Bin Chen1, Liming Wu1, Bing Zhang2, Simin Liu3, and Hua Guo3
1Purdue University Northwest, Hammond, IN, United States, 2Nanjing University Medical School, Nanjing, China, 3Tsinghua University, Beijing, China
Synopsis
Recent advances show promising fetal
brain reconstruction results through image motion correction and super
resolution from a stack of unregistered images consisting of in-plane motion
free snapshot slices acquired by fast imaging methods. Most motion correction
and super resolution techniques for 3D volume reconstruction require accurate
fetal brain segmentation as the first step of image analysis. In this study, a
customized U-Net based deep learning method was implemented for automatic fetal
brain segmentation. The high accuracy of deep learning based semantic
segmentation improves the performance in volume registration as well as quantitative
studies of brain development and group analysis.
Introduction
Fetal
MRI has been increasingly used in quantitative brain development studies as
well as normal/abnormal development diagnosis for the gestational age (GA) over
20 weeks [1]. While fast imaging methods are capable of acquiring in-plane motion
free 2D snapshot, unconstrained fetal motion in uterus between slice acquisitions
usually leads to unregistered image volumes. Recent advances show promising
volume reconstruction results from a stack of unregistered images through fetal
image motion correction and super resolution volume reconstruction. The
performance of most techniques in 3D volume registration reconstruction usually
depends on the accuracy of fetal brain segmentation [2,3]. In this study, a customized
U-Net [4] based deep learning method was implemented for automatic fetal brain
segmentation. The results are evaluated by commonly used measures in medical image
segmentation, and
compared with FSL’s popular brain extraction tool (BET).
Methods
A
modified U-Net model was implemented to better fit the fetal brain
segmentation task. The dimension for the input layer was resized from 480×480
to 512×512 pixels for convenient feature concatenation in implementation. The
convolution was set to have the kernel size of 3×3, stride 2, and zero padding.
The batch normalization was adopted for faster and more stable network
convergence. Manually
segmented images in 52 volumes from 19 fetuses with GA between 30-33 weeks
were served as the training dataset. The binary cross-entropy loss was
calculated between the ground truth and the output
prediction. The SDG optimizer was set to a learning rate of 0.01 with a decay
rate of 50% for every 10 epochs. The network converged and the loss started to
plateau after 20 epochs. The training time was approximately 5 hours on a
computer with two Nvidia Geforce GTX 1080 graphic cards. Both the trained
network and FSL’s BET program were applied to the test dataset consisting of the subjects not in the training dataset for
comparison. The results from
FSL’s BET with the fractional intensity threshold of 0.80, the best
segmentation results by visual inspection with the threshold ranging from 0.5
to 0.9, were used as a reference. The performance of segmentation was
evaluated by intersection over union (IOU), Dice coefficient, sensitivity and specification.Results
Figure
1 shows the segmented areas of a fetal brain with the gestational age of 31 weeks in the test dataset. The green contours represent the
manual segmented ground truth. The red contours are predicted segmentation by
the trained deep learning network and FSL’s BET program. The overlapped contours
of the ground truth and predicted segmentation are shown in yellow. The results show that deep
learning based segmentation has high resemblance in shape and size to the
manual segmentation. The quantitative measures for performance evaluation are
shown in Table 1. Other than the Dice coefficient, the IOU score provides an
intuitive way to illustrate the resemblance between the ground truth and predicted
segments. Both techniques have high sensitivity and specificity scores.Discussions
Segmentation
usually serves as a first step for quantitative analysis. Deep learning based
techniques have demonstrated their performances in challenging tasks of medical
image segmentation in recent years. Our results show that the trained network can
accurately extract the brain regions of a fetus with similar
GA weeks. A further study may evaluate the segmentation performance between a large one-fit-all network for all gestational ages and multiple trained networks for different GA stages.Acknowledgements
No acknowledgement found.References
-
Makropoulos
A, Counsell S, Rueckert D. A Review on Automatic Fetal and Neonatal Brain MRI
Segmentation. NeuroImage. 2018;170:231-248.
- Keraudren
K, Kuklisova-Murgasova M, Kyriakopoulou V, Malamateniou C, Rutherford MA, Kainz
B, Hajnal JV, Rueckert D. Automated fetal brain segmentation from 2D MRI slices
for motion correction. Neuroimage. 2014; 101:633-43.
- Tourbier
S, Velasco-Annis C, Taimouri V, Hagmann P, Meuli R, Warfield S, Cuadra M,
Gholipour A. Automated Template-based Brain Localization and Extraction for
Fetal Brain MRI Reconstruciton. NeuroImage. 2017; 155:460-472.
- Ronneberger
O, Fischer P, Brox T. U-Net: Convolutional Networks for Biomedical Image
Segmentation. 2015. Medical Image Computing and Computer-Assisted Intervention
(MICCAI), Springer, LNCS, 9351: 234-241.