TITLE: A study for developing, optimizing, and evaluating a 4D maximum a posteriori rescaled-block iterative (MAP-RBI) EM image reconstruction method for gated myocardial perfusion SPECT
AUTHORS: Taek-Soo Lee, W. Paul Segars, and Benjamin M. W. Tsui
PURPOSE: We developed and optimized a 4D maximum a posteriori rescaled-block iterative (MAP-RBI) EM image reconstruction algorithm with space-time Gibbs priors and compensation of image degradation effects for application to gated myocardial perfusion (GMP) SPECT for improved simultaneous visualization of myocardial perfusion and motion defects. We also investigated the use of image bias and noise level as image quality metrics to evaluate the 4D MAP-RBI-EM method by utilizing human observer study.
METHODS AND MATERIALS: The 4D MAP-RBI-EM method includes 4D space-time Gibbs priors which enforce smoothing in both space and time domains while preserving the edges in the reconstructed image. To describe the Gibbs prior, we defined clique structures and a generalized potential function (GPF) with its three parameters, i.e., d, a and g, which control the shape of the potential function. To evaluate the 4D MAP-RBI-EM method for application to Tc-99m Sestamibi GMP SPECT, we used simulated projection data from the 4D NURBS-based Cardiac-Torso (NCAT) phantom whose cardiac cycle was divided into 8, 16, 24 and 36 gates. For each gate, noise-free and Poisson noise-added SPECT projection data were generated using an analytical projector that included the effects of attenuation, collimator-detector response (CDR) and scatter. To seek the optimal ranges of the parameters for the GPF, we investigated wide ranges of the parameter combination for the space-time Gibbs priors and determined the optimized values of the parameters as those that minimize both bias and noise level. We also determined iteration number beyond which the noisy reconstructed image did not change significantly. We calculated bias between the noise-free image estimate and its corresponding phantom slices, and normalized standard deviation (NSD) as measures for noise level. Then, The 4D MAP-RBI-EM reconstructed images using these optimal parameters with and without corrections of attenuation, CDR and scatter were compared to those from the 3D FBP and 3D OS-EM methods with and without corrections followed by an optimized 4D linear filter. The performance of 4D MAP-RBI-EM method was also evaluated through the use of task-based human observer study. We used a population of realistic 4D NCAT phantoms modeling variations in patient anatomy, organ uptakes, and cardiac motion. Half the population was normal and the other half had hypokinetic cardiac motion abnormalities. The noise-free and noisy projection data with 16 cardiac gates were generated and reconstructed using the same procedure described above. The observers were trained to the simulated gated SPECT images animated with a realistic real-time frame rate and were instructed to rate their confidence on the absence or presence of a motion defect on a continuous scale from 1 to 5. We applied receiver operating characteristic (ROC) analysis and used the area under the ROC curve as an index of comparison.
RESULTS: The results show that both 3D OS-EM and 4D MAP-RBI-EM methods lowered reconstructed image noise level as compared to FBP without correction. With the optimal parameters for each gating scheme, 4D MAP-RBI-EM was found to give significantly reduced image noise level while retaining resolution at higher iterations and at higher number of gates, as compared to 3D OS-EM with the post filter. The addition of attenuation, CDR, and scatter correction provided additional substantial improvement in image quality in terms of improvement in both reconstruction image resolution and noise. The results of human observer study showed significant differences in detection performance among the different bias-noise level combinations. Images obtained from the optimized 4D MAP-RBI-EM with corrections, which showed lower bias and lower noise level, gave better human observer detection performance among the other image reconstruction methods.
CONCLUSIONS: We conclude that, with the selection of optimal parameters, the 4D MAP-RBI-EM method with corrections of image degrading factors provides significant improvement in resolution and reduced noise level in the gated myocardial SPECT images allowing the possibility of improved temporal resolution through the use of more cardiac gates. Also, the bias-noise level model can be used to guide the determination of optimal parameters for the 4D MAP-RBI-EM method.
FUNDING SOURCES: This work was supported by the Public Health Service grant R01 HL68075.