hide
Free keywords:
-
Abstract:
Introduction: Previously we proposed a supervised learning
approach to automatically generate MR sequences from scratch
without providing sequence programming rules, called MRzero [1]. In
the present work, we develop an auto-encoder for T1 by performing a
joint optimization of sequence parameters and a neural network using
MRzero.
Subjects/methods: The fully differentiable MRI pipeline is simulated
end-to-end with Bloch parameters as input and T1 as target. We
utilize known operator learning [2] in the reconstruction to reduce the
number of trainable parameters in our NN by keeping the adjoint
formalism [1] as known operator in the image reconstruction. The T1
training dataset consist of ten T1 maps with matrix size 32 9 32. For
each target sample a non-zero PD rectangle with matrix size 16 9 16
at varying spatial location with voxel-wise randomly assigned PD, T1,
T2 and B0 is defined, resulting in a total training data size of 2560
samples. A three-hidden-layer multilayer perceptron is used for T1
quantification.
The MR sequence is based on a 180 deg inversion prepared 2D
FLASH sequence with matrix size 32 9 32, TR = 15 ms, TE = 8 ms,
FA = 5 deg, repeated 6 times with varying TI and Trec. Together
with the NN parameters, all TI and Trec times are optimized to find
the best sequence for T1 mapping and are initialized with 0. Additionally,
a penalty for longer times was applied to enforce shorter
sequences. The optimization process (Fig. 1) interleaves the sequence and NN optimization after 50 and 5000 iterations, respectively. In
total 500 iterations of sequence optimization are performed.
Simultaneously optimized sequence parameters and trained NN
are applied on a higher resolution with matrix size 126 9 126 and
parallel imaging (GRAPPA acceleration factor 3) for in vivo measurements
at 3T. Results/discussion: The T1 map of a healthy subject generated by the
final optimized sequence is displayed in Fig. 2. Figure 3 shows the
different stages of sequence optimization. The acquired T1 values of
CSF, white matter and grey matter for later iterations match well to
literature values at 3T [3]. A standard inversion recovery sequence
was used as reference. The obtained maps match well with the reference,
but the acquisition time could be reduced from 63.3 s to
19.2 s. Optimized TI and Trec times range from 0.5 s to 1.8 s and
0.5 s to 1.1 s, respectively.
The simultaneous sequence optimization and NN training was performed
solely on synthetic data at low resolution, but inference on
higher resolution on in vivo data provided high quality T1 maps.
Preliminary results at low resolution were shown in [1]. The T1 autoencoder
is a proof-of-concept that can be extended also to multiparametric
mapping—similar to MR fingerprinting—yielding PD, T1,
and T2, as well as B1 and B0 inhomogeneity maps.
