ROC study of maximum likelihood estimator human brain image reconstructions in PET clinical practice

This paper reports on the progress to date in carrying out Receiver Operating Characteristics (ROC) studies comparing Maximum Likelihood Estimator A(MLE) and Filtered Backprojection (FBP) reconstructions of normal and abnormal human brain PET data in a clinical setting. A previous statistical study of reconstructions of the Hoffman brain phantom with real data indicated that the pixel-to-pixel standard deviation in feasible MLE images is approximately proportional to the square root of the number of counts in a region, as opposed to a standard deviation which is high and largely independent of the number of counts in FBP A preliminary ROC study carried out with 10 non-medical observers performing a relatively simple detectability task indicates that, for the majority of observers, lower standard deviation translates itself into a statistically significant detectability advantage in MLE reconstructions. The initial results of ongoing tests with four experienced neurologists/nuclear medicine physicians are presented. Normal cases of 18F-fluorodeoxyglucose (FDG) cerebral metabolism studies and abnormal cases in which a variety of lesions have been introduced into normal data sets have been evaluated. The authors report on the results of reading the reconstructions of 90 data sets, each corresponding to a single brain slice. It has become apparent that the design of the study based on reading single brain slices is too insensitive and the authors propose a variation based on reading three consecutive slices at a time, rating only the center slice