Elsevier

NeuroImage

Volume 65, 15 January 2013, Pages 97-108
NeuroImage

AdaPT: An adaptive preterm segmentation algorithm for neonatal brain MRI

https://doi.org/10.1016/j.neuroimage.2012.08.009Get rights and content

Abstract

Advances in neonatal care have improved the survival of infants born prematurely although these infants remain at increased risk of adverse neurodevelopmental outcome. The measurement of white matter structure and features of the cortical surface can help define biomarkers that predict this risk. The measurement of these structures relies upon accurate automated segmentation routines, but these are often confounded by neonatal-specific imaging difficulties including poor contrast, low resolution, partial volume effects and the presence of significant natural and pathological anatomical variability. In this work we develop and evaluate an adaptive preterm multi-modal maximum a posteriori expectation-maximisation segmentation algorithm (AdaPT) incorporating an iterative relaxation strategy that adapts the tissue proportion priors toward the subject data. Also incorporated are intensity non-uniformity correction, a spatial homogeneity term in the form of a Markov random field and furthermore, the proposed method explicitly models the partial volume effect specifically mitigating the neonatal specific grey and white matter contrast inversion. Spatial priors are iteratively relaxed, enabling the segmentation of images with high anatomical disparity from a normal population. Experiments performed on a clinical cohort of 92 infants are validated against manual segmentation of normal and pathological cortical grey matter, cerebellum and ventricular volumes. Dice overlap scores increase significantly when compared to a widely-used maximum likelihood expectation maximisation algorithm for pathological cortical grey matter, cerebellum and ventricular volumes. Adaptive maximum a posteriori expectation maximisation is shown to be a useful tool for accurate and robust neonatal brain segmentation.

Highlights

► Very preterm birth increases the risk of subsequent learning difficulties. ► Accurate segmentation may allow predictive biomarkers to be established. ► Adaptive segmentation allows improved segmentation in pathological cases. ► Comparison of the algorithm to manual segmentation shows significant improvement.

Introduction

Babies born very preterm (less than 32 weeks of gestation) are at increased risk of a range of cognitive and learning problems that become more frequent with lower gestations at birth (Marlow et al., 2005, Mathur and Inder, 2009, Ment et al., 2009). Although survival rates of preterm infants have improved significantly over the last few decades, this has not been accompanied by a reduction in rates of neurodisability, or improvement in the cognitive outcome for survivors. As a result, a large proportion of very preterm children have disabilities and special educational needs with consequent high societal costs (Mangham et al., 2009). Following very preterm birth, brain development occurs outside the normal protective fetal environment, exposing the infant to a range of external stimuli and problems of homeostasis. Changes seen on MRI at term equivalent age are likely to represent the amalgamation of cellular injury and disturbance of normal brain development (Khwaja and Volpe, 2008, Volpe, 2009).

Studies using advanced magnetic resonance techniques over the past decade have begun to identify differences in size and structure of the brains of preterm infants imaged at term compared with term born controls and correlation with neurodevelopmental outcome at 2 years (Boardman et al., 2010). Fundamental to performing volumetric and morphometric studies is the ability to classify different brain tissues. In contrast to adults, the segmentation of the neonatal brain is complicated due to a combination of: low-signal-to noise ratio; increased voxel partial volume (PV) as a result of adapting the resolution to the smaller neonatal head; and the existence of both natural and pathological hypo- and hyper‐intensities. In addition, there is substantial natural and pathological variability due to the effects of prematurity and the spectrum of preterm brain injury (Kapellou et al., 2006, Rutherford et al., 2010).

Maturational dependent intensity differences are well-known between the adult and neonatal brain. There is a dynamic natural developmental variability which arises due in part to the receding germinal matrix and progressive myelination, which manifests itself as an apparent reversal of the signal intensities of grey and white matter (GM/WM) on neonatal T1- and T2-weighted MRI. As myelination proceeds during the first months of life, contrast between the two tissue types progressively changes until an adult intensity pattern emerges around 2 years of age. The complexity of the cortical surface also increases rapidly over the period between 20 and 40 weeks gestational age, corresponding to the preterm period raising the possibility that cortical surface analysis may provide an independent predictive biomarker of neurological outcome (Dubois et al., 2008).

