Efficient Training of 3D Unrolled Neural Networks for MRI Reconstruction Using Small Databases

3D MRI encodes volumetric information, typically offering improved contiguous coverage and resolution than 2D MRI. However, 3D MRI data acquisition is lengthy, and requires accelerated imaging techniques. Deep learning methods have recently emerged as a powerful strategy for MRI reconstruction. Among such methods, unrolled networks have proven powerful with their ability to incorporate the forward encoding operator directly. These methods are largely applied in a 2D setting, but 3D processing has the potential to further improve reconstruction quality for volumetric imaging by capturing multi-dimensional interactions. Nevertheless, implementing 3D unrolled networks is challenging because of memory limitations on GPUs, as well as the lack of large databases of 3D k-space data. To tackle both of these issues, we propose a data augmentation strategy that generates smaller sub-volumes from large volumetric datasets. Subsequently, these augmented datasets are used to train a 3D unrolled network, and compared to their 2D counterpart. The results show that our 3D processing provides improved reconstruction results on volumetric data than 2D processing.