Performance of ordered-subset reconstruction algorithms under conditions of extreme attenuation and truncation in myocardial SPECT

We studied the bias and variance characteristics of the ordered-subset expe ctation maximization (OSEM) and rescaled block-iterative EM (RBIEM) iterati ve reconstruction algorithms in myocardial SPECT under extreme, but realist ic, conditions. Methods: We used the 2-dimensional mathematic cardiac torso phantom to simulate 2 patient anatomies: a large male with a raised diaphr agm and a female with large breast size, approximating extreme cases of att enuation conditions found in the clinic. For each anatomy, realistic (TI)-T -201 projection data were simulated for a 180 degrees acquisition are. Thre e cases of truncation for a 90 degrees-configured dual detector system were simulated: no truncation, moderate truncation, and extreme truncation. For each case, an ensemble of 250 noise simulations was generated, and each no isy dataset was reconstructed with the OSEM and RBIEM algorithms. The recon structions modeled only the effects of nonuniform attenuation and used a ra nge of subset configurations. Over the ensemble, we computed means and vari ances of activity in 8 regions of interest (ROIs) in the heart as a functio n of iteration. Results: Under conditions of no truncation and moderate tru ncation, the results from OSEM and RBIEM were very close to those from maxi mum-likelihood EM (MLEM); in all cases, the difference in ROI means was <2. 5%, For extreme truncation, the errors increased to as much as 11% with OSE M, but these were no greater than the errors for MLEM under the same condit ions. The OSEM algorithm with 2 views per subset was found to result in muc h higher variance of ROI estimates for the same bias as compared with RBIEM or OSEM with 4 or more views per subset. Conclusion: The OSEM and RBIEM al gorithms are at least as robust to highly attenuating patients and truncati on as MLEM algorithm and can be adequate substitutes for MLEM, even in extr eme cases. Clinical users should apply the smallest number of subsets that can be accommodated by allowable processing time to reduce image noise and variance in quantitative estimates.