Abstract
Tau PET imaging is an essential imaging modality for the diagnosis and monitoring of Alzheimer's disease and related dementias. To enable tau PET imaging-based longitudinal monitoring of disease progression, further reducing the injected dose during each scan is important. In this work, we developed a novel deep learning approach that incorporated cross-modality transformer blocks to integrate both PET and MR prior information to further improve low-dose tau PET imaging. Both spatial and channel information were utilized during the calculation of cross-modality self-attention maps. Performance of the proposed method was evaluated based on the early-frame and late-frame images from 139 dynamic (18)F-MK-6240 tau PET datasets. Results showed that the proposed network can outperform other reference networks which concatenated PET and MR images together as the network input.