An Efficient Data-Model Dual-Drive Algorithm for Compressed Sensing MRI

Traditional compressed sensing MRI methods that focus on constructing better prior regularizations or numerical iterative optimizations usually suffer from heavy computational burden.Recently developed deep learning based approaches rely too much on the selection of training data and deep architecture,thus have poor abilities of generalization.To address these issues,we propose an efficient and robust algorithm to achieve the balance between reconstruction accuracy and efficiency.We construct a model-driven priori expression process and a data-driven prediction process for details restoration and artifacts correction,in a complementary perspective,realizing an integration of domain knowledge and deep representation.Further,the iteratively alternating mechanism ensures that the output propagation can be corrected in time and guided towards the desired solution in expected direction.Detailed experiments on T1 and T2 weighted data demonstrate that compared with the state-of-the-art,our method achieves higher reconstruction accuracy for all three kinds of sampling patterns and five sampling ratios,as well as higher computation efficiency on both GPU and CPU.Further experiments show that our method provides stronger robustness to data variations and noise pollution.