SCMR/ISMRM Co-Provided Workshop
SCMR 22nd Annual Scientific Sessions
Background: Cardiac MRF (cMRF) jointly maps T1, T2, and M0 by matching time-varying signals to a dictionary of simulated timecourses . However, a new dictionary must be created after each scan to model the subject’s heart rhythm, which is challenging for clinical translation and online reconstruction . Inspired by recent work [3–5], this study uses a neural network to rapidly output dictionaries for arbitrary heart rhythms.
Methods: A shallow neural network is proposed that takes as inputs T1, T2, and the RR intervals. The network has two fully-connected layers with 300 nodes and outputs a cMRF signal evolution. The network was trained with signals simulated using the Bloch equations with different T1, T2, and cardiac rhythms, including corrections for slice profile and inversion efficiency . There are 500 simulated cardiac rhythms with different average heart rates (40-120bpm) and amounts of Gaussian noise added to each RR interval (standard deviations 0-50% of the mean RR). The training set has a total of 2 million signals, since 4000 T1 (50-3000ms) and T2 (5-600ms) combinations are modeled for each cardiac rhythm. The T2 array of the ISMRM/NIST system phantom was scanned at 3T (Siemens Skyra) . Several artificial heart rates were simulated on the scanner: (1) constant 40-120bpm (step size 10bpm) (2) constant 60bpm that switched to 80bpm after 8 heartbeats, and (3) constant 80bpm that switched to 60bpm. Dictionaries were generated both with the Bloch equations and the neural network with the same T1 and T2 discretization. Parameter maps were reconstructed by dot product matching . A similar analysis was performed for in vivo scans from 6 volunteers with ROIs in the septal myocardial wall.
Results: After training, the time to generate a dictionary with 26,680 entries was 0.8s (neural network) vs 158s (Bloch equations using parallelized Mex code). In the NIST phantom, the network estimates agree with reference values—for cases with constant heart rates, all R2>0.999 for T1 and 0.996 for T2; for abrupt changes in heart rate, all R2>0.999 for T1 and 0.998 for T2. Note that these cardiac rhythms were not present in the training data. In vivo maps are presented in Figure 3A. Figures 3B-3C show a Bland-Altman analysis comparing myocardial T1 and T2 using a dictionary simulated from the Bloch equations vs the neural network. The Bland-Altman statistics are: T1 bias 3.9ms, 95% limits of agreement (-7.5,15.3)ms; T2 bias 0.0ms, 95% limits of agreement (-0.9,0.9)ms.
Conclusion: A neural network can rapidly generate cMRF dictionaries for arbitrary heart rates, eliminating the need to simulate a new dictionary for every scan. This may aid clinical translation by reducing computation times. Corrections for effects like slice profile, which are computationally intensive, may be included upfront in the training. Rapid dictionary generation may enable efficient comparison of cMRF sequences optimized for studying different pathologies.