Abstract
1685
Objectives: To improve our previous work (S.M. Kim, et al, J. Nucl. Med., 48(Sup2), 91P, 2007) for 3-D Compton camera reconstruction, we evaluated performance of a row-action maximum likelihood algorithm (RAMLA) proposed by Browne and De Pierro (IEEE Trans. Med. Imag., 15(5), pp. 687-699,1996), and developed a method for choosing relaxation parameters.
Methods: In RAMLA, performance of the algorithm depends highly on the choice of relaxation parameters. Here the relaxation parameter was set to a value inversely proportional to the element of the system matrix so that the geometrical characteristic of the camera is applied to the relaxation parameter. To test 3-D RAMLA, Compton scattered data acquisitions were simulated through a 6-cylinder software phantom. Performance of RAMLA was evaluated with percentage error of the reconstruction and compared with that of the expectation maximization (EM) algorithm.
Results: Initial results indicate significantly better performance for RAMLA. Our proposed method for choosing relaxation parameters revealed dramatic improvement in convergence rate as well as in quality of reconstruction. For one-directional acquisitions with a fixed detector pair, 1 iteration of RAMLA performed as well as 89 iterations of EM. For three-directional acquisitions in the x-, y- and z-axes, 1 iteration of RAMLA performed as well as 143 iterations of EM. Computational costs of one iteration of RAMLA and EM were about the same.
Conclusions: RAMLA significantly outperforms EM in convergence rate. It can also be deduced that RAMLA with optimal relaxation parameters can even outperform OSEM (ordered subsets EM) with moderate subsets. Since RAMLA converges to an ML solution, while OSEM does not, it will be clearly more useful in 3-D Compton camera reconstruction.
- Society of Nuclear Medicine, Inc.