Optimising Subset Selection in Synergistic Emission Tomography Reconstruction

This study investigates subset selection strategies for the synergistic reconstruction of Yttrium-90 PET/CT and SPECT/CT NEMA phantom data acquired on a triple-modality scanner. We previously introduced a smoothed directional total nuclear variation (TNV) prior, which extends the standard TNV prior with an anatomically informed, directionally dependent operator and can be used in gradient-based reconstruction algorithms. We used a modified block-sequential regularised expectation maximisation algorithm and its stochastic variants. Stochastic variance reduction improved convergence rates for PET but showed slower convergence in all but the smallest NEMA sphere for SPECT. Treating modalities as separate subsets yielded convergence rates comparable to joint updates, suggesting the number of subsets can be optimised for each modality. However, treating the prior as a separate subset was less effective, particularly for the largest sphere due to overshooting. This could be mitigated in future work with improved step size selection and preconditioning.