White matter lesion extension to automatic brain tissue segmentation on MRI
Introduction
Brain tissue segmentation on structural magnetic resonance imaging (MRI) has received considerable attention. Quantitative analysis of MR images of the brain is of interest in order to study the aging brain in epidemiological studies, to better understand how diseases affect the brain and to support diagnosis in clinical practice. Manual quantitative analysis of brain imaging data is a tedious and time-consuming procedure, prone to observer variability. Therefore, there is a large interest in automatic analysis of MR brain imaging data, especially segmentation of cerebrospinal fluid (CSF), gray matter (GM) and white matter (WM). In the last decade several automatic brain tissue segmentation methods have been proposed, often based on T1-, T2- or proton density-weighted MR images. Some use a fixed set of labeled samples, that were derived from manual segmentations, to train the classifier (Amato et al., 2003, Anbeek et al., 2005). This has, however, some disadvantages as it is dependent on the MRI sequence, requires a laborious training stage and is limited to the MRI intensity variations captured in the training set. Therefore, some studies developed methods to obtain subject-specific training samples labeled by, for example, clustering (Harris et al., 1999, Barra and Boire, 2000), Gaussian mixture models (Ruan et al., 2000, Lemieux et al., 2003) or atlas registration (Cocosco et al., 2003, Song et al., 2006). These methods are independent of intersubject intensity variations and MRI sequence. Another option, often used nowadays, is updating both the classification and the model parameters in an iterative process (Van Leemput et al., 1999, Zhang et al., 2001, Kovacevic et al., 2002, Ashburner and Friston, 2005, Ruf et al., 2005, Awate et al., 2006). This type of method is more complicated but it is also independent of intersubject intensity variations and MRI sequence. Several of these studies have evaluated their method on a reasonably large group varying between 34 to 71 subjects (Harris et al., 1999, Kovacevic et al., 2002, Cocosco et al., 2003, Lemieux et al., 2003, Song et al., 2006, Vrooman et al., 2007). The method developed by Zhang et al. (2001) is incorporated as FSL's brain tissue segmentation method, FAST, and is used in multiple studies, similar to the SPM brain tissue segmentation method by Ashburner and Friston (2005).
Besides automatic brain tissue segmentation, automatic WML segmentation has also received considerable interest. White matter lesions (WML) are commonly found in elderly subjects and are associated with cognitive decline (de Groot et al., 2002) and increased risk of stroke (Vermeer et al., 2003) and dementia (Prins et al., 2004). Recent studies often use T2-weighted or fluid-attenuated inversion recovery (FLAIR) scans in which white matter lesions are hyperintense. Several automatic segmentation methods have been developed based on intensity alone (Jack et al., 2001, Admiraal-Behloul et al., 2005, DeCarli et al., 2005) or including also spatial- (Anbeek et al., 2004), texture- (Kruggel et al., 2008) or shape-information (Alfano et al., 2000). It is difficult to compare the reported accuracies of these WML segmentation methods. Often different evaluation measures are used and some of these measures depend on the WML load of the subject (Admiraal-Behloul et al., 2005). Furthermore, automatic WML segmentations are often evaluated by comparison to manual segmentations and the evaluation is therefore influenced by the manual segmentation protocol. The robustness of an automatic segmentation method can be demonstrated by applying the method to a large dataset. Only some studies evaluated their WML segmentation method on datasets of 100 or more subjects (Admiraal-Behloul et al., 2005, Kruggel et al., 2008, Maillard et al., 2008).
In this paper a fully automated method for CSF, GM and WM segmentation based on multimodal MRI data is optimized and extended with WML segmentation. The contribution of this paper to the existing literature is threefold. Firstly, we evaluate different atlas registration methods for a brain tissue segmentation method presented by Cocosco et al. (2003) and Vrooman et al. (2007) where atlas registration is used to automatically train a k-nearest neighbor classifier. Different types of registration are compared: single- versus multiple-atlas registration; affine versus B-spline based non-rigid registration at different control point spacings; and registration of a varying number of atlases. Secondly, the method is extended with an automatic WML segmentation. This segmentation method uses the GM classification to determine a white matter lesion intensity threshold value in the FLAIR scan. Thirdly, the method is qualitatively validated on a large dataset of 209 elderly subjects. A quantitative evaluation is performed on a small subset using manual segmentations.
