3D MR ventricle segmentation in pre-term infants with post-hemorrhagic ventricle dilatation (PHVD) using multi-phase geodesic level-sets
Introduction
The mild enlargement of the cerebral ventricles, called ventriculomegaly (VM), is often seen in neonates born prematurely with the highest risk population being those born at < 32 weeks gestation or with very low birth weight (< 1500 g). One of the primary non-congenital causes of ventriculomegaly is intraventricular hemorrhage (IVH), which is brain bleeding that occurs in 15–30% of very low birth weight pre-term neonates and is predictive of an adverse neurological outcome (Synnes et al., 2001, Adams-Chapman et al., 2008, Klebermass-Schrehof et al., 2013). Infants with IVH are at risk of developing progressive dilatation of the ventricles, a pathology called hydrocephalus. About 30–50% of infants with a severe IVH develop post-hemorrhagic ventricular dilatation (PHVD) and around 20–40% of those patients will consequently develop hydrocephalus, necessitating to be treated with a permanent ventriculo-peritoneal shunt (Adams-Chapman et al., 2008, Klebermass-Schrehof et al., 2013). Preterm infants are at risk of white matter injury, due to either a unilateral parenchymal hemorrhage or a more diffuse bilateral white matter damage, and the development of PHVD increases the risk of an adverse neurodevelopmental outcome. Around 45–60% of infants with PHVD have marked cognitive impairment (developmental quotient equivalent of less than 70) (Adams-Chapman et al., 2008, Klebermass-Schrehof et al., 2013). Any patient with VM will be followed for 1–2 years, as they are at risk of developing hydrocephalus. However, it is difficult to determine how quickly ventricle dilatation is progressing through currently used qualitative viewing of the MR images. During the follow-up, ultrasound (US) or MR images of their brains are acquired to quantitatively monitor the ventricular size. However, while clinicians have a good sense of what VM looks like compared to normal ventricles, there is little information available on what quantitative volumetric normal ranges are for ventricle volume of the cerebral ventricles in pre-term neonates as they mature, and only a few investigative studies (Haiden et al., 2005, Knickmeyer et al., 2008) have been conducted. In addition, VM is considered to be related to diagnosed anomalies of the central nervous system (CNS) (Guibaud, 2009). Although the sources and outcomes of VM have been extensively investigated (Weisglas-Kuperus et al., 1992, Gaglioti et al., 2005, Breeze et al., 2007, Beeghly et al., 2010), much remains unknown due to its variable etiology, complicated pathophysiology, and poorly understood natural history (Gaglioti et al., 2005, Beeghly et al., 2010). Volumetric analysis based on volumetric and morphological biomarkers of the developing neonatal brain can greatly improve the diagnosis, prognosis, and treatment of VM, which requires a robust and accurate automatic segmentation algorithm.
Previous brain segmentation algorithms have been extensively investigated in adult brain MR images, leading to several successful methods and software packages, such as SPM (Ashburner and Friston, 2005) and FreeSurfer (Fischl et al., 2002).
However, there is limited publicly available software for neonatal brain MR image segmentation. Neonatal brain MR images tend to suffer from high image noise, low tissue contrast, and considerable inter-subject anatomical variability, which make most of the methods used for the adult population not applicable to these images (Gui et al., 2012). Segmentation is even more difficult in infants with IVH as their ventricles contain heterogeneous cerebrospinal fluid (CSF) due to the hemorrhage. Wang et al. (2011) developed a coupled level-set method initialized by convex optimization technique for automatic neonatal MR images. Shi et al. (2010a) built a subject-specific tissue probabilistic atlas to segment longitudinal neonatal brain MR images. Wang et al. (2014a) integrated this probabilistic atlas into a patch-driven level-set framework for more accurate segmentation of neonatal MR images. Three different image sets of preterm infants provided in the NeoBrainS12 study (http://neobrains12.isi.uu.nl) (Isgum et al., 2015) were set up to allow a comparison of 8 brain tissue segmentation methods. The results demonstrated that the participating methods were able to segment all tissue classes well, except myelinated white matter. Even though several methods tend to perform better than others, areas of poor segmentation were common to all methods.
Most of previous work (Gui et al., 2012, Wang et al., 2011, Wang et al., 2014a, Shi et al., 2010a, Isgum et al., 2015) segmented neonatal brain MR images into different brain tissues, such as white matter (WM), gray matter (GM), and CSF, but did not focus on the ventricular system. Moreover, these methods were only validated on high-quality healthy neonate images generated for research purposes only while patients were sedated. Gholipour et al. (2012) developed a ventricle segmentation method of fetal brain MRIs using 13 normal brains and 12 fetuses with mild VM, which is the most similar study to ours. However, unlike the healthy brains and the brains with mild VM, the MR images of neonatal patients with IVH and hydrocephalus used in this study provide greater challenges for segmentation. The poor physical condition of these patients does not allow for a normal acquisition time or patient sedation, leading to a much poorer image quality. Furthermore, heterogeneous CSF with degrading blood products and dilatation of the ventricles cause large brain deformation and drastic inter-subject variations in the anatomy from patient to patient (Fig. 1). The blood clots from IVH are not only in the ventricles, but also at the junction between the ventricles and the brain, and infiltrating the brain creating heterogeneous imaging properties. For some severe cases, the ventricle volume is > 70% of the whole brain compared to < 30% of patients with mild VM, making it much more challenging to capture such large shape variability for most multi-atlas registration-based approaches (Warfield et al., 2004, Aljabar et al., 2009). Although manual segmentation is an option, it is too arduous and time consuming to be clinically feasible. Thus, an accurate and robust automatic cerebral ventricle segmentation algorithm from IVH neonatal MRIs is still highly desirable in clinical practice in order to handle images of poor quality and large shape variability.
