Sampling Pattern Augmentation to Enhance Deep Learning-based Image Reconstruction of MRI

The application of compressed sensing (CS) to MRI enables the acceleration of imaging times. CS can approximate an original signal from measurements that do not satisfy the sampling theorem. While CS can reduce the sampling time of k-space signals by reconstructing an MR image from limited measurement signals, there are still issues such as reconstruction time and optimal parameter search due to the CS optimization algorithm. In contrast, unrolling-based network models that can learn the parameters of the CS optimization algorithm to appropriate values for the subject are proposed, and the reconstruction performance is improved by applying the model to CS. However, the unrolling-based network models are trained based on a specific sampling pattern, and the reconstruction performance is adversely affected when the reconstruction is carried out using an untrained pattern. In this study, a sampling pattern augmentation method that employs multiple sampling patterns for learning is proposed as a means of developing a model that is not dependent on specific sampling patterns. Simulation experiments demonstrated that the reconstructed images using the sampling pattern augmentation exhibited superior quality compared to those reconstructed using conventional methods, while requiring a significantly shorter learning time. Moreover, it was confirmed that the proposed method yields high reconstruction performance even when reconstructing with untrained patterns.