Longitudinal Multi-Dataset PET Image Reconstruction

In positron emission tomography (PET), a subject may be scanned two or more times to monitor longitudinal functional changes. These images often appear very similar except for localised, relatively small regions of change. This observation led to the proposal of the maximum a posteriori simultaneous longitudinal reconstruction (MAP-SLR) method, which reconstructs longitudinal datasets together with regularisation to encourage sparse difference images between pairs of scans. In this work we extend MAP-SLR to application to a multi-scan treatment response simulation study. To do this, five 2D [18F]fluorodeoxyglucose head scan datasets (designed to emulate a brain tumour longitudinal study) were simulated and then reconstructed with MAP-SLR. The resulting images were compared to: maximum likelihood expectation-maximisation (MLEM) reconstructions; longitudinally smoothed MLEM reconstructions; and MLEM applied to a reference dataset with five times the number of counts. When using MAPSLR, the noise (in terms of regional coefficient of variation) in a longitudinally unchanging white matter region was reduced and with sufficient regularisation these noise levels approached the high counts reference case. In the tumour, whilst a longitudinal bias is obtained with MAP-SLR, the bias is much smaller than that obtained when performing a noise-matched longitudinal smooth on MLEM reconstructions. With an appropriate level of regularisation the tumour bias is small enough to produce reconstructed images which preserve the longitudinal changes seen in the independent dataset MLEM reconstructions, but with noise reduction of 40% in regions which do not change. The results suggest that MAP-SLR is a simple and effective way of achieving noise reduction in longitudinal PET imaging. Future work will involve application specific testing and investigation into the inclusion of other longitudinally defined penalties into the simultaneous reconstruction.