Optimization of high resolution PET iterative reconstruction with resolution modeling for image derived input function

Quantification of brain PET is traditionally carried out using arterially sampled input functions, IFs. Bloodless alternatives in the form of image-derived input functions, IDIFs, have thus far been unable to provide accurate IFs for brain PET studies due to partial volume effects. Presently, no study has been carried out to estimate how many iterations should be used with iterative reconstruction incorporating resolution modelling for the extraction of IDIFs from high resolution FDG brain PET data. In this study, IDIFs were obtained in three subjects from the carotid arteries, CA, and the superior sagittal sinus, SSS, using varying numbers of iterations. IDIFs were extracted as the mean value within each region (CA-A and SSS-A) and after applying a threshold to each region (CA-B and SSS-B). The IDIFs were compared in terms of area under the curve, AUC, and the influx constant Ki to a population-based input function, PBIF, scaled using three late venous blood samples. All IDIFs underestimated the AUC of the PBIF and generally showed closer agreement for the SSS IDIFs than the CA IDIFs. For both blood pools, method B resulted in a larger AUC. The Ki estimates obtained with the SSS IDIF approached convergence around 40 iterations, coming on average to within 3% of the PBIF K_i at 40 iterations. The Ki estimates from the CA IDIFs didn't converge in two of the three subjects and even after 120 iterations there remained a 20% difference with the PBIF. These initial investigations show that IDIFs for FDG could be extracted from the SSS on images acquired with the HRRT scanner and reconstructed using motion correction and resolution modeling with 40 or more iterations. Larger group sizes must be used to determine the accuracy of this method and confirm the convergence properties observed.