Purpose: To compare the performance of the deep learning (DL) based method and the total generalized variation (TGV) based iterative reconstruction (IR) method for fast magnetic resonance imaging (MRI).
Methods: T1 and T2-weighted images of brain, breast, cervix and prostate were included in the evaluation. The under-sampled k-space data were retrospectively sparsely sampled from the full k-space data using a cartesian-based under-sampling mask that is to simulate an accelerated image acquisition by a factor of 4, mimicking 75% reduction of scan time. Then, the under-sampled data were reconstructed to yield images using a deep-cascaded (DC-) convolutional neural network (CNN) and the TGV method, respectively. For DC-CNN, the model was trained separately for each site, and the evaluation data were not included in the training data. Images reconstructed from the fully sampled data were used as the reference for comparison. The reconstructed images by the DC-CNN and the TGV methods were evaluated quantitatively using structure similarity (SSIM) and total relative error (TRE).
Results: The images reconstructed by DC-CNN generally show improved quality than those by the TGV method. Specifically, DC-CNN reconstructed images show less blurred effect and less artifacts than the TGV-based reconstruction for all the four clinical sites. Quantitatively, images reconstructed using DC-CNN have greater SSIM and less TRE values than those by TGV in all cases. The reconstruction time for DC-CNN is much less than the TGV method, which is 0.2s and 176.2s for each slice, respectively. Furthermore, the DC-CNN method demonstrates good generalizing ability across various clinical sites.
Conclusion: The results have demonstrated that DL-based method shows superior to the TGV method for fast MRI in both image quality and reconstruction time. Yet DL-based method should be further tested for more clinical sites such as spine and other clinical applications such as diffusion imaging, MR angiography and DCE-MRI.