Jayse Merle Weaver1,2, Marissa DiPiero2,3, Patrik Goncalves Rodrigues2, Hassan Cordash2, and Douglas C Dean III1,2,4
1Medical Physics, University of Wisconsin-Madison, Madison, WI, United States, 2Waisman Center, University of Wisconsin-Madison, Madison, WI, United States, 3Neuroscience Training Program, University of Wisconsin-Madison, Madison, WI, United States, 4Pediatrics, University of Wisconsin-Madison, Madison, WI, United States
Synopsis
Keywords: Machine Learning/Artificial Intelligence, Artifacts, Quality Control
A
three-dimensional convolutional neural network was trained to detect motion
artifacts on a volume level for two pediatric diffusion MRI datasets acquired
between 1 month and 3 years of age. Accuracies of 95% and 98% were achieved
between the two datasets. Additionally, the effects of motion-corrupted volumes
on quantitative parameter estimation was examined. Data was processed without
quality control and with quality control performed by the neural network. DTI
and NODDI metrics were calculated and compared between methods. Significant
differences were found for both individual and group results.
Introduction
DTI and NODDI
are quantitative MRI techniques capable of probing microstructural changes
using diffusion MRI (dMRI) data1,2. The metrics obtained from fitting
the DTI and NODDI models provide sensitive markers for brain development and
disorders but may be confounded by artifacts in the dMRI images, such as those
from subject motion. Hence, quality control (QC) measures are needed to
identify such motion and other artifacts from dMRI data to either correct or
exclude these data from further analysis. The gold standard for QC of dMRI
data, especially pediatric dMRI, is visual inspection and removing
artifact-corrupted DWIs from further processing. However, this is
time-consuming and prone to subjective error. Thus, automated QC is desired.
Convolutional neural networks (CNNs)
have been used to perform automated QC of dMRI data3,4,5,6,7,8.
However, no prior works have performed automated QC using a CNN on pediatric
dMRI in the age range of 1 month to 3 years, a period of life targeted in
recent neuroimaging studies due to the rapid development of white matter9.
In this work, we propose a 3D-CNN for detecting motion artifacts in dMRI data
acquired during the first 3 years of life.Methods
For this work,
we used two datasets (Dataset A and B, Table 1) acquired with different
acquisition parameters and on populations with different age ranges. Dataset A
contains 151 dMRI datasets acquired at two time points, 1 and 24 months of age.
Dataset B contains 26 dMRI datasets acquired between 2 and 35 months of age.
All data were acquired during natural, non-sedated sleep.
Dataset A was used to create the
training and testing sets using an identical number of motion-corrupted and
motion-free volumes. To prevent data leakage, training and testing sets were
split by subject rather than volume. Dataset B served as an additional unseen
testing set. All volumes were resized to a common size (128x128x70) and
intensity normalized between 0 and 1.
The 3D-CNN architecture was
implemented in Python using Keras and Tensorflow. An overview of the network
architecture is depicted in Figure 1. The network consists of four feature
extraction blocks, with each block containing a 3D convolutional layer with an
increasing number of filters, a kernel size of 3x3x3, and ReLU activation. Each
convolutional layer is followed by a 3D max pooling layer with a pool size of 2
and batch normalization. The final output is flattened and passed to two fully
connected layers with a total dropout rate of 50%. Lastly, a dense layer with 1
neuron and sigmoid activation is used for binary classification.
The network was trained with a batch
size of 8, binary cross-entropy as the loss function, and the Adam optimizer
with an initial learning rate of 1e-3 and learning rate decay of 0.01. K-fold
cross-validation with k=4 was used to test the model’s dependence on the input
data.
The model with the highest accuracy
across both datasets was selected for use in a pre-processing and analysis
pipeline developed by our lab. All one-month-old subjects in the testing
dataset (n=24) were used for analysis. The pipeline was repeated three times
for each subject with different QC methods: motion-corrupted volumes were
identified by a human reader and removed (manual QC), identified by the neural
network and removed (model QC), or not removed at all (no QC). The data were
fit to the DTI model using the DIPY package10 and the NODDI model
using the Dmipy toolbox11.
A study-specific template was
created from FA maps using ANTs12. Corpus callosum and internal
capsule ROIs were obtained from the JHU Neonate Atlas13. ROIs were
then warped into each subject’s native space. FA, RD, and ICVF values were
extracted using the ROIs, and intrasubject and group differences between QC
methods were examined using t- and F-tests.Results
Mean accuracies
of 95.4% and 98.5% were achieved on Datasets A and B, respectively (Table 2).
Box plots displaying the mean FA, RD, and ICVF values for each region and QC
method are shown in Figure 2. The F-tests revealed significant
differences in the variances between QC methods. The results of individual t-
and F-tests (Table 3) show significant differences in the mean of at
least one quantitative measure for 19 of the 24 subjects.Discussion & Conclusion
While CNNs have
been previously used to perform automated QC of diffusion data, no studies have
trained and evaluated a network on data acquired from infants and toddlers. The
proposed CNN identifies motion artifacts with an accuracy greater than 95%. Additionally,
the removal of motion-corrupted volumes by either manual QC or neural network
QC causes significant differences in a subset of DTI and NODDI metrics on an
individual and group level.
The proposed network performs binary
classification focusing on detecting motion artifacts, the most common artifact
when scanning young children during natural, non-sedated sleep. Future work
will employ additional public datasets during network training to perform
multi-label classification of several types of artifacts on a wider range of
pediatric data.Acknowledgements
We sincerely thank our research
participants and their families who participated in this research as well as
the dedicated research staff who made this work possible. This work was
supported by grants P50 MH100031, 5R00 MH110596-05, and 1U01 DA055370-01 from
the National Institute of Mental Health, National Institutes of Health.
Infrastructure support was also provided, in part, by grant U54 HD090256 from
the Eunice Kennedy Shriver NICHD, National Institutes of Health (Waisman
Center).
References
-
Alexander, A. L., Lee, J. E., Lazar,
M., & Field, A. S. (2007). Diffusion tensor imaging of the brain. Neurotherapeutics
: the journal of the American Society for Experimental NeuroTherapeutics, 4(3),
316–329. https://doi.org/10.1016/j.nurt.2007.05.011
- Zhang, H., Schneider, T.,
Wheeler-Kingshott, C. A., & Alexander, D. C. (2012). NODDI: practical in
vivo neurite orientation dispersion and density imaging of the human brain. NeuroImage,
61(4), 1000–1016. https://doi.org/10.1016/j.neuroimage.2012.03.072
- Kelly,
C., Pietsch, M., Counsell, S., & Tournier, J. D. (2017, April). Transfer
learning and convolutional neural net fusion for motion artefact detection. In Proceedings
of the Annual Meeting of the International Society for Magnetic Resonance in
Medicine, Honolulu, Hawaii (Vol. 3523).
- Graham, M. S., Drobnjak, I., &
Zhang, H. (2018). A supervised learning approach for diffusion MRI quality control
with minimal training data. NeuroImage, 178, 668–676.
https://doi.org/10.1016/j.neuroimage.2018.05.077
- Samani, Z. R., Alappatt, J. A.,
Parker, D., Ismail, A., & Verma, R. (2020). QC-Automator: Deep
Learning-Based Automated Quality Control for Diffusion MR Images. Frontiers
in neuroscience, 13, 1456. https://doi.org/10.3389/fnins.2019.01456
- Ahmad,
A., Parker, D., Samani, Z. R., & Verma, R. (2021). 3D-QCNet--A pipeline for
automated artifact detection in diffusion MRI images. arXiv preprint
arXiv:2103.05285.
- Ettehadi, N., Zhang, X., Wang, Y.,
Semanek, D., Guo, J., Posner, J., & Laine, A. F. (2021). Automatic
Volumetric Quality Assessment of Diffusion MR Images via Convolutional Neural
Network Classifiers. Annual International Conference of the IEEE Engineering
in Medicine and Biology Society. IEEE Engineering in Medicine and Biology
Society. Annual International Conference, 2021, 2756–2760.
https://doi.org/10.1109/EMBC46164.2021.9630834
- Ettehadi, N., Kashyap, P., Zhang,
X., Wang, Y., Semanek, D., Desai, K., Guo, J., Posner, J., & Laine, A. F.
(2022). Automated Multiclass Artifact Detection in Diffusion MRI Volumes via
3D Residual Squeeze-and-Excitation Convolutional Neural Networks. Frontiers
in human neuroscience, 16, 877326. https://doi.org/10.3389/fnhum.2022.877326
- Lebel, C., & Deoni, S. (2018).
The development of brain white matter microstructure. NeuroImage, 182,
207–218. https://doi.org/10.1016/j.neuroimage.2017.12.097
- Garyfallidis,
E., Brett, M., Amirbekian, B., Rokem, A., van der Walt, S., Descoteaux, M.,
Nimmo-Smith, I., & Dipy Contributors (2014). Dipy, a library for the
analysis of diffusion MRI data. Frontiers in neuroinformatics, 8,
8. https://doi.org/10.3389/fninf.2014.00008
- Fick,
R., Wassermann, D., & Deriche, R. (2019). The Dmipy Toolbox: Diffusion MRI
Multi-Compartment Modeling and Microstructure Recovery Made Easy. Frontiers
in neuroinformatics, 13, 64.
https://doi.org/10.3389/fninf.2019.00064
- Avants,
B. B., Tustison, N. J., Song, G., Cook, P. A., Klein, A., & Gee, J. C.
(2011). A reproducible evaluation of ANTs similarity metric performance in
brain image registration. NeuroImage, 54(3), 2033–2044.
https://doi.org/10.1016/j.neuroimage.2010.09.025
- Oishi,
K., Mori, S., Donohue, P. K., Ernst, T., Anderson, L., Buchthal, S., Faria, A.,
Jiang, H., Li, X., Miller, M. I., van Zijl, P. C., & Chang, L. (2011).
Multi-contrast human neonatal brain atlas: application to normal neonate
development analysis. NeuroImage, 56(1), 8–20.
https://doi.org/10.1016/j.neuroimage.2011.01.051