Joint attention deep learning reconstruction of highly-accelerated pre- and post-contrast T1-weighted 3D images of brain tumors

Motivation: The efficiency of brain tumor MRI can be improved by exploiting correlations between pre- and post-contrast repetitions. Goal(s): Develop a deep joint reconstruction network to exploit correlations between pre- and post-contrast 3D T1-weighted without explicit k-space data consistency for highly-accelerated brain tumor MRI. Approach: A multi-scale Transformer was employed to exploit the attention mechanism for long-distance correlations within and across image contrasts. 8-fold accelerated brain tumor patient data was used to train and evaluate the network. Results: The joint-attention transformer reconstruction outperformed convolutional Unet and transformer trained on individual contrasts. Impact: A joint-attention deep learning reconstruction method can exploit correlations across sequences and enable significant reductions in MRI protocols.