Deep Learning Initialized Compressed Sensing (Deli-CS) in Volumetric Spatio-Temporal Subspace Reconstruction.

Introduction: Spatio-temporal MRI methods enable whole-brain multi-parametric mapping at ultra-fast acquisition times through efficient k-space encoding, but can have very long reconstruction times, which limit their integration into clinical practice. Deep learning (DL) is a promising approach to accelerate reconstruction, but can be computationally intensive to train and deploy due to the large dimensionality of spatio-temporal MRI. DL methods also need large training data sets and can produce results that don't match the acquired data if data consistency is not enforced. The aim of this project is to reduce reconstruction time using DL whilst simultaneously limiting the risk of deep learning induced hallucinations, all with modest hardware requirements. Methods: Deep Learning Initialized Compressed Sensing (Deli-CS) is proposed to reduce the reconstruction time of iterative reconstructions by "kick-starting" the iterative reconstruction with a DL generated starting point. The proposed framework is applied to volumetric multi-axis spiral projection MRF that achieves whole-brain T1 and T2 mapping at 1-mm isotropic resolution for a 2-minute acquisition. First, the traditional reconstruction is optimized from over two hours to less than 40 minutes while using more than 90% less RAM and only 4.7 GB GPU memory, by using a memory-efficient GPU implementation. The Deli-CS framework is then implemented and evaluated against the above reconstruction. Results: Deli-CS achieves comparable reconstruction quality with 50% fewer iterations bringing the full reconstruction time to 20 minutes. Conclusion: Deli-CS reduces the reconstruction time of subspace reconstruction of volumetric spatio-temporal acquisitions by providing a warm start to the iterative reconstruction algorithm.

Time-encoded pseudo-continuous arterial spin labeling: Increasing SNR in ASL dynamic angiography.

PURPOSE: Dynamic angiography using arterial spin labeling (ASL) can provide detailed hemodynamic information. However, the long time-resolved readouts require small flip angles to preserve ASL signal for later timepoints, limiting SNR. By using time-encoded ASL to generate temporal information, the readout can be shortened. Here, the SNR improvements from using larger flip angles, made possible by the shorter readout, are quantitatively investigated. METHODS: The SNR of a conventional protocol with nine Look-Locker readouts and a 4 × $$ \times $$ 3 time-encoded protocol with three Look-Locker readouts (giving nine matched timepoints) were compared using simulations and in vivo data. Both protocols were compared using readouts with constant flip angles (CFAs) and variable flip angles (VFAs), where the VFA scheme was designed to produce a consistent ASL signal across readouts. Optimization of the background suppression to minimize physiological noise across readouts was also explored. RESULTS: The time-encoded protocol increased in vivo SNR by 103% and 96% when using CFAs or VFAs, respectively. Use of VFAs improved SNR compared with CFAs by 25% and 21% for the conventional and time-encoded protocols, respectively. The VFA scheme also removed signal discontinuities in the time-encoded data. Preliminary data suggest that optimizing the background suppression could improve in vivo SNR by a further 16%. CONCLUSIONS: Time encoding can be used to generate additional temporal information in ASL angiography. This enables the use of larger flip angles, which can double the SNR compared with a non-time-encoded protocol. The shortened time-encoded readout can also lead to improved background suppression, reducing physiological noise and further improving SNR.

Highly accelerated vessel-selective arterial spin labeling angiography using sparsity and smoothness constraints.

PURPOSE: To demonstrate that vessel selectivity in dynamic arterial spin labeling angiography can be achieved without any scan-time penalty or noticeable loss of image quality compared with conventional arterial spin labeling angiography. METHODS: Simulations on a numerical phantom were used to assess whether the increased sparsity of vessel-encoded angiograms compared with non-vessel-encoded angiograms alone can improve reconstruction results in a compressed-sensing framework. Further simulations were performed to study whether the difference in relative sparsity between nonselective and vessel-selective dynamic angiograms was sufficient to achieve similar image quality at matched scan times in the presence of noise. Finally, data were acquired from 5 healthy volunteers to validate the technique in vivo. All data, both simulated and in vivo, were sampled in 2D using a golden-angle radial trajectory and reconstructed by enforcing image domain sparsity and temporal smoothness on the angiograms in a parallel imaging and compressed-sensing framework. RESULTS: Relative sparsity was established as a primary factor governing the reconstruction fidelity. Using the proposed reconstruction scheme, differences between vessel-selective and nonselective angiography were negligible compared with the dominant factor of total scan time in both simulations and in vivo experiments at acceleration factors up to R = 34. The reconstruction quality was not heavily dependent on hand-tuning the parameters of the reconstruction. CONCLUSION: The increase in relative sparsity of vessel-selective angiograms compared with nonselective angiograms can be leveraged to achieve higher acceleration without loss of image quality, resulting in the acquisition of vessel-selective information at no scan-time cost.