Comparison of Gamma-Regularized Bayesian Reconstruction to Segmentation-Based Reconstruction for Transmission Tomography

For PET and SPECT, it is important to obtain accurate attenuation ( ) maps of thorax under conditions where counts are low. We previously proposed (Hsiao, Rangarajan, Gindi; Proc. IEEE Med. Img. Conf., 1998) a Bayesian reconstruction method in which the prior is base on a gamma mixture model. This method models the object as comprising attenuation coefficients that cluster into few (e.g. soft tissue, lung, bone) groups, and delivers both a reconstruction and a segmentation. We compared the performance of this algorithm with that of a more conventional method, segmentation attenuation correction (SAC) (Xu, Luk, Cutler, Digby; IEEE Tran. Nuc. Sci., v.41, pp. 1532-1537, 1994) that implicitly makes similar assumptions about the object. We simulated transmission acquisitions through a 2D thorax phantom at 140 KeV and 511 KeV with 1M counts in each case, and with (cm ) set to (0.15=soft tissue, 0.038=lung, 0.203=bone for 140KeV) and (0.096=soft tissue, 0.025=lung for 511KeV). Only the effects of low counts were considered. Performance (see table) was summarized at two energies (hi, lo) with two metrics: rmse of the reconstruction, and percentage of misclassified pixels (PMP). Initial results indicate better performance for the Bayesian method. We are currently investigating whether this advantage translates to the emission scan when tested with more realistic physical models and phantoms.