In this study, we propose a multi-region segmentation approach for delineating the brain ventricle system of pre-term IVH neonate patients from acquired 3D T1 weighted (T1w) MR images. The proposed algorithm is initialized by non-linearly registering multiple pre-segmented patient images onto the subject image using a duality-based convex optimization registration scheme. A variational time-implicit multiphase geodesic level-sets (MGLS) is then proposed to automatically extract the ventricle system, which incorporates an intensity probability density function (PDF) and a probabilistic labeling map as shape priors generated in the initialization procedure. In particular, the major components of the whole segmentation pipeline are implemented using general-purpose programming on graphics processing units (GPGPU) to obtain a high computational efficiency.
The proposed MGLS algorithm is the extension of a 2015 EMMCVPR paper (Rajchl et al., 2015), which only provided a segmentation framework, and was evaluated on one 3D adult MR image. Compared to the previous version based on classic level sets (Rajchl et al., 2015), the geodesic level-sets (Criminisi et al., 2008) is incorporated into the surface evolution to improve the segmentation accuracy, especially for the hydrocephalus patient images. The improved algorithm has been evaluated on a larger number of patient images. To the best of our knowledge, this paper is the first study focusing on automated ventricle segmentation of pre-term neonatal MR images with VM and hydrocephalus.
Section snippets
Segmentation pipeline
Fig. 2 shows an overview of the proposed segmentation pipeline. An input subject image is first classified into a specified patient group (patients with mild enlarged ventricles or hydrocephalus patients) based on the Bhattacharyya distance (Michailovich et al., 2007) between the background and foreground defined by registered pre-segmented labels. Multiple manually pre-segmented patient images from the specified group are registered to the subject image using affine and deformable registration
Image acquisition
All patients were imaged following a protocol approved by the ethics review board at the University of Western Ontario (REB #100315). After an initial diagnosis of IVH on a clinical 2D US exam, the parents of the patient were approached to enroll their child in the study. Imaging sessions would be canceled if patients were not stable as judged by the attending team and nursing staff. Preterm-born patients with different IVH grades were imaged upon reaching term-equivalent age with a 1.5T Signa
Discussion and conclusion
This paper proposes an accurate solution to a challenging segmentation problem of the pre-term IVH neonatal cerebral ventricle from 3D MR images. The proposed segmentation algorithm makes use of the convex optimization technique associated with spatial shape prior, which is built via a multi-atlas non-linear registration scheme. The proposed segmentation pipeline is used for monitoring the ventricular size of patients with mild ventricle enlargement and hydrocephalus. The experimental results
Acknowledgments
The authors are grateful for the funding support from the Canadian Institutes of Health Research (CIHR) and Academic Medical Organization of Southwestern Ontario (AMOSO).
References (82)
- et al.
Multi-atlas based segmentation of brain images: atlas selection and its effect on accuracy
NeuroImage
(2009) - et al.
Probabilistic segmentation of brain tissue in MR imaging
NeuroImage
(2005) - et al.
Unified segmentation
NeuroImage
(2005) - et al.
Measurement of progressive cerebral ventriculomegaly in infants after grades III and IV intraventricular hemorrhages
J. Pediatr.
(1990) - et al.
A review of atlas-based segmentation for magnetic resonance brain images
Comput. Methods Prog. Biomed.
(2011) - et al.
Patch-based segmentation using expert priors: application to hippocampus and ventricle segmentation
NeuroImage
(2011) - et al.
Brain templates and atlases
NeuroImage
(2012) - et al.
Whole brain segmentation: automated labeling of neuroanatomical structures in the human brain
Neuron
(2002) - et al.
Distribution metrics and image segmentation
Linear Algebra Appl.
(2007) - et al.
Multi-atlas multi-shape segmentation of fetal brain MRI for volumetric and morphometric analysis of ventriculomegaly
NeuroImage
(2012)
Morphology-driven automatic segmentation of MR images of the neonatal brain
Med. Image Anal.
A spatiotemporal atlas of MR intensity, tissue probability and shape of the fetal brain with application to segmentation
NeuroImage
3-d ultrasonographic imaging of the cerebral ventricular system in very low birth weight infants
Ultrasound Med. Biol.