Section snippets
Atlas data
Twelve atlases have been obtained by manual segmentation of scans from the Rotterdam Scan Study (de Leeuw et al., 2001), that were acquired in 1995–1996. This population-based imaging study is aimed at investigating determinants of age-related neurologic diseases among elderly persons. The twelve subjects were female and had a mean age (± standard deviation) of 64 (± 1.8) years. MR brain imaging was performed on a 1.5 T Siemens scanner using a quadrature head coil. An inversion recovery double
Optimization of α and β
The two WML segmentation parameters α and β were optimized on six of the 20 subjects who had only a manual WML segmentation. These subjects were not used in any of the other experiments. For these six subjects WML segmentations were obtained by the automatic method with non-rigid registration of 11 atlases with 2.5 mm control point spacing and a range of α from 2.0 to 3.1 at intervals of 0.1 and of β from 0.02 to 0.30 at intervals of 0.02. For every subject the SI between the automatic and the
Comparison of different types of atlas registration
The results for the accuracy study, as a function of the atlas registration method used in training, are shown in Fig. 3. The given SIs are averages of twelve SIs obtained by comparing the automatic segmentations of the six subjects with the manual segmentations by two observers. Fig. 4 shows the average SIs of the experiment using different numbers of atlases. The SIs obtained by registering 12 atlases are given as comparison. Non-rigid registration using 12 atlases and a control point spacing
Discussion and conclusion
A fully automated method for CSF, GM and WM segmentation has been optimized, extended with WML segmentation, and quantitatively and qualitatively validated. The different brain tissues, CSF, GM and WM, are segmented by an automatically trained kNN classifier using atlas registration. The quantitative evaluation comparing the automatic segmentations to manual segmentations showed similarity indices close to the interobserver similarity index. The quantitative evaluation was performed primarily
Acknowledgments
The Rotterdam Study is supported by the Erasmus Medical Center and Erasmus University Rotterdam, the Netherlands Organization for Scientific Research (NWO), the Netherlands Organization for Health Research and Development (ZonMw), The Research Institute for Diseases in the Elderly (RIDE), the Ministry of Education, Culture and Science, the Ministry of Health, Welfare and Sports, the European Commission (DG XII) and the Municipality of Rotterdam. This study was further financially supported by
References (35)
- et al.
Fully automatic segmentation of white matter hyperintensities in MR images of the elderly
Neuroimage
(2005) - et al.
Segmentation of magnetic resonance brain images through discriminant analysis
J. Neurosci. Methods
(2003) - et al.
Automatic segmentation of different-sized white matter lesions by voxel probability estimation
Med. Image Anal.
(2004) - et al.
Probabilistic segmentation of brain tissue in MR imaging
Neuroimage
(2005) - et al.
Unified segmentation
Neuroimage
(2005) - et al.
Adaptive Markov modeling for mutual-information-based, unsupervised MRI brain-tissue classification
Med. Image Anal.
(2006) - et al.
A fully automatic and robust brain MRI tissue classification method
Med. Image Anal.
(2003) - et al.
Measures of brain morphology and infarction in the Framingham Heart Study: establishing what is normal
Neurobiol. Aging
(2005) - et al.
Brain tissue volumes in the general elderly population. The Rotterdam Scan Study
Neurobiol. Aging
(2008) - et al.
A robust method for extraction and automatic segmentation of brain images
Neuroimage
(2002)
Texture-based segmentation of diffuse lesions of the brain's white matter
Neuroimage
Tract-based spatial statistics: voxelwise analysis of multi-subject diffusion data
Neuroimage
Multi-spectral brain tissue segmentation using automatically trained k-Nearest-Neighbor classification
Neuroimage
Automated segmentation and measurement of global white matter lesion volume in patients with multiple sclerosis
J. Magn. Reson. Imaging
Tissue segmentation on MR images of the brain by possibilistic clustering on a 3D wavelet representation
J. Magn. Reson. Imaging
Periventricular cerebral white matter lesions predict rate of cognitive decline
Ann. Neurol.
Prevalence of cerebral white matter lesions in elderly people: a population based magnetic resonance imaging study. The Rotterdam Scan Study
J. Neurol. Neurosurg. Psychiatry
Cited by (254)
Small vessel disease burden and functional brain connectivity in mild cognitive impairment
2024, Cerebral Circulation - Cognition and BehaviorWhere is VALDO? VAscular Lesions Detection and segmentatiOn challenge at MICCAI 2021
2024, Medical Image AnalysisSocial Health Is Associated With Tract-Specific Brain White Matter Microstructure in Community-Dwelling Older Adults
2023, Biological Psychiatry Global Open ScienceThe role of cognitive and brain reserve in late-life depressive events: The Rotterdam Study
2023, Journal of Affective DisordersAssociations of neuroimaging markers with depressive symptoms over time in middle-aged and elderly persons
2023, Psychological Medicine