As a result of the crucial role of segmentation, a number of authors have produced techniques specifically for neonatal MRI, primarily by adapting and enhancing well-established techniques in the adult brain. Prastawa et al. (2005) developed a technique based on the canonical expectation maximisation (EM) method of Van Leemput et al. (1999), explicitly modelling additional classes of white matter to account for the myelination process, adding an extra tissue class for myelinated white matter which they found primarily in the deep brain making this technique a useful possibility for studies of myelination changes during infancy. The authors subsequently applied their technique to larger volumes of data in Gilmore et al. (2007). Xue et al. (2007) proposed a series of improvements to neonatal segmentation specifically for cortical segmentation, combining the brain extraction technique with deep grey matter structure removal and particularly highlighting the problem of misclassified partial volume (PV) pixels (e.g. unlikely white matter pixels appearing between dark cortical grey-matter and light cerebral spinal fluid on T2-weighted MRI) through morphological operations without implicitly modelling PV within the EM procedure. This addition is specific to neonatal MR contrast, but equivalent to the corresponding adult problem of unlikely grey matter pixels appearing between white matter and ventricular CSF. The authors subsequently used the segmentation results to extract and analyse the cortical surface using a level-set routine for further processing. Crucially, in the absence of a neonatal atlas, the authors initialised the segmentation with the results of a k-means clustering. Both Prastawa et al. (2005) and Xue et al. (2007) applied their techniques to infants with normal anatomical appearance. More recently, Shiee et al. (2011) addressed the problem of the segmentation of cases that are far from the space of the priors, specifically adult ventriculomegaly, by allowing an iterative ‘relaxation’ of the anatomical priors on the data, an approach likely to be useful for the diverse anatomy of neonatal cohorts. This procedure thus allows pixels far from the prior atlas to become progressively incorporated in the classification allowing the segmentation of pathological cases. An alternative segmentation methodology was proposed by Weisenfeld and Warfield (2009), employing template propagation and fusion to estimate the most likely tissue classifications (including classes for unmyelinated white matter and subcortical grey matter). In this algorithm, sub-structures that will not be identified in a three or four class expectation-maximisation routine may be extracted and analysed and the method mitigates the effect of low signal and contrast to noise ratios. However, template driven segmentation methods rely on the availability and registration of well-defined templates in order to propagate knowledge of a particular population to a patient specific space. This is problematic in pathological cases, as their anatomy may be markedly different from the one derived from a normal population atlas and the process may not be mitigated by registration.

More recent advances in segmentation methodology have seen a number of atlas-driven segmentation methods proposed as a result of a number of groups developing and making available specific neonatal atlases, averaged over a number of subjects: Knickmeyer et al. (2008) developed a three-class atlas-based approach from birth to 2 years, analysing population changes over this period and thus not distinguishing between myelinated and unmyelinated white matter; Oishi et al. (2011) made use of diffusion-weighted data to produce a multi-contrast neonatal brain atlas; whilst Kuklisova-Murgasova et al. (2011) have made available a comprehensive neonatal atlas from 29 to 44 weeks generated from T2-weighted data with intensity classes for grey matter, white matter and CSF and spatial sub-classes for sub-cortical grey matter, cerebellum and brainstem. Song et al. (2007) proposed a simpler intensity based classification method based on K-means and augmented by a population atlas and Yu et al. (2010) proposed a Parzen windows based hidden Markov random field algorithm, optimised by an expectation-maximisation algorithm, again augmented by a population atlas. Atlas-driven methods may have problems in patients with variable morphology or injury due to reliance on image registration for the propagation of the templates. We attempt to address this phenomena by combining the non-rigid registration of an atlas with the adaptive prior relaxation EM strategy described in Shiee et al. (2011) and Cardoso et al. (2011a).

Specifically in this study we seek to include those infants with structural abnormalities such as ventriculomegaly so that future studies of preterm brain development and injury can be more comprehensive, thus we combine a number of the steps described above to facilitate neonatal segmentation. Here we propose an adaptive segmentation pipeline specifically for preterm neonates based on Cardoso et al. (2011a) incorporating a novel maximum a posteriori expectation-maximisation (MAP-EM) based probabilistic segmentation technique that includes intensity non-uniformity (INU) correction, spatial dependence via a Markov random field (MRF) and implicit correction of PV containing voxels; we name this algorithm AdaPT.2 The algorithm specifically segments six classes representing the cortical grey matter, unmyelinated white matter, cerebrospinal fluid (CSF) space, cerebellum, deep grey matter (the myelinated white matter is included within this region) and brainstem/pons. Key features of this algorithm are: the development of a population specific T1 template for brain masking, subsequently combined with tissue class priors from the T2-weighted atlas developed by Kuklisova-Murgasova et al. (2011) and a combination of registration steps with the prior relaxation strategy of Shiee et al. (2011) that acts to adapt the population template for each subject. The algorithm includes not only adaptive priors on the tissue spatial distribution, but also priors over the expected intensities themselves, a novel feature applied here to the unique problem of neonatal tissue contrast. In addition we use a novel implicit partial volume correction strategy also specific for neonatal image contrast and provide a comparison with manual segmentation substantially beyond that found in Cardoso et al. (2011a). The described algorithm is evaluated in both normal and pathological subjects by comparison with manual segmentation of the six tissue classes and compared to a widely-used implementation of Van Leemput et al. (1999) incorporating an MRF, bias field correction and the same 4D anatomical neonatal priors.

