Automated CSF Detection for Post-hemorrhagic Hydrocephalus in Preterm Infants Using 3D U-Net
Li Zhao1, Xue Feng2, Craig Meyer2, Kushal Kapse1, Matthew T. Whitehead1, Adre J. du Plessis3, and Catherine Limperopoulos1

1Diagnostic Imaging and Radiology, Childrens National Medical Center, Washington, DC, United States, 2Biomedical Engineering, University of Virginia, Charlottesville, VA, United States, 3Fetal Medicine, Childrens National Medical Center, Washington, DC, United States


Post-hemorrhagic hydrocephalus is a prevalent and severe neurological complication in very premature infants. Converging evidence suggests that increased ventricular size is an important and potentially modifiable risk factor for adverse neurological outcomes. MRI measures of CSF volume often rely on manual measurements to quantify ventricular size because automatic neonatal brain segmentation methods often fail in the setting of severe brain injury. In this pilot study, we proposed and validated a deep convolutional neural network method, 3D U-Net, to automatically identify the lateral ventricular system and the external cerebrospinal fluid regions. The proposed method showed superior accuracy in a preliminary cohort of 19 scans of very preterm infants compared to a conventional method.


Hydrocephalus is a central nervous system disease caused by the imbalance between cerebrospinal fluid (CSF) production and absorption1. Post-hemorrhagic hydrocephalus (PHH) commonly occurs in the preterm infant and it can result in devastating neurological complications including cognitive deficits and cerebral palsy. Currently, hydrocephalus diagnosis mainly relies on the measurement of the ventricular size based on neuro-images. Although ventricle segmentation methods are well established for the healthy adult brain, reliable methods for the preterm brain with significant brain injury are still lacking. The methods based on probability maps2 can fail in the HPP cases due to rapid brain tissue and ventricular growth, and the presence of brain lesions.

U-Net is a powerful method for automatic segmentation utilizing a deep convolutional neural network. Its performance has been previously demonstrated3. In this work, we showed our initial results for CSF segmentation using 19 MRI studies from preterm infants with PHH.


We prospectively enrolled very preterm infants <32 weeks and <1500 grams at birth who underwent two MRI scans: once in the ‘preterm period’ as soon as they were medically stable and the second at term equivalent age. Preterm infants requiring temperature monitoring were scanned using an MRI-compatible incubator in a GE 1.5T scanner. Preterm babies that were clinically stable or term-equivalent age were scanned in a GE 3.0T scanner. 19 scans were performed on 14 infants. The mean gestational age at MRI was 38±4.7 weeks. PHH was confirmed as atrial diameter enlargements in one or both lateral ventricles.

T2-weighted images were acquired with fast spin echo (FSE) sequences. In preterm infants, 2D single-shot FSE sequences were acquired in coronal, sagittal and axial planes with matrix size 160x160, slice thickness 2mm, and TE/TR = 160/1100ms. Each scan was repeated twice to avoid possible motion artifacts in further processing. High-resolution 3D images were reconstructed from 2D images with the slice-to-volume method4. In term equivalent age infants, 3D FSE sequences were acquired with 1mm isotropic voxel size. The echo train length was 100 and TE/TR was 64/2500ms.

Conventional CSF segmentations were performed by the Developing brain Region Annotation With Expectation-Maximization (Draw-EM) tool5. The algorithm segments the brain into nine tissue-labels, however, only extra-axial CSF and lateral ventricle CSF were used in this study. Based on the conventional segmentation, the regional tissue labels were manually corrected by experienced technicians, which served as the ground truth in the following 3D U-Net model training and validation. The third and the fourth ventricles were treated extra-axial CSF in the templates of the conventional segmentation, which were kept the same in the manual corrections.

The ventricle segmentation was performed using a modified 3D U-Net structure. In this work, the U-Net was implemented with a parametric rectified linear unit (PReLU), instead of the rectified linear unit. The threshold level was also determined during the training. Furthermore, the first layer was changed to 96 filters. In the training process, 64x64x64 patches were extracted from the 3D T2-weighted images randomly. For argumentations, images were flipped randomly along the left-right direction. The cross-entropy loss was calculated based on the ground truth with manual corrections. The extra-axial and ventricular CSF regions were identified in this model. 15 MRI scans were used to train the model and the remaining 4 scans were used for validation. Dice coefficient was used to evaluate the performance. The proposed method was implemented on Tesla P100 graphics processing units.


In the validation data of the 4 infants with PHH, the proposed method yielded a dice coefficient of 0.825 for the extra-axial CSF regions and 0.944 for the lateral ventricular CSF. Figure 1 shows one example from a 43 4/7-week-old preterm infant with PPH in the validation data. The yellow arrows highlighted the regions of ventricles that were mislabeled as extra-axial CSF by the conventional segmentation method. The green arrows highlighted the eye region that was mislabeled as the extra-axial CSF regions in the conventional method. The proposed method of 3D U-Net provided very comparable outputs to the manually corrected ground truth and it also avoided the mislabeled regions of the conventional method. The accuracy of the proposed method is considerable, although the differences from the ground truth require further evaluations.


We demonstrated the performance of modified 3D U-Net on the accuracy of CSF segmentation of preterm infants with PHH. The proposed method showed superior performance for CSF delineation in the PHH cohort, compared to the conventional method. The proposed method may provide a reliable and efficient tool for automatic detection and quantitative analysis of progressive PHH in high-risk preterm infants.


This work was partly supported by DC-IDDRC 1U54HD090257 grant.


1. Vinchon M, Rekate H, Kulkarni A V. Pediatric hydrocephalus outcomes: a review. Fluids Barriers CNS 2012;9:1 doi: 10.1186/2045-8118-9-18.

2. Ashburner J, Friston KJ. Unified segmentation. Neuroimage 2005;26:839–851 doi: 10.1016/j.neuroimage.2005.02.018.

3. Ronneberger O, Fischer P, Brox T. U-Net: Convolutional Networks for Biomedical Image Segmentation. 2015:1–8 doi: 10.1007/978-3-319-24574-4_28.

4. Kainz B, Steinberger M, Wein W, et al. Fast Volume Reconstruction from Motion Corrupted Stacks of 2D Slices. IEEE Trans. Med. Imaging 2015;34:1901–1913 doi: 10.1109/TMI.2015.2415453.

5. Makropoulos A, Gousias IS, Ledig C, et al. Automatic whole brain MRI segmentation of the developing neonatal brain. IEEE Trans. Med. Imaging 2014;33:1818–1831 doi: 10.1109/TMI.2014.2322280.


Figure 1 Internal and external CSF identification on a neonate with hydrocephalus at the equivalent gestational age of 43 4/7 weeks. a) high-resolution T2 weighted images, b) segmentation of conventional template method, c) manual corrected mask, and d) the proposed method with 3D U-Net. The red regions represent extra-axial CSF and the blue regions represent ventricles.

Proc. Intl. Soc. Mag. Reson. Med. 27 (2019)