Image-Domain Bootstrapping of PET Time-Course Data for Assessment of Uncertainty in Complex Regional Summaries of Mapped Kinetics

Imaging biomarkers extracted from diagnostic PET imaging scans are increasingly used in clinical decisions about the treatment of individual cancer patients. In this setting, measurements of uncertainty in the biomarker information presented for an individual patient may have an important role. In theory, the non-parametric bootstrap, based on resampling list-mode data, provides a solution to this problem. However, the computations have limited its use in practical setting, particularly for dynamic studies and scanners using iterative reconstruction. In recent work, our group has developed an image-domain bootstrapping technique that gives the ability to efficiently process 4-D dynamic PET data. This has been used to map uncertainties in images of tracer kinetics. This paper explores the utility of this approach for the evaluation of sampling characteristics of more complex imaging biomarkers - regional maximum and regional coefficient of variation - of mapped kinetics. A series of numerical simulations matched to two dynamic PET imaging studies with FDG and FLT in brain and breast cancer patients are carried out. A large (>650) collection of ROIs with varying size and location are considered. True target values of uncertainty are evaluated by study replication. Both the non-parametric projection-domain and the novel image-domain bootstraps are evaluated. Comparisons across a range of mapped kinetic variables are considered. The results find that the accuracy of the image-domain assessment of uncertainty is very acceptable - within 10% of the accuracy of the non-parametric bootstrap approach. The image-domain bootstrap gives a potential to practically evaluate the uncertainties of complex biomarkers recovered from the analysis of information in an individual patient PET study.