Automatic anatomical brain MRI segmentation combining label propagation and decision fusion
NeuroImage
Improving intersubject image registration using tissue-class information benefits robustness and accuracy of multi-atlas based anatomical segmentation
NeuroImage
Evaluation of automatic neonatal brain segmentation algorithms: the NeoBrainS12 challenge
Med. Image Anal.
Evaluation of prostate segmentation algorithms for MRI: the PROMISE12 challenge
Med. Image Anal.
Fast free-form deformation using graphics processing units
Comput. Methods Prog. Biomed.
Multi-contrast human neonatal brain atlas: application to normal neonate development analysis
NeuroImage
Incidence and evolution of subependymal and intraventricular hemorrhage: a study of infants with birth weights less than 1,500 gm
J. Pediatr.
Using the logarithm of odds to define a vector space on probabilistic atlases
Med. Image Anal.
Dual optimization based prostate zonal segmentation in 3D MR images
Med. Image Anal.
Mammography segmentation with maximum likelihood active contours
Med. Image Anal.
Neonatal brain image segmentation in longitudinal MRI studies
NeuroImage
Construction of multi-region-multi-reference atlases for neonatal brain MRI segmentation
NeuroImage
Variations in intraventricular hemorrhage incidence rates among Canadian neonatal intensive care units
J. Pediatr.
New geodesic distance transforms for gray-scale images
Pattern Recogn. Lett.
Automatic segmentation of neonatal images using convex optimization and coupled level sets
NeuroImage
Segmentation of neonatal brain {MR} images using patch-driven level sets
NeuroImage
Neonatal cerebral ultrasound, neonatal neurology and perinatal conditions as predictors of neurodevelopmental outcome in very low birthweight infants
Early Hum. Dev.
The effect of template choice on morphometric analysis of pediatric brain data
NeuroImage
User-guided 3D active contour segmentation of anatomical structures: significantly improved efficiency and reliability
NeuroImage
Neurodevelopmental outcome of extremely low birth weight infants with posthemorrhagic hydrocephalus requiring shunt insertion
Pediatrics
Curvature-driven flows: a variational approach
SIAM J. Control. Optim.
Probabilistic brain tissue segmentation in neonatal magnetic resonance imaging
Pediatr. Res.
The Isometric Log-Ratio Transform for Probabilistic Multi-Label Anatomical Shape Representation
Neurodevelopmental outcome of fetuses referred for ventriculomegaly
Ultrasound Obstet. Gynecol.
An experimental comparison of min-cut/max-flow algorithms for energy minimization in vision
IEEE Trans. Pattern Anal. Mach. Intell.
Fast approximate energy minimization via graph cuts
IEEE Trans. Pattern Anal. Mach. Intell.
An integral solution to surface evolution pdes via geo-cuts
Obstetric and neonatal outcomes in severe fetal ventriculomegaly
Prenat. Diagn.
Cited by (22)
Automatic determination of ventricular indices in hydrocephalic pediatric brain CT scan
2023, Interdisciplinary Neurosurgery: Advanced Techniques and Case ManagementCitation Excerpt :Accurate automatic tracking of ventricular indices would significantly improve the lifelong clinical management of hydrocephalic patients. There is more ventricular shape heterogenicity and distortion in pediatric hydrocephalic patients than adults and automated methods for evaluating pediatric brain sub-structures have also been limited to small cohorts [17,18]. According to the mentioned limitations of the existing methods, we propose image processing and analysis algorithms for assessment of pediatric hydrocephalus based on CT scans.
Automatic segmentation and location learning of neonatal cerebral ventricles in 3D ultrasound data combining CNN and CPPN
2021, Computers in Biology and MedicineCitation Excerpt :proposed a CNN which performs segmentation based on features from three independent branches with different patch size as input and [18] combined spatial and non-spatial dictionnary learning to achieve CVS segmentation. An active contour based method was also proposed by Ref. [19]. It used a multiphase geodesic level-set which utilizes a spacial prior obtained from multi-atlas registration.
Automated brain extraction from head CT and CTA images using convex optimization with shape propagation
2019, Computer Methods and Programs in BiomedicineNovel and powerful 3D adaptive crisp active contour method applied in the segmentation of CT lung images
2017, Medical Image AnalysisCitation Excerpt :On the other hand, geometric ACMs move the curve by minimizing the energy required based on a function of statistical probability (Leninisha and Vani, 2015; Mesejo et al., 2015; Rebouças et al., 2016). There are different versions of these models in 2D and 3D, which are commonly called Level Set models, including the Geodesic model (Diciotti et al., 2011; Qiu et al., 2015) and other models developed to optimize performance (Wang et al., 2011, 2014). This paper proposes a new parametric 3D active contour model specifically to segment complex objects such as the lung, and not only objects with cylindrical topology and regular shape, as the one proposed in Schmitter et al. (2015).
AI-based medical e-diagnosis for fast and automatic ventricular volume measurement in patients with normal pressure hydrocephalus
2023, Neural Computing and Applications
- 1
Contributed equally.