Qiang Zhang1, Pan Su2, Ying Liao3, Rui Guo1, Haikun Qi4, Zhangxuan Hu1, Hanzhang Lu2, and Huijun Chen1
1Tsinghua University, Beijing, China, 2Johns Hopkins University, Baltimore, MD, United States, 3New York University, New York, NY, United States, 4School of Biomedical Engineering and Imaging Sciences, King's College London, London, London, United Kingdom
Synopsis
The purpose of this study is to develop a MRF-ASL reconstruction
algorithm using deep learning (DeepMARS). Compared with the traditional dictionary
matching, our DeepMARS achieved higher intra-class correlation (ICC) in B1
(0.971 vs 0.921) and BAT (0.926 vs 0.761), similar ICC in T1 (0.957 vs 0.964)
and CBF (0.936 vs 0.948) in the reproducibility test with much shorter
calculation time per voxel (0.368 ms vs 2.899 s), suggesting that our DeepMARS
may be a better alternative than the conventional MR dictionary matching
approach.
Introduction:
Recently, an MR
fingerprinting (MRF) based ASL technique (MRF-ASL) [1] was developed to take advantage of, rather than
being limited by, the multiparametric nature of ASL. However, the reconstruction
of MRF-ASL is
time consuming, and the reproducibility needs to be improved.
In this work, we demonstrate that deep learning based reconstruction method can achieve high reproducibility with short reconstruction
time in
MRF-ASL.Experiments:
Eight healthy subjects (age 26
3.7
years, 4 males and 4 females) were imaged using MRF-ASL
sequence [1], and
each subject was scanned three times to test the reproducibility, and three patients
with Moyamoya disease (age 39
2.1years,
1 male) were included. The sequence parameters were: multi-slice EPI, matrix
size = 64×64, voxel size = 2.8mm×2.8mmx10mm, 500 dynamics, scan duration=3.7min.Data analysis:
The
MRF-ASL signal was considered under a single-compartment model [1] and the goal is to estimate four
quantitative parameters simultaneously: B1, T1,
cerebral blood flow (CBF), and bolus arrive time (BAT). The primarily focus of
this study is to compare a novel deep learning (DL) based parametric estimation
method to a traditional dictionary matching (DM) [2]:
DM
method: First, a
MRF dictionary was simulated with 6149000 combinations of B1 values in the
range of 0.6~1.1 (in an increment of 0.02), T1 values in the range of 500~5000
ms (in an increment of 20 ms within 2000 ms, and 300 ms beyond 2000 ms), CBF
values in the range of 3~150 ml/100g/min (in an increment of 3 ml/100g/min),
and BAT values in the range of 300~3000 ms (in an increment of 50 ms). Then, for
a given MRF signal time course, its cross-correlations (cc) with all entries in
the dictionary were calculated and the highest cc was identified.
DL
method: The architecture of our DeepMARS was
illustrated in Fig. 1, and five models were trained: Model 1 used dictionary without
noise and mean square error (MSE) as the loss function, Model 2 used the
dictionary contaminated with Gaussian noise and MSE as loss function, Model 3 used
the dictionary with Gaussian noise and mean absolute percentage error (MAPE) as
loss function, Model 4 used Model 1 as the pre-trained model and was further
trained by the dictionary with Gaussian noise and used MSE as loss function.
Model 5 used Model 1 as the pre-trained model and was further trained by the
dictionary with Gaussian noise and used MAPE as loss function. The signal to noise
ratio of the Gaussian noise was randomly selected from 20 dB to 60 dB. The MSE
was defined as: (∥TruePara−PredPara∥2)2, and the MAPE was defined as:∑p∈Para|Truep−Predp|/Truep,
where Para was [B1,T2,CBF,BAT]. The voxel-wise intra-class correlation (ICC) [3] was calculated for each slice to test
the reproducibility. The calculation time per voxel of DM and DeepMARS was
compared on a 3.2 GHz Inter Core i5 with 26 GB RAM.
Results and Discussion:
The
ICCs (mean±standard) of four quantitative parameters for all
methods were shown in Table 1. Compared with the DM method, the Model 1 showed
lower ICCs in all four parameters, while Model 2/3/4/5 all showed higher ICCs
in B1 and BAT, similar ICC in T1 and CBF, suggesting that the proposed DeepMARS
outperformed the traditional DM method, and adding noise in the training promoted
robustness of learning. It can also be observed that using pre-trained model
improved the reproducibility of B1 and BAT (ICC: Model 4 > Model 2, Model 5 >
Model 3), and the MAPE showed similar ICC of B1/T1/CBF and lower ICC of BAT
compared with MSE (Model 3 vs Model 2, Model 5 vs Model 4). Notably, the
calculation time of DM was more than 7500 times than that of DeepMARS in the
same computer. Fig. 2 illustrated the results for all methods from one patient,
who had bilateral MCA and right PCA stenosis. Accordingly, the prolonged BAT and
lower CBF can be observed at both DM and DeepMARS methods. As shown in CBF
mapping (Fig. 2), the MAPE loss function (Model 3 & 5) made it less noise
in the delineated lesion than the MSE loss function (Model 2 & 4). Therefore,
the Model 5 might be more suitable than other models for MRF-ASL
reconstruction. Fig. 3 and Fig. 4 showed the comparison of DM and DeepMARS (using
model 5) in one healthy subject and two patients. It can be seen that DeepMARS
derived parametric maps showed better quality compared to DM method. Conclusion:
We have successfully demonstrated that the DeepMARS can provide MRF-ASL
higher reproducibility in B1 and BAT, and similar reproducibility in T1 and CBF
with much shorter calculation time than traditional dictionary matching method.Acknowledgements
No acknowledgement found.References
1. Pan, S., et al., Multiparametric estimation of brain hemodynamics with MR fingerprinting
ASL. Magnetic Resonance in Medicine, 2017. 78(5): p. 1812-1823.
2. Ma, D., et al., Magnetic resonance fingerprinting.
Nature, 2013. 495(7440): p. 187-192.
3. McGraw, K. and S.P.
Wong, Forming Inferences About Some
Intraclass Correlation Coefficients. Vol. 1. 1996. 30-46.