Alexandra Grace Roberts1,2, Yi Wang1,2, Pascal Spincemaille2, and Thanh Nguyen2
1Electrical Engineering, Cornell University, Ithaca, NY, United States, 2Radiology, Weill Cornell Medicine, New York, NY, United States
Synopsis
Keywords: Machine Learning/Artificial Intelligence, Brain
3D super-resolution of QSM is feasible using the VDSR and U-net
architectures using a fraction of the required number of epochs as previous 2D super-resolution
networks. Additionally, this method both reduces whole brain MSE and ROI MSE and
increases the apparent resolution as compared to the interpolated input
Introduction
Quantitative Susceptibility Mapping (QSM) is a contrast in magnetic
resonance imaging (MRI) that quantifies the induced magnetization of substances
in the brain such as calcium deposits or iron found in blood products [1]. The
sensitivity of this contrast allows the detection of microbleeds linked to
COVID-19, multiple sclerosis, and vascular dementia [2]. Additionally, this
quantification allows longitudinal monitoring in patients with multiple
sclerosis, Parkinson’s disease [3] or Alzheimer’s disease [4]. Current clinical
protocols that aim for 5-minute acquisitions are compromising spatial
resolution to allow sufficient sampling of echo times. Given the compromise
between resolution and acquisition time[5], [6], the use of neural networks to learn
mappings between low and high-resolution images (known as super-resolution) is
of interest.
Generative
adversarial networks (GANs) have been used to obtain super-resolved T1w images
[7] and convolutional neural networks (CNNs) increase the apparent resolution
of double-echo steady state (DESS) images [8] and susceptibility
weighted-images (SWI) [9]. Since different contrasts in MRI visualize the same
anatomical structures differently, the mappings are contrast-specific, there is benefit in examining the mapping between low and high-resolution QSM
datasets. Here, the feasibility of 3D QSM super-resolution (SuperQ) is
demonstrated.Theory
The
network architecture used here is a Very Deep Super Resolution
(VDSR) neural network [14], the goal of which is to learn the mapping from a
downsampled high-resolution image (i.e. a low-resolution image) and its
residual. The residual is defined as the difference in the high-resolution
image and the cubic interpolation of the downsampled image. The residual
consists of features that cubic interpolation fails to represent, improving the
image sharpness and overall quality. As mentioned in the original work for 2D
natural images, learning the residual image rather than the high-resolution
image requires less memory and allows faster convergence of stochastic gradient
descent. The VDSR convergence is also compared to a U-net trained on a similar
number of parameters. Training
the network amounts to minimizing the mean-squared error (MSE) loss function $$E_R (θ)= \frac{1}{2} ((h_θ(x)-y)^2+\lambda(w^Tw))$$
Where
$$$\theta$$$ represents network parameters, $$$h_{\theta}(x)$$$ is the learned residual between the cubic
interpolation of the downsampled input $$$x$$$ and the high-resolution image $$$x_0$$$, and $$$y$$$ is the ground truth residual; $$$\lambda$$$ is an L2 regularization parameter that
prevents outsized influence of a particular filter kernel in weight vector $$$w$$$.
The
VDSR network architecture is comprised of alternating convolution and Rectified
Linear Unit (ReLU) layers with a final regression layer.Method
The network depth
totals 41 layers, 20 convolution layers with filter size [3 3 3], stride of [1
1 1], and zero-padding of [1 1 1] with
20 ReLU layers. The network was initialized according to He’s method [15] and
trained with a stochastic gradient descent optimizer with a momentum of 0.9 and
a gradient threshold 0.1 for 100 epochs with an epoch interval 1. The
batch size was 64, training was conducted with a piecewise learning schedule
with initial learning rate 0.01, learning rate factor 0.1, learn
rate drop period of 10 epochs and an L2 regularization factor 0.00001.
Nine
cases featuring 3T QSMs (of matrix size 512 x 512 x 152 and voxel size 0.43 x
0.43 x 1 mm with echo time (TE) = 20 ms) were acquired and split 7:1:1 into
training, testing, and validation subsets. Training and validation sets
were divided into 16 x 16 x 32 patches, which was found to be the optimal size
via grid search. The training volume was downsampled using k-space cropping
by a factor of 2, 3, and 4. The downsampled volume was cubically interpolated
(upsampled) by the same factor. The residual was calculated by subtracting the
cubic interpolation from the initial high-resolution. The MSE and apparent resolution [16] between the ground truth and super-resolved image (the
sum of the cubic interpolation and residual) was calculated on the brain
volume and regions of interest (ROIs) - the red nucleus, substantia
nigra, caudate nucleus, globus pallidus, dentate nucleus, and putamen.