Section snippets

Methods

The segmentation procedure combines a number of steps which are now outlined in detail: the brain extraction procedure, expectation maximisation framework (incorporating a novel prior on the expected tissue intensities), the spatial regularisation process, the prior relaxation strategy and finally the neonate-specific partial volume correction strategy. In order to simplify the segmentation process, the neonatal brain volumes must first be extracted from the full image. Due to the white/grey

Data

The data were acquired on a 1.5 T Siemens Avanto. Infants were sedated with an oral dose of chloral hydrate (Rosemont Pharmaceuticals, Leeds, UK) and imaged within a transparent MR-compatible pod. T1-weighted data were acquired with TR = 17 ms, TE = 6 ms and flip angle of 21°. In total 92 T1-weighted volumes are analysed with a resolution of 0.39 × 0.391 mm. The mean gestational age at birth was 27.0 ± 2.7 weeks (range 22.9–32.2 weeks), mean birth weight 966 ± 380 g (range 447–2470 g) and mean post-menstrual age

Discussion and conclusion

This work has developed a tool for the accurate segmentation of challenging neonatal MRI with specific emphasis on pathological cases. The method uses a MAP-EM algorithm with a prior relaxation strategy combined with a semi-conjugate prior over the intensities and an explicit PV model in order to mitigate the problem of misclassified PV voxels. Both these additions are important and complement each other; the prior relaxation lessens the spatial constraints whilst the priors over the parameters

Acknowledgments

M. Jorge Cardoso was supported by a scholarship from the Fundação para a Ciência e a Tecnologia, Portugal (scholarship number SFRH/BD/43894/2008). Andrew Melbourne was supported by the UK registered charity SPARKS (grant held by Neil Marlow). Giles Kendall is a National Institute for Health Research (NIHR) funded Clinical Lecturer. Nicola Robertson and Neil Marlow receive partial funding from the Department of Health NIHR Biomedical Research Centres funding scheme at the University College

References (36)

  • P.A. Yushkevich et al.

    User-guided 3D active contour segmentation of anatomical structures: significantly improved efficiency and reliability

    Neuroimage

    (2006)
  • M. Cardoso et al.

    Adaptive neonatal brain segmentation

  • M.J. Cardoso et al.

    Locally ranked STAPLE for template based segmentation propagation

    MICCAI workshop on Multi-Atlas Labeling and Statistical Fusion

    (2011)
  • J. Dubois et al.

    Primary cortical folding in the human newborn: an early marker of later functional development

    Brain

    (2008)
  • J.H. Gilmore et al.

    Regional gray matter growth, sexual dimorphism, and cerebral asymmetry in the neonatal brain

    J. Neurosci.

    (2007)
  • O. Kapellou et al.

    Abnormal cortical development after premature birth shown by altered allometric scaling of brain growth

    PLoS Med.

    (2006)
  • O. Khwaja et al.

    Pathogenesis of cerebral white matter injury of prematurity

    Arch. Dis. Child. Fetal Neonatal Ed.

    (2008)
  • A. Kitamoto et al.

    Image classification using probabilistic models that reflect the internal structure of mixels

    Pattern Anal. Appl.

    (1999)
  • Cited by (75)

    • Deep CNN ensembles and suggestive annotations for infant brain MRI segmentation

      2020, Computerized Medical Imaging and Graphics
      Citation Excerpt :

      This is particularly problematic in the case of infant brain segmentation, due to isointense contrasts and the high spatial variability of the infant population. To overcome the limitations of atlas-based methods, parametric (Cardoso et al., 2013; Ledig et al., 2012; Prastawa et al., 2005; Wu and Avants, 2012; Xue et al., 2007) or deformable (Wang et al., 2011) models can be used in a refinement step. Parametric models typically state segmentation as the optimization of an energy function, which integrates pixel/voxel probabilities from the atlas with priors restricting the shape or pairwise interactions of brain tissues (Cardoso et al., 2013; Ledig et al., 2012; Melbourne et al., 2012; Xue et al., 2007).

    • Imaging biomarkers in alzheimer’s disease

      2019, Handbook of Medical Image Computing and Computer Assisted Intervention
    View all citing articles on Scopus
    1

    The first two authors contributed equally to this work.

    View full text