Magnetic Resonance Fingerprinting (MRF) excites the imaging object with an inversion recovery (IR) RF pulse followed by a sequence of RF pulses with randomized flip angles (FA) and repetition times (TR) [1]. A single image is reconstructed from data collected from each short TR due to the non-repeatable magnetization history. The highly-undersampled single-shot imaging leads to high levels of noise and artifacts, and requires increased temporal frames in order to achieve accurate fingerprint matching. Saturation recovery (SR) is a commonly used preparation scheme in conventional MR to reset the magnetization to zero, which can be utilized to achieve a “multi-shot” MRF acquisition without waiting for magnetization recovery. However, the SR, instead of IR, may decrease the dynamic range of the magnetization in MRF. In this work, we show that an SR prepared multi-shot MRF can achieve similar or even better accuracy than the original IR prepared MRF with the same amount of data collected.

SR prepared MRF Sequence:

An SR module including a 90˚ RF pulse with spoiling gradients applied before and after the 90˚ pulse was used to reset the magnetization spins to zero before each MRF repetition (Figure 1). No relaxation waiting time is needed between each repetition. A different part of k-space can then be sampled with each repetition, making this a multi-shot acquisition. To compensate for the decreased dynamic range due to the saturation recovery, an SR and IR (SR+IR) hybrid MRF sequence was also proposed. In the SR+IR case, an SR module was applied at the beginning of each repetition, and one or more IR pulses were used during each repetition (Figure 1).

Experiment design:

For each of the IR, SR and SR+IR MRF sequences, 1000 TRs were simulated using a balanced SSFP acquisition. Pseudorandom FAs with the same pattern and amplitude were used for all the sequences. A single spiral was acquired within each TR (~14ms). Repetitions of 2 and 4 were simulated for SR based MRF, resulting in 2 spirals/image x 500 images and 4 spirals/image x 250 images, respectively (Table 1). Multiple spirals were distributed uniformly within each image and were rotated by 2pi/48 for succeeding images. Ground truth T1, T2 and proton density brain maps were obtained from BrainWeb [2], zero padded to matrix size 256x256. Off-resonance values ranged linearly from -60Hz to 60Hz.

Reconstruction:

An accelerated iterative MRF (AIR-MRF) reconstruction method [3], integrated with dictionary compression and fast searching were developed and used to reconstruct the MRF data. Briefly, the reconstruction compressed the images and the fingerprints in the dictionary along the temporal direction, and iteratively updated estimates in the compressed image space and the tissue parameter space. An approximate nearest neighbor search [4] was used instead of an exhaustive search to speed up the reconstruction. The dictionary contained around 51k atoms, covering normal brain tissue T1, T2 and off-resonance ranges. Dictionary compression through SVD [5] was used to compress the fingerprint length (see Table 1).The number of iterations was fixed at 5 with sufficient convergence typically observed.

The reconstructed parameter maps were compared to the ground truth and normalized mean square error (NMSE) was utilized to quantify the results.

Figure 2 shows example T1, T2 maps from several sequences, along with ground truth and corresponding error maps. Zoomed-in T1 and T2 maps from all tested sequences are shown in Figure 3. Less noise can be observed on the SR+IR results in the grey and white matter regions, compared to the IR and SR. The SR+IR presented similar quality on the dura mater region as the IR, better than the SR. Figure 4 shows the NMSE analysis and the reconstruction time of all the tested sequences. The SR+IR at different repetitions achieved better performances than IR.In this study, an SR preparation module was introduced to MRF, enabling multi-shot MRF without a waiting time for magnetization recovery. Thanks to the non-steady-state property of MRF, IR pulses were combined with SR to compensate (to some extent) for the shortage of the decreased magnetization range and the hybrid SR+IR sequence achieved better parameter estimation than using SR alone. The multi-shot SR+IR MRF achieved similar or even better parameter estimation than the single-shot IR MRF. With the multi-shot scheme, MRF can be readily extended to Cartesian trajectory sampling, 3D imaging, etc., using conventional MRI techniques. The SR preparation may open the door for MRF to more applications and easy implementation. Future study will include implementing the sequence and validating on real acquired data.