BACKGROUND

Magnetic resonance fingerprinting (MRF) is a rapid imaging technique for simultaneous mapping of multiple tissue properties such as T1 and T2 relaxation times. However, conventional pattern matching reconstruction and iterative low rank reconstruction methods may not take full advantage of the spatiotemporal content of MRF data and can require significant computational resources with long reconstruction times. Deep learning reconstruction using a three-dimensional (3D) convolutional neural network (CNN)-based method may enable high-quality, rapid MRF reconstruction. Evaluation of such proposed deep learning reconstruction methods for MRF is needed to clarify whether deep learning techniques adapted from other MR image reconstruction problems will yield benefits when employed in MRF applications. The objective of this study is to design and evaluate a novel deep learning framework (SuperMRF) that directly transforms undersampled parameter-weighted 3D Cartesian MRF data into quantitative T1 and T2 maps, bypassing traditional pattern-matching in MRF.

METHODS

In contrast to conventional MRF where only the temporal evolution of each voxel is used for quantification, SuperMRF exploits both two-dimensional spatial and one-dimensional temporal information with a 3D CNN for reconstruction. Controlled simulation experiments were performed using reference parameter maps from knee scans of healthy volunteers. To evaluate the robustness to noise, we trained our network using clean data and tested it on simulated noisy data. Conventional inner product-based pattern matching and state-of-the-art iterative low rank reconstruction techniques were used for comparison. The performance of all methods was evaluated with respect to structural similarity index (SSIM), peak signal-to-noise ratio (PSNR), and normalized mean squared error (NMSE). Prospective real-world MRF scans were performed in four volunteer subjects using the trained network from simulations and cartilage and muscle T1 and T2 values were compared between conventional pattern matching, low rank reconstruction, and SuperMRF.

RESULTS

SuperMRF estimated accurate T1 and T2 mapping in a highly accelerated scan (15× undersampling in k-space with a 20-fold reduction in the number of acquired MRF frames) with low error (NMSE of 5%) and high resemblance (SSIM of 94%) to reference quantitative maps. SuperMRF was observed to be superior to the conventional and low rank MRF reconstruction methods in terms of NMSE, SSIM, and robustness to noise. In prospective real-world data, SuperMRF provided comparable T1 and T2 maps as compared to low rank MRF. The only significantly different cartilage and muscle values in prospective data across the three reconstruction methods were those from conventional MRF T2.

CONCLUSIONS

Our results demonstrate that the proposed SuperMRF can achieve rapid, robust reconstruction with reduced frames in addition to k-space undersampling, outperforming the conventional and state-of-the-art reconstruction methods in simulation and providing comparable results to low rank reconstruction in prospective real-world subjects.