Christopher S Parker1, Daniel C Alexander1, and Hui Zhang1
1Centre for Medical Image Computing, University College London, London, United Kingdom
Synopsis
Keywords: AI Diffusion Models, Quantitative Imaging, parallel imaging; apparent diffusion coefficient; IVIM; parameter estimation
Motivation: The distribution of reconstructed MRI signals, used as input for quantitative MRI with self-supervised deep learning, depends on the number of receiver coils. Current loss functions do not account for this, leading to bias.
Goal(s): Develop a non-central chi likelihood (NLC) loss that accounts for the distribution of MRI measures in the most common scenario of parallelised acquisitions.
Approach: Implement and evaluate the NLC loss and compare its performance against the MSE and Rician likelihood loss in simulated data.
Results: The NLC improves performance compared to the Rician likelihood and MSE loss for the mono-exponential ADC model in simulated data.
Impact: The
NLC loss permits fast inference of parameters from MRI signals reconstructed
from parallelised acquisitions and may reduce bias compared to the Rician and
MSE loss. The NLC loss is widely applicable due to the abundance of
parallelised MRI acquisitions.
Introduction
The
network training loss function has recently been shown to have a significant
impact on parameter estimation performance in quantitative MRI with self-supervised
deep learning [1]. This may be explained by its implicit assumptions on the
distribution of MRI signals. Mean squared error (MSE) assumes
Gaussian-distributed signal measures and is associated with biased parameter
estimates at low SNR [1,2]. To overcome this, we developed the Rician
likelihood loss that captures the distribution of magnitude signals
reconstructed from a single receiver coil [1]. However, more typically signal magnitudes
are reconstructed from multiple receiver coils which acquire data in parallel [3,4].
In this case, the distribution is more accurately described by a non-central
chi distribution [5]. In this work, we develop and introduce the negative log non-central
chi likelihood loss and assess its parameter estimation performance in synthetic
data.Methods
Theory
The sum-of-squares (SoS) reconstruction is commonly
used to compute the signal magnitude from multiple coils and is defined
as the square root of the sum-of-squares of the real and imaginary
components of complex signals [6]. For $$$N$$$ receiver coils, the
magnitude $$$M$$$ is computed as $$$M=\sqrt{\sum_{i=1}^{N} S_{R,i}^2 +
S_{I,i}^2}$$$, where $$$S_{R,i}$$$ and $$$S_{I,i}$$$ are
Gaussian-distributed with standard deviation $$$\sigma$$$.
For uncorrelated signals, the distribution of $$$M$$$ then
follows a non-central chi distribution with probability density function
$$$ p(M; v, \sigma, N) = \frac{m^N}{\sigma^2 v^{N-1}}
exp(\frac{-(M^2+v^2)}{2\sigma^2}) I_{N-1}(\frac{Mv}{\sigma^2})$$$ [5],
where $$$v$$$ is the noise-free signal we wish to estimate, which
depends both on the biophysical parameters and acquisition settings.
The maximum likelihood estimate (MLE) of $$$v$$$ is
known to be asymptotically unbiased. The network training loss function
was
therefore defined as the negative log of the non-central chi likelihood
(NLC) over the batch of $$$N$$$ voxels, each with $$$N_z$$$ signal
measures:
$$$ -\sum_{n=1}^{N} \sum_{z=1}^{N_z} log(p(M_{n,z};v_{n,z}, \sigma, N)
$$$.
Experimental evaluation
We
developed a PyTorch compatible implementation of the NLC loss. Performance was
evaluated against the MSE and NLR losses using the apparent diffusion coefficient
(ADC) and intra-voxel incoherent motion (IVIM) model as exemplar quantitative
MRI models.
For synthetic data experiments, complex noise-free MRI
data was generated from a uniform distribution of model parameters covering the
range of biophysically plausible values and i.i.d. Gaussian noise
($$$\sigma$$$=10 or 30) added. The S0 magnitude of each single coil was set equal 1.
An autoencoder with IVIM biophysical model
decoder was trained on 105 voxels [1].
