitative accuracy of damped MLEM algorithm for SPECT imaging

The most disadvantageous artifacts that may accompany the maximum likelihood expectation maximization (MLEM) reconstruction are noise amplification and image deterioration near the edges of an area with sharp intensity change. In this paper, we investigate the performance of a damped MLEM (dMLEM) algorithm with a modified likelihood function [White R.L., J.SPIE Records, 2198, pp. 1342–1348, 1994] using simulated SPECT data. In particular, a trade-off between artifact suppression and quantitative accuracy of images is of our special interest. In our experiments, both noiseless and noisy numeric phantom data were created. Small cylindrical object with a 5cm diameter was located off-center in a large cylinder having a diameter of 16cm. Acquisition parameters were selected to model a typical SPECT clinical protocol. The images were reconstructed using both dMLEM (with two flattering parameters N=2 and N=3) and a conventional MLEM (corresponding to N=0) algorithm with up to 100 iterations. The modification of the likelihood function in dMLEM led to the suppression of both ringing and noise artifacts. This effect strongly depends on the damping parameter N. With increasing of N, the appearance of artifacts is delayed in the iteration process, but also the quantitative accuracy of images is degraded. Our simulations showed that slight damping of likelihood function in MLEM algorithm might serve as a reasonable tradeoff by allowing users for the suppression of artifacts while preserving the quantitative accuracy at an acceptable level. Especially, dMLEM method with N=2 allowed us to improve reconstruction convergence and avoid ringing artifacts in the first 20–30 iterations while quantitative accuracy of the recovered activity was degraded by only 3–4%.