MR-Resolution Kernel Method for PET Reconstruction

Conventional PET reconstruction produces noisy images. Recently proposed techniques such as the MR-guided kernel method have been employed to reduce the impact of noise, whilst retaining important image details. However, this can lead to over smoothing of PET unique features. To address this issue, this work extends the MR-guided kernel method to use MR resolution basis functions, which are extracted from an MR image at its native resolution. Furthermore, this MR-resolution kernel method is modified to produce spatially constrained basis functions in order to limit the smoothing of PET-unique features whilst still reducing the impact of noise. The MR-resolution kernel reconstruction is compared to MLEM and conventional PET resolution kernel methods for tumour contrast recovery. These methods are applied to real patient FDG data augmented with simulated tumours. The proposed kernel method shows an improved contrast to noise ratio compared to the conventional kernel method for all tumour sizes. However, MLEM attained a higher contrast to noise ratio for the small tumour. In summary, the MR resolution spatially constrained kernel method maintains the noise reduction properties of the conventional kernel method implementation, whilst better retaining the features unique to the PET data.