Unified Deep Learning for Simultaneous Cardiac Cine MRI Reconstruction, Motion Estimation and Segmentation

Deep learning methods have achieved great success in cardiac cine MRI reconstruction, motion estimation, and segmentation. However, most studies considered the three tasks separately. While benefits have been shown for dual-task learning, such as motion-compensated cine MRI reconstruction and motion-guided cine segmentation, a framework considering all three tasks to more effectively exploit inter-task dependency remains uninvestigated. To fill this gap, we propose a unified approach that adopts groupwise image registration to estimate motion between cine frames, which is leveraged to assist reconstruction and segmentation. In particular, motion-augmented images enhance the reconstruction task, and motion-generated annotations offer additional supervision for the segmentation task. The improved reconstruction and segmentation will in turn help to yield more accurate motion estimation. We embed the three tasks in an unrolled framework, enabling iterative coarse-to-fine refinement of predictions across all tasks. Experimental results on a cine MRI dataset show the superior performance of the proposed method compared with the single-task and dual-task baseline methods.