Ga-67 Tumor Detection using Penalized-EM with Nonanatomical Regularizers

Results from localization ROC (LROC) observer studies evaluating a penalized-likelihood SPECT reconstruction algorithm are presented. Both human and numerical observers were employed for the task of Ga-67 tumor detection in simulated chest images. Projection data from a mathematical phantom were generated with the SIMIND Monte Carlo program. Reconstructions using De Pierro's modified-EM algorithm [1] with the 3D quadratic and Huber penalty functions were performed. Iterative corrections for nonuniform attenuation, collimator response, and in-patient scatter were included. The LROC studies with these images provided an initial validation of a channelized nonprewhitening (CNPW) numerical observer that explicitly performs the tumor search task, and at the same time demonstrated the usefulness of frequency-selective channels for modeling human detection-task performance. Results drawn from the CNPW observer indicate that for the range of iterations tested, relatively low levels of noise regularization are desirable in the penalized-likelihood images.