Image acquisition and reconstruction

All exams were acquired on two 3T clinical MR scanners (Discovery MR750 and Discovery MR750W, GE Healthcare, Waukesha, WI) using an eight channel receive-only RF head coil. MRF data acquisition was performed using a three-dimensional steady state free precession sequence with a novel multi‐axis spiral trajectory and slab excitation [9]. Adiabatic inversion pulses were used before each acquisition. A ramp flip angle schedule was used ranging from 0.778° to 70°. The sequence details can be found in [9]. A volumetric k-space data set consisting of 256 × 256 × 256 samples and a FOV of 25.6 × 25.6 × 25.6 cm³ resulted in reconstructed relaxometry maps with 1 mm isotropic resolution. The total acquisition time for the whole brain was 4:38 (minutes: seconds). The T₁ for the dictionary ranged from 10 to 3000 ms and T₂ from 10 ms up to 2000 ms. Each T₁ and T₂ dictionary step range follow the exponential nature of T₁ and T₂ relaxation curves. For example, 10:5:100 means step size 5 range from 10 to 100 ms. As such, a few examples for T₁ curve steps are 10:5:100, 110:10:1000, 1050:50:2000, 2100:100:3000 and for T₂ = 10:1:100, 105:5:500, 525:25:1000, 1100:100:2000. Fingerprint reconstruction and dictionary matching were performed offline in Matlab (Mathworks, Natick, Massachusetts) on a Linux workstation equipped with two 8‐core Intel Xeon Gold 6244 central processing unit @ 3.60 GHz, 376 GB system memory, and a NVIDIA Tesla V100 graphical processing unit. The reconstruction pipeline has been described elsewhere [35]. Briefly, the undersampled data were anti-aliased with a k-space-weighted view-sharing algorithm and [36, 37] then the view-shared data were compressed with singular value decomposition algorithm and finally the first 15 singular value decompensation coefficients of the temporal signals were kept for parametric maps reconstruction. GPU gridding and fast Fourier transform were performed on the compressed non-Cartesian k-space data. T₁, T₂ and proton density maps were computed via dictionary matching on the general processing unit and interpolated to 512 × 512 × 512 image matrix for display. The computation time was approximately 10 min for each dataset.