Performance
was evaluated in terms of accuracy (bias), precision (standard deviation) and
total error (RMSE) on 105 unseen test voxels acquired both high SNR ($$$\sqrt{S_{R}^2(b=0)+
S_{I}^2(b=0)} = 30$$$) and low SNR ($$$\sqrt{S_{R}^2(b=0)+ S_{I}^2(b=0)} = 10$$$)
and for signals reconstructed with N=2 and 10 receiver coils.Results
For
both qMRI models, patterns of performance differences between losses were
similar at high and low SNR, with all showing improved performance (higher
accuracy and precision and lower total error) at high SNR in simulated data. We
therefore focus on performance at low SNR.
For
the ADC model, the NLC loss shows higher accuracy than NLR or MSE losses for
both ADC and S0 parameters on simulated data reconstructed from N=2 and N=10
receiver coils, with higher precision for S0 (the signal magnitude in each
receiver coil). Higher precision for ADC was observed at N=10 for all losses compared
to N=2. Lower accuracy for S0 was observed at N=10 compared to N=2 for MSE and
NLR losses, whereas the NLC accurately estimated S0 consistently.
For
the IVIM model, Dp shows high accuracy at N=10 and over-estimation at N=2 for all
losses. For both N=2 and N=10, Dt is underestimated with MSE loss, whereas accuracy
is high for the NLC and NLR losses. Fp estimation is relatively accurate and precise
for all losses at both N=2 and N=10.Discussion
For simple mono-exponential signal decay, NLC
loss’ higher accuracy is explained by its more accurate modelling of the
distribution
of reconstructed MRI signals. As the number of coils used in the SoS
reconstruction
increases, the distribution of signals becomes increasingly shifted
towards
higher values relative to the noise-free signal (see preview picture
where $$$v$$$=1). As N increases, the MSE and NLR losses will predict
noise-free signals that
are closer to the centre of the measured signals, causing
under-estimation of
ADC. Higher precision of estimates reported at high N is explained by
the
relatively higher number of signal averages in the reconstructed signal.Conclusion
The
NLC loss is designed for training self-supervised deep networks for quantitative
MRI on parallelised acquisitions. Results in simulated data show improved
parameter estimation for ADC estimation and improved S0 estimation for IVIM.Acknowledgements
CSP, DCA and HZ are supported by the Medical ResearchCouncil (MR/T046473/1).References
[1] Parker, C.S., Schroder,
A., Epstein, S.C., Cole, J., Alexander, D.C. and Zhang, H., 2023. Rician
likelihood loss for quantitative MRI using self-supervised deep learning. arXiv
preprint arXiv:2307.07072.
[2]
Barbieri, S., Gurney‐Champion, O.J., Klaassen, R. and Thoeny, H.C.,
2020. Deep learning how to fit an intravoxel incoherent motion model to
diffusion‐weighted MRI. Magnetic resonance in medicine,
83(1), pp.312-321.
[3]
Griswold, M.A., Jakob, P.M., Heidemann, R.M., Nittka, M., Jellus, V., Wang, J.,
Kiefer, B. and Haase, A., 2002. Generalized autocalibrating partially parallel
acquisitions (GRAPPA). Magnetic Resonance in Medicine: An Official Journal
of the International Society for Magnetic Resonance in Medicine, 47(6),
pp.1202-1210.
[4]
Pruessmann, K.P., Weiger, M., Scheidegger, M.B. and Boesiger, P., 1999. SENSE:
sensitivity encoding for fast MRI. Magnetic Resonance in Medicine: An
Official Journal of the International Society for Magnetic Resonance in
Medicine, 42(5), pp.952-962.
[5]
Constantinides, C.D., Atalar, E. and McVeigh, E.R., 1997. Signal‐to‐noise
measurements in magnitude images from NMR phased arrays. Magnetic resonance
in medicine, 38(5), pp.852-857.
[6]
Sotiropoulos, S.N., Moeller, S., Jbabdi, S., Xu, J., Andersson, J.L., Auerbach,
E.J., Yacoub, E., Feinberg, D., Setsompop, K., Wald, L.L. and Behrens, T.E.J.,
2013. Effects of image reconstruction on fiber orientation mapping from
multichannel diffusion MRI: reducing the noise floor using SENSE. Magnetic
resonance in medicine, 70(6), pp.1682-1689.
[7]
Epstein, S.C., Bray, T.J., Hall-Craggs, M. and Zhang, H., 2022. Choice of
training label matters: how to best use deep learning for quantitative MRI
parameter estimation. arXiv preprint arXiv:2205.05587