Task-based requirements for resampled data in nuclear medicine

Bootstrap resampling offers a means of efficiently generating multiple sets of tomographic data from a single acquisition. We have investigated how the quality of this acquired data affects the performance of tumor quantitation and detection tasks carried out with the reconstructed images. The Poisson parametric bootstrap was applied in a simulation that modeled SPECT Ga-67 thoracic imaging. A set of low-noise Monte-Carlo projections for the MCAT phantom was the basis for creating classes of data defined by Poisson means with varying amounts of resampling bias. Images were reconstructed with the OSEM algorithm. Quantitation of tumor activity in the images was performed with an ROI estimator, while localization ROC detection studies employed a model observer. With either task, the results indicate relatively small differences in performances due to the resampling. A definition of "noise-free" Monte-Carlo acquisitions for this particular imaging simulation was also determined. These study procedures are currently being extended to nonparametric bootstrap methods applied to PET and time-of-flight PET imaging.