Infrastructure for efficient multi-component uncertainty estimation of SUVR values in amyloid PET imaging

Objectives In PET imaging, any statistic derived from a measurement, like the standardised uptake value ratios (SUVR), is subject to uncertainty due to limited statistics. This work addresses the problem of efficient uncertainty estimation of any PET image statistic in the presence of a highly complex chain of image reconstruction and processing. This is particularly useful in the study of longitudinal SUVR changes, which are subtle and sensitive to noise.

Methods The uncertainty estimation is achieved through the development of an innovative, self-contained and efficient GPU-based infrastructure with all the necessary components such as: (1) histogramming, which rapidly generates bootstrap resampled sinograms from raw PET list mode datasets; (2) detector normalisation and attenuation correction using pseudo-CT mu-maps for the PET/MR scanner used; (3) novel 3D scatter estimation for large axial field of views; (4) variance-reduced random events estimation; (5) fast matched forward and back projectors, adapted for accelerated image reconstruction using 14 ordered projection subsets. (6) robust MR-PET image registration which uses symmetric affine scheme based on block-matching approach; (7) MR-based image parcellation into 144 brain regions using multi-atlas segmentation propagation strategy. The estimated uncertainties are found by accounting for random errors as they propagate through all the above components. The components can be considered jointly or individually to study their influence on the final noise properties of any given image statistic (here the SUVR). The above pipeline was applied to the list-mode data of a 50 minute dynamic PET scan with F-18 florbetapir injected to a patient with a probable Alzheimer's disease and acquired using the Siemens Biograph mMR. Two time frames were considered: (1) an early 15 second time frame with low count level and (2) 10 minute time frame at the end of the scan. 100 bootstrap realisations were generated for both time frames based on which the distributions of uncertainties were formed.

Results The whole processing time for each bootstrap sample with all the above components takes approximately 4 minutes. Voxels in such areas like the cerebellum and brain stem had higher variance than the posterior or anterior cingulate gyrus, however, when mean regional values were considered, cerebellum grey matter was one of the least variable due to its bigger size. The SUVR value for the 10 minute time frame and the whole cingulate gyrus was 1.127 and the estimated standard deviation (SD) was +/- 0.014 with all the components of the pipeline being accounted for. The SD was reduced to +/- 0.006 when the estimated uncertainty was affected by only the random correction component. The regional and voxel SD was approximately 10 times higher in the 15 second time frame, which had a significant impact on the stability of the PET-MR co-registration, causing an isotropic registration wobble of up to 4 mm. This can have a considerable impact on the performance of partial volume correction which uses MR-based tissue segmentation as input.

Conclusions It was demonstrated, that the devised framework of highly complex processing components used for generating and processing bootstrap samples from PET list mode data is feasible. It enables joint and individual estimation of uncertainties for any component within a reasonable time. This multi-component uncertainty estimation can prove very useful in finding optimal image reconstruction and processing algorithms. Such an infrastructure can help in robust estimation and comparison of SUVR values especially in longitudinal studies of neurodegenerative disease.

• Download figure
• Open in new tab
• Download powerpoint