A
U-net with a similar number of parameters was trained for 50 epochs for
comparison. A test cast outside the (healthy subjects) training distribution
consisting of a patient with multiple sclerosis (MS) was evaluated.Results
SuperQ
demonstrates lower MSE in the whole brain volume and ROIs as shown in
Table 1. The improvement in the visual quality of the test case is shown
(Figures 1, 2). A U-net architecture with a smaller set of parameters outperformed
the VDSR architecture after 50 epochs of training (Figure 3). The improved performance by
the U-net may offer insight into the quantitative super-resolution problem. Additionally,
SuperQ is trained on 3D data and requires less than a 10th of the
number of epochs in the previously proposed 2D method and generalizes to a case outside the training distribution (Figure 4).Discussion
3D super-resolution of QSM is feasible using the VDSR and U-net
architectures using a fraction of the required number of epochs as 2D super-resolution
networks. Additionally, this method reduces whole brain and ROI MSE and
increases the apparent resolution compared to the interpolated input. Learning
the residual rather than the high-resolution image grants
generalization to the mapping, as is seen with the MS case. Future work includes k-space
domain training and prospective super-resolution.Acknowledgements
No acknowledgement found.References
-
Wang, Y. and T. Liu (2015).
"Quantitative susceptibility mapping (QSM): Decoding MRI data for a tissue
magnetic biomarker." Magnetic Resonance in Medicine 73(1): 82-101.
- M. Ayaz, A. S. Boikov, E. M.
Haacke, D. K. Kido, and W. M. Kirsch, “Imaging cerebral microbleeds using
susceptibility weighted imaging: one step toward detecting vascular dementia:
Imaging Microbleeds Using SWI,” J. Magn. Reson. Imaging, vol. 31, no. 1, pp.
142–148, 2010.
- Du, G., et al. (2016). "Quantitative
susceptibility mapping of the midbrain in Parkinson's disease." Movement
Disorders 31(3): 317-324.
- Kim, H.G., et al.
(2017). "Quantitative susceptibility mapping to evaluate the early stage
of Alzheimer's disease." NeuroImage. Clinical 16: 429-438.
- M. Kozak, C. Jaimes, J.
Kirsch, and M. S. Gee, “MRI techniques to decrease imaging times in children,”
Radiographics, vol. 40, no. 2, pp. 485–502, 2020.
- J. Yeung, “Spatial resolution (MRI),”
Radiopaedia.org. [Online]. Available:
https://radiopaedia.org/articles/spatial-resolution-mri-2?lang=us. [Accessed:
28-Sep-2021].
- Y. Chen, A. G. Christodoulou, Z. Zhou, F. Shi,
Y. Xie, and D. Li, “MRI super-resolution with GAN and 3D multi-level DenseNet:
Smaller, faster, and better,” arXiv [cs.CV], 2020.
- A. S. Chaudhari et al., “Super‐resolution
musculoskeletal MRI using deep learning,” Magn. Reson. Med., vol. 80, no. 5,
pp. 2139–2154, 2018.
- A.G. Roberts, P. Spincemaille, I. Kovanlikaya,
John T., T. Nguyen, and Y. Wang. SWISeR: Multi-field susceptibility-weighted
images super-resolution. In Proceedings of International Society for Magnetic
Resonance in Medicine, 2022
- K. Zeng, H. Zheng, C. Cai, Y. Yang, K. Zhang,
and Z. Chen, “Simultaneous single- and multi-contrast super-resolution for
brain MRI images based on a convolutional neural network,” Comput. Biol. Med.,
vol. 99, pp. 133–141, 2018.
- S. Roy, A. Jog, E. Magrath, J. A. Butman, and
D. L. Pham, “Cerebral microbleed segmentation from susceptibility weighted
images,” in Medical Imaging 2015: Image Processing, 2015.
- B. Bachrata, S. Bollmann, G. Grabner, S.
Trattnig, and S. Robinson Isotropic QSM in seconds using super-resolution 2D
EPI Imaging in 3 Orthogonal Planes. In Proceedings of International Society for
Magnetic Resonance in Medicine, 2022.
- A. Moevus, M. Dehaes, B. De Leener. Improving
Deep Learning MRI Super-Resolution for Quantitative Susceptibility Mapping.
International Society for Magnetic Resonance Imaging, Magnetic Resonance in
Medicine, 2021.
- J. Kim, J. K. Lee, and K. M. Lee, “Accurate
image super-resolution using very deep convolutional networks,” arXiv [cs.CV],
2015.
- He, K., et al. Delving Deep into Rectifiers:
Surpassing Human-Level Performance on ImageNet Classification, IEEE.
- Eskreis-Winkler S, Zhou D, Liu T, Gupta A, Gauthier SA, Wang Y,
Spincemaille P. On the influence of zero-padding on the nonlinear operations in
Quantitative Susceptibility Mapping. Magn Reson Imaging. 2017 Jan;35:154-159.
doi: 10.1016/j.mri.2016.08.020. Epub 2016 Aug 29. PMID: 27587225; PMCID:
PMC5160043.