Multi-stage segmentation of white matter hyperintensity, cortical and lacunar infarcts
Introduction
Magnetic resonance imaging (MRI) has been widely used to detect a variety of cerebral abnormalities, such as white matter hyperintensity (WMH), cortical infarct (CI), and lacunar infarct (LI), that are of clinical importance. The WMH is thought to reflect small vessel cerebrovascular disease (Pantoni, 2002, Young et al., 2008), and may contribute to age-associated cognitive decline (Carmichael et al., 2010, He et al., 2010, Jokinen et al., 2009, Marquine et al., 2010, Ota et al., 2009, Vannorsdall et al., 2009), and increase the risk of dementia (Debette and Markus, 2010). CIs, as the name suggests, are infarcts in the cortical regions caused by cerebral artery occlusion most commonly by emboli, whereas LIs are subcortical infarcts due to the blockages of small penetrating arteries in the deep brain region with sizes up to approximately 15 mm (Ropper, 2005, Xavier et al., 2003). Previous studies suggest that both types of cerebral infarcts are associated with cognitive decline and increase the likelihood of dementia and stroke (Bennett et al., 2005, Carey et al., 2008, Jokinen et al., 2011, Jokinen et al., 2009, Schneider et al., 2004, Tripathi et al., 2011). Hence, it is of clinical importance to identify these cerebral abnormalities for potential early prevention, diagnosis, and treatment in cerebrovascular and neurodegenerative diseases.
Due to the complex mechanisms underlying cerebrovascular diseases, the appearance of the WMH, CI, and LI is heterogeneous in terms of their location, size, and image intensity on MRI. Up to now, visual inspection is still a common approach used to quantify the severity of these cerebral abnormalities. However, it is laborious and time consuming and therefore impractical in large-scale imaging studies. In addition, visual inspection is biased to raters and hence, highly dependent on inter and intra rater reliabilities (Kapeller et al., 2003, Prins et al., 2004, Vannorsdall et al., 2009), which in turn decrease the sensitivity in subsequent statistical analyses (Garrett et al., 2004, Van Straaten et al., 2006).
In recent years, major progress has been made on the development of semi- or fully automated segmentation for the WMH. Wen and Sachdev (2004) and Ramirez et al. (2011) introduced semi-automated approaches by choosing empirical thresholds based on the descriptive statistics of the image intensity and then manually modifying false WMH areas. Jack et al. (2001) and Gibson et al. (2010) developed fully automated approaches that first employ empirical thresholds before applying linear fitting or fuzzy clustering to segment the WMH. However, both approaches are based only on fluid attenuated inversion recovery (FLAIR) images that are less sensitive in the posterior fossa. In addition, they may overestimate the WMH due to its typical high intensity appearance in cortical areas, the septum, pellucidum, and flow artifacts in the 4th ventricle where a large percentage of the false positive WMH is detected. To partially address these issues, Gibson et al. (2010) further applied a white matter mask to remove this false positive WMH. More advanced methods have been developed based on Markov random field model (Schwarz et al., 2009), k-nearest neighbor (Anbeek et al., 2004, Wen et al., 2009), and neural classification (Dyrby et al., 2008), which require training images with the WMH labels. The segmentation accuracy of these methods relies on the representative training data that may be difficult to select due to the heterogeneous nature of the WMH. Different from these methods with a need of training samples, Admiraal-Behloul et al. (2005) proposed a fuzzy inference system to classify the WMH based on both anatomical locations and intensity values from the T2-weighted MRI and FLAIR. This approach is robust to a wide range of image intensities and contrasts. However, as it uses the prior masks of the intra-cranial, white matter, gray matter, and cerebrospinal fluid (CSF) in the Montreal Neurological Institute (MNI) brain template, segmentation accuracy is highly dependent on the alignment of individual subjects to the MNI template. Similar approaches using the above mentioned machine learning techniques have been proposed for the automated segmentation of multiple sclerosis lesions as well (Shiee et al., 2010, Warfield et al., 2000, Wu et al., 2006, Zijdenbos et al., 2002).
The CI and LI have thus far been manually identified by neuroradiologists in most of the existing studies (Bennett et al., 2005, Carey et al., 2008, Jokinen et al., 2011, Jokinen et al., 2009, Schneider et al., 2004, Tripathi et al., 2011). A few fully automated LI segmentation approaches have been proposed using the T1- and T2-weighted images (Uchiyama et al., 2007, Yokoyama et al., 2007). Uchiyama et al. (2007) first applied the top hat transform and then binarized the T2-weighted MR image for labeling LI voxels. Next, support vector machine classification was used to eliminate the LI false positives. Yokoyama et al. (2007) searched for LI candidates using a binarization approach at multiple threshold levels and then removed false positive LIs based on intensity thresholds and the shape of the LI. Since the size of the LI is relatively small and its appearance in the T1 and T2-weighted MRI is similar to CSF and the WMH, the method is limited to identifying LIs using only T1- and T2-weighted MR images. Sasaki et al. (2008) demonstrated that combining the FLAIR with the T1- and T2-weighted MR images increases the segmentation accuracy of the LI.
In this paper, we employ multi-modal MR images and present a multi-stage segmentation approach to automatically delineate the WMH, CI, and LI. Since the T1-, T2-weighted, and FLAIR images are commonly used in hospitals for determining cerebral abnormalities and have been recommended by previous studies (Debette and Markus, 2010, Jokinen et al., 2011, Jokinen et al., 2009, Kapeller et al., 2003, Sasaki et al., 2008, Van Straaten et al., 2006), our approach takes these three MRI modalities in order to increase the sensitivity and specificity of the abnormal white matter classification. Moreover, we developed our segmentation algorithm based on a series of simple image analysis operations, including Gaussian mixture models, region growing, and morphological operations, without the need of a training set. Hence, our method overcomes the difficulties in the selection of a representative training set, customarily faced by other existing approaches (Anbeek et al., 2004, Dyrby et al., 2008, Schwarz et al., 2009, Wen et al., 2009). Furthermore, our framework automatically labels the WMH, CI, and LI at the same time whereas existing methods often segment the WMH (Admiraal-Behloul et al., 2005, Anbeek et al., 2004, Dyrby et al., 2008, Gibson et al., 2010, Jack et al., 2001, Schwarz et al., 2009, Wen and Sachdev, 2004, Wen et al., 2009) or the LI alone (Uchiyama et al., 2007, Yokoyama et al., 2007). Finally, we evaluate the segmentation accuracy of the WMH, LI, and CI through comparison with manual labels and visual grading using a dataset of 272 old adults.
Section snippets
Methods
We now present a multi-stage segmentation technique (Fig. 1) based on T1-, T2-weighted, and fluid attenuated inversion recovery (FLAIR) magnetic resonance (MR) images. After correcting for intensity inhomogeneity, removing the brain skull, and aligning within-subject T2-weighted and FLAIR images to the corresponding T1-weighted image, the brain tissues (white matter, gray matter, CSF) and hyperintense regions are respectively identified using the T1-weighted MRI and FLAIR. We then further
Results
To evaluate the segmentation accuracy, we selected 272 subjects (131 males and 141 females; age: 70.7 ± 6.3 years; age range: 60 to 86 years) from the ongoing epidemiological aging cohort recruited by the Memory Aging & Cognition Center at the National University of Singapore. Every subject underwent MRI scans that were performed on a 3T Siemens Magnetom Trio Tim scanner using a 32-channel head coil at the Clinical Imaging Research Center of the National University of