LEADERS: LEARNABLE DEEP RADIAL SUBSAMPLING FOR MRI RECONSTRUCTION

Recently, deep learning approaches have shown great promise in learning MRI subsampling. The majority of existing works have focused on optimizing Cartesian or equipmentconstrained Gaussian-like subsampling, ignoring the question of learning radial subsampling. This paper proposes a simple learnable radial subsampling technique for compressed sensing MRI. The proposed approach exploits a radial subsampling for direct estimation of all radial spokes' weights from radial sampling space. The proposed Learnable Deep Radial Subsampling (LEADERS) method can be easily integrated with any deep learning-based reconstruction algorithm. This method can provide reliable estimates in a deep learning manner. The effectiveness of the generated radial subsampling patterns is verified on two deep learning-based reconstruction models, with a large-scale, publicly available brain MRI datasets for two downsampling factors (R = 4 and 8). The numerical and visual experiments demonstrate that the learned radial subsampling patterns can be applied for different deep learning reconstruction models with different subsampling rates, and shows more efficient and effective results than the ones reconstructed using existing handcrafted radial subsampling patterns.