Jeremiah W Sanders1, Jason M Johnson2, Jong Bum Son1, Zijian Zhou1, Henry Szu-Meng Chen1, Joshua Yung1, Jason Szu-Meng Stafford1, Melissa Chen2, Maria Gule-Monroe2, Ho-Ling Liu1, Mark D Pagel3, and Jingfei Ma1
1Imaging Physics, University of Texas MD Anderson Cancer Center, Houston, TX, United States, 2Neuroradiology, University of Texas MD Anderson Cancer Center, Houston, TX, United States, 3Cancer Systems Imaging, University of Texas MD Anderson Cancer Center, Houston, TX, United States
Synopsis
Neurocognitive function is often associated with structural differences
in the brain for patients with neurofibromatosis type-1 (NF1), and studies have
shown that NF1 is associated with larger subcortical volumes and thicker
cortices of certain brain structures. Routine monitoring of NF1 patients would
be possible with tools that enable rapid whole-brain segmentation in standard of care
T1w MRI. Modern machine learning techniques, including fully convolutional
networks (FCNs), have demonstrated the ability to rapidly perform segmentation tasks
across a range of applications. In this work, we investigate the performance of
different FCNs for rapid whole-brain segmentation in pediatric
T1w brain MRI.
Introduction
Neurocognitive function is often associated with structural differences
in the brain for patients with neurofibromatosis type-1 (NF1), and studies have
shown that NF1 is associated with larger subcortical volumes and thicker
cortices of certain brain structures [1]. Routine monitoring of NF1 patients
would be possible with tools that enable whole-brain segmentation in standard
of care T1w MRI. Methods exist that will allow whole-brain segmentation in T1w
MRI, but these methods are currently time and resource-intensive. Modern
machine learning techniques, including fully convolutional networks (FCNs),
have demonstrated the ability to perform segmentation tasks rapidly across a range of
applications. The purpose of this work was to investigate the performance of
different FCNs for rapid and automated whole-brain segmentation in pediatric
T1w brain MRI.Methods
Data used in the preparation of this article were obtained from the
Adolescent Brain Cognitive Development (ABCD) Study (https://abcdstudy.org),
held in the NIMH Data Archive (NDA). This is a multisite, longitudinal study
designed to recruit more than 10,000 children age 9-10 and follow them over ten
years into early adulthood. 107 pediatric patients from this archive were
scanned with a 3D spoiled gradient-echo MR sequence. Typical scan parameters
were TR/TE = 7.0/2.9 ms, flip angle = 8°, FOV = 26 × 21 cm, and voxel size = 1
× 1 × 1 mm. Brain segmentation masks for each patient volume were computed
using FreeSurfer [2]. A total of 43 tissue classes were segmented (plus an
additional class for background for a total of 44). Each slice of the patient
image volumes was normalized to the range [0, 1]. The patient images were
randomly split into 82 and 25 cohorts for training/validation and testing,
respectively.
Seven different constructs of FCNs were constructed (Figure 1).
Each was created with a convolutional encoder and a convolutional decoder, with
skip connections between the encoder and decoder at equivalent resolution
scales. We investigated 7 different convolutional encoders including compact
modules (MobileNet), residual modules (ResNet50 and ResNet50V2), inception
modules (IncepionResNetV2), xception modules (Xception), dense modules
(DenseNet121), and stacked convolutions (VGG16). The decoder remained fixed
across all of the models and was similar to that of the original U-Net
architecture [3].
Eighteen different brain structures of primary interest were
identified. These included the right (R) and left (L) cerebral cortex white
matter (WM), R and L cerebral cortex grey matter (GM), R and L lateral
ventricle, R and L cerebellum WM, R and L cerebellum GM, R and L thalamus, R
and L caudate, R and L putamen, and R and L pallidum. The predicted
segmentation masks were evaluated for pixel accuracy, Dice similarity, and
intersection over union. For visualization, an open-source mesh generation
package [4] was used to create 3D volumetric meshes of the ground truth and
predicted brain segmentation masks.
All models were constructed and trained with a deep learning
application engine [5] on an NVIDIA DGX-1 workstation. The mini-batches of size
22 were distributed across 2 of the V100 GPUs during training. All models were
trained using cross-entropy loss and an Adam optimizer with a learning rate (LR)
of 0.0001. The LR was decayed by 20% when no improvement in the validation loss
occurred after five epochs, and training was terminated after 11 epochs with no
improvement in the validation loss.Results
Using stacked convolutions for the convolutional encoder
produced the highest overall pixel accuracy of 98.96% across all models and
brain structures (Table 1). Stacked convolutions also yielded the highest
accuracy across the organ-wise similarity metrics (Table 2, Dice coefficient).
Of the 18 brain structures analyzed, 12 of the predicted brain structure
segmentation masks had a mean Dice similarity of ≥ 90.0% with the FreeSurfer
predictions. The lowest Dice similarity was observed for the R palladium
(86.1%±5.8%), whereas the highest Dice similarity was observed for the L
cerebral cortex WM (96.6%±0.9%). All predictions (volumetrically) took <5 s
on an NVIDIA V100 GPU, and <1 min on an NVIDIA GTX 1050 GPU.Discussion and conclusion
Software tools exist that enable accurate, whole-brain segmentation in
T1w MRI. However, these techniques are currently resource-intensive and
time-consuming, which reduces their feasibility for routine implementation in
clinical workflows. We have investigated FCNs for rapid whole-brain
segmentation in T1w pediatric MRI. This approach provides accurate segmentation
masks in a fraction of the time required by currently available open-source
software techniques. It may enable implementation in routine clinical workflows
for multiple applications, including NF1 evaluation and image quality
evaluation in T1w MRI. Future work will focus on incorporating adult T1w MRIs
into the model for whole-brain segmentation across a range of ages and human
brain morphologies.Acknowledgements
No acknowledgement found.References
[1]
Barkovich MJ, Tan CH, Nillo RM, et al. “Abnormal morphology of select cortical
and subcortical regions in neurofibromatosis type 1,” Radiology, 2018; 289(2): 499-508.
[2]
Fischl B. “FreeSurfer,” Neuroimage,
2012; 62(2), 774–781.
[3]
Ronneberger O, Fischer P, Brox T. “U-Net: Convolutional networks for biomedical
image segmentation,” arXiv:1505.04597, 2015.
[4] Fang Q, Boas D,
"Tetrahedral mesh generation from volumetric binary and gray-scale
images," Proc IEEE International
Symposium on Biomedical Imaging, 2009; 1142-1145.
[5] Sanders JW,
Fletcher JR, Frank SJ, et al. “Deep learning application engine (DLAE):
development and integration of deep learning algorithms in medical imaging,” SoftwareX, 2019; 10:100347.