Yuan Zheng1, Tao Feng1, Sirui Li2, Wenbo Sun2, Qing Wei3, Samo Lasic4, Danielle van Westen5, Karin Karin Bryskhe4, Daniel Topgaard4,5, and Haibo Xu2
1UIH America, Houston, TX, United States, 2Zhongnan Hospital of Wuhan University, Wuhan, China, 3United Imaging Healthcare, Shanghai, China, 4Random Walk Imaging, Lund, Sweden, 5Lund University, Lund, Sweden
Synopsis
Multidimensional diffusion MRI (dMRI) is a powerful tool that even in its simplest form provides
more detailed microstructural information than conventional dMRI, such as
microscopic anisotropy (µFA) unconfounded by orientation dispersion. However,
it requires multiple diffusion encoding modes (usually directional and isotropic
encodings) and, for the more advanced versions, prolonged scan and post-processing
times. We proposed using convolutional neural networks (CNN) to accelerate multidimensional dMRI data
acquisition and analysis, and have demonstrated that satisfactory µFA maps can
be generated in real-time with only 50% of the encodings, which might help to better
adapt multidimensional dMRI to clinical practices.
Introduction
Multidimensional diffusion MRI (dMRI)1 is a novel imaging modality
that provides more detailed microstructural information than conventional dMRI,
for instance independent quantification of microscopic fractional anisotropy (µFA)
and orientation dispersion2 which are two fundamentally different
properties that are entangled in conventional methods.
However, in addition to conventional directional encoding,
multidimensional dMRI requires more advanced diffusion encoding modes such as
isotropic encoding2, requiring somewhat longer scan times. While simple estimation of µFA with
the generalized cumulant inversion3 requires no more time than for conventional
DTI analysis, more advanced post-processing relying on Monte Carlo inversion currently
is too time-consuming for real-time diagnosis.4 Both the long
acquisition and analysis time may limit the clinical use of the more advanced
forms of multidimensional dMRI.
Great potential has been demonstrated using convolutional neural networks
(CNN) with deep learning in many aspects of MRI. In this study, we trained a
CNN model using only half of the total diffusion encodings to predict µFA maps.
High quality µFA maps were generated in real time, demonstrating the possibility
of using CNN to significantly accelerate multidimensional dMRI data acquisition
and post-processing.Methods
Eighty-five patients with
glioma, meningioma and some other brain diseases were included in this study.
MRI scans were performed on a uMR 790 3.0 T scanner (United Imaging
Healthcare, Shanghai, China) with a 24-channel head coil. The multidimensional
dMRI sequence has 80 diffusion gradient encodings total, including 40 directional
and 40 isotropic (Fig.1). For both encoding schemes, images were acquired at a
b=100, 700, 1400, 2000 s/mm2, with 6, 6, 12, 16 directions or averages.
Imaging parameters were identical for both kinds of encoding as detailed in
Fig.1 caption. The number of slices were chosen to cover the whole brain, and
the total scan time was approximately 5 min. Images were analyzed using the
procedure described by Lasič to extract µFA2,
assuming a Gamma distribution of the diffusivity.
Slices below the nasal cavity were excluded to avoid EPI artifact caused
by the complex air-tissue boundaries, and a total of 1444 slices were included
in CNN training (1155) and test (289). We retrospectively undersampled both directional
and isotropic encodings at each b value by 50% while aiming to maintain a relatively
uniform direction distribution.
The 40 diffusion encodings left were used as the network input (112×112×40)
and the µFA maps generated from all the 80 encodings with the Gamma model were
used as the output (112×112) for model training. Structure of the CNN is shown
in Fig.2. Eight convolutional layers were included. The number of channels for
the layers were 80,120,180,180,120,60,30 and 1. A 3×3 filter was used
for all layers. Leaky rectified linear units (ReLU) were applied for all layers
except the last one, which was followed by conventional ReLU activation. The
normalized squared error was used as the loss function and Adam5 was used
as the optimizer. The CNN was implemented using Tensorflow.
Data augmentation including horizontal/vertical flips and rotation were
applied, and each batch contained 9 slices. The number of epochs for model
training was determined using 10-fold cross-validation of the training set. The
model was then trained on the full training dataset and evaluated on the test
dataset.Results
The training and validation loss during cross validation were shown in
Fig.3. Both losses became flat after 200 iterations with no major difference.
The epoch # for final model training was thus chosen to be 200.
The model performance was evaluated on the test dataset, and the normalized
squared error was ~2.2%.
Fig.4a shows the FA map of a meningioma case calculated using directionally
encoded images with the DTI model. Low FA value is ambiguous since it could be
caused either by orientation dispersion or lack of microanisotropy. On the
other hand, the basic implementation of multidimensional dMRI showed elevated µFA
of the tumor (Fig.4b,c), revealing strong subvoxel microanisotropy (Fig.4d).
The µFA maps reconstructed with the conventional method and the proposed CNN
approach showed similar results. The former took 80 diffusion encodings (5 min scan time) and 8 min for data analysis on a intel 8-core 3.4 GHz CPU,
while the latter took only 40 encodings (2.5 min scan time) and under 2 sec to
generate the µFA map with a Nvidia GTX960 GPU.Discussion
We acquired 80 diffusion encodings in this study because of clinical
scan time concerns, while some other multidimensional dMRI studies used larger
numbers of diffusion encodings4. Therefore our µFA maps calculated
using the conventional method were slightly noisy. However, they were still
used for model training and evaluation. In this case the normalized loss does
not accurately reflect the model performance, since denoising effect of the CNN
makes the prediction deviate from the “gold standard” but is desired. Conclusion
Multidimensional dMRI is a powerful tool that even in its simplest
incarnation provides more detailed microstructural information than
conventional dMRI. However, the more advanced versions require prolonged scan
and post-processing time. We proposed using CNN to accelerate multidimensional
dMRI data acquisition and analysis, and have demonstrated that satisfactory µFA
maps can be generated in real-time with only 50% of the encodings, which might help
to better adapt this technique to clinical settings.Acknowledgements
Data collection was approved by the Wuhan
hospital ethics committee and with written informed consent from all
participants. This work was financially supported by the national key research and development plan of China (2017YFC0108803), the Swedish Foundation for
Strategic Research (AM13-0090, ITM17-0267) and the Swedish Research Council
(2018-03697). Daniel Topgaard owns shares in Random Walk Imaging AB (Lund,
Sweden, http://www.rwi.se/), holding patents related to the described methods.References
- D.
Topgaard. Multidimensional diffusion MRI. J Magn. Reson. 275:98-113 (2017).
- S.
Lasič et.al., Microanisotropy imaging: quantification of microscopic diffusion
anisotropy and orientational order parameter by diffusion MRI with magic-angle
spinning of the q-vector. Front. Physics. Feb 27 (2014).
- Sjölund J, Szczepankiewicz
F, Nilsson M, Topgaard D, Westin C-F, Knutsson H. Constrained optimization of
gradient waveforms for generalized diffusion encoding. J Magn Reson
2015;261:157-168.
- Westin C-F, Knutsson H,
Pasternak O, Szczepankiewicz F, Özarslan E, van Westen D, Mattisson C, Bogren
M, O'Donnell L, Kubicki M, Topgaard D, Nilsson M. Q-space trajectory imaging
for multidimensional diffusion MRI of the human brain. Neuroimage
2016;135:345-362.
- Topgaard D. Diffusion tensor
distribution imaging. NMR Biomed 2019;32:e4066.
- F.
Szczepankiewicz et.al., The link between diffusion MRI and tumor heterogeneity:
Mapping cell eccentricity and density by diffusional variance decomposition
(DIVIDE). Neuroimage. 142:522-532 (2016).
D. P. Kingma et.al., Adam: A method
for stochastic optimization, arXiv Prepr. arXiv1412.6980, 2014.