Spatial covariance characteristics in a collection of 3-D PET scanners used in clinical imaging trials

An NCI-sponsored program to allow qualification of PET imaging sites for use in Cancer Clinical trials has created a data set, assembled and maintained by the American College of Radiology Imaging Network (ACRIN), with a rich collection of PET phantom measurements assessing imaging quality. Previous work with the data has focused on systematic deviations between the actual activity value in the phantom and its measurement by an average of PET-recorded voxel values in the region of interest corresponding to the phantom in the scanner. But the data also allow for more detailed evaluation of imaging characteristics. Our work focuses on components of random variation. Dynamic data considered are obtained from 3-D sequential scanning, typically with variations on OSEM for reconstruction, of a uniform cylindrical phantom over a 25 minute period. We evaluate axial, transaxial and temporal patterns in variance and covariance. Variance characteristics are dominated by effective counts - these are lower over short time-frames and in the axial extremes of the scanner bed. After adjustment for variance the spatial auto-correlation patterns in 3-D are evaluated. Auto-correlation is decomposed as a product of axial and trans-axial effects. The trans-axial pattern follows the structure generally associated with standard 2-D filtered back-projection reconstruction - i.e. largely determined by the impulse response. Having a phantom based measurement of the variance and auto-correlation patterns gives the possibility to make more efficient use of region of interest data from patient scans. Usual regional averages can be replaced by weighted averages, with weights inversely proportional to the local variance. In addition, an approximate standard error for an ROI average can take account of the phantom derived measurement of the auto-correlations. Thus information from routine phantom scanning would practially enhance the value of information recovered for patient studies.