JOINT OPTIMIZATION OF K-SPACE SAMPLING AND RECONSTRUCTION FOR MULTI-CONTRAST MRI

Deep learning-based multi-contrast MRI reconstruction can have outstanding performances because it utilizes the complementary information between images of different contrasts. However, existing methods only use feature fusion in the re construction network and rarely consider joint optimization of the sampling trajectories and the reconstruction. Moreover, due to the probable scanning alterations between visits of the same patient and the issues of data storage and communication, it can be difficult to pair the previously acquired images to assist the reconstruction of new scans. In this paper, we propose a method that jointly optimizes the sampling trajectories and the reconstruction networks of multi-contrast MR images in the same visit to obtain the optimal reconstruction. In the reconstructor, feature fusion from different images is placed in the decoder and the temporal information is added. Meanwhile, for a specific acceleration factor, we propose learnable acceleration ratio which can allocate different sampling budgets to each acquisition of the different contrasts. Our proposed method outperforms the baseline approach for multi-contrast MRI reconstruction.