Simultaneous Multi-Dataset PET Image Reconstruction Using a Progressive Mutually Weighted Quadratic Penalty

There are many contexts in which a patient is scanned with positron emission tomography (PET) multiple times, for example in oncology treatment monitoring scans, or to obtain complementary information from multiple radiotracers. As is common in PET, such scans often suffer from high noise due to the counts-limited nature of PET acquisitions. Reduction of this noise while preserving scan-specific features by sharing information between datasets is the subject of ongoing research. In this work we adapt a recently proposed joint PET- MR reconstruction method to the multi-dataset PET context in order to reduce PET image noise across data-series. The method calculates similarity weights between neighbouring voxels in each image at each iteration and uses these to update the weighted quadratic penalty (WQP) for use in the next iteration. By iterating the two-step process of weights calculation and a penalised PET reconstruction update, the mutually weighted quadratic penalty (MWQP) method is defined. The proposed method was applied to a 2-scan, 2D, [18F]fluorodeoxyglucose brain cancer treatment response simulation study. Results indicate that using the MWQP method allows for reduced reconstructed image noise (up to 89%) while avoiding bias within scan-unique features, such as tumours. In this respect the method is superior to maximum-likelihood expectation maximisation reconstructions with a Gaussian smooth, total-variation regularised reconstructions, and traditional guided image reconstruction techniques that only allow a one-way transfer of information between datasets. Reconstructed visual image quality reflects the quantitative results, demonstrating lower image-wide noise and higher tumour contrast These results suggest that using a MWQP reconstruction approach improves reconstructed image quality beyond that achieved with conventional methods. Future work will involve further characterisation of the method and testing on clinical datasets.