ABSTRACT
INTRODUCTION: Deep learning-based approaches are increasingly being used for the reconstruction of accelerated MRI scans. However, presented analyses are frequently lacking in-detail evaluation of basal measures like resolution or signal-to-noise ratio. To help closing this gap, spatially resolved maps of image resolution and noise enhancement (g-factor) are determined and assessed for typical model- and data-driven MR reconstruction methods in this paper. METHODS: MR data from a routine brain scan of a patient were undersampled in retrospect at Râ¯=â¯4 and reconstructed using two data-driven (variational network (VN), U-Net) and two model based reconstructions methods (GRAPPA, TV-constrained compressed sensing). Local resolution was estimated by the width of the main-lobe of a local point-spread function, which was determined for every single pixel by reconstructing images with an additional small perturbation. G-factor maps were determined using a multiple replica method. RESULTS: GRAPPA showed good spatial resolution, but increased g-factors (1.43-1.84, 75% quartile) over all other methods. The images delivered from compressed sensing suffered most from low local resolution, in particular in homogeneous areas of the image. VN and U-Net show similar resolution with mostly moderate local blurring, slightly better for U-Net. For all methods except GRAPPA the resolution as well as the g-factors depend on the anatomy and the direction of undersampling. CONCLUSION: Objective image quality parameters, local resolution and g-factors have been determined. The examined data driven methods show less local blurring than compressed sensing. The noise enhancement for reconstructions using CS, VN and U-Net is elevated at anatomical contours but is drastically reduced with respect to GRAPPA. Overall, the applied framework provides the possibility for more detailed analysis of novel reconstruction approaches incorporating non-linear and non-stationary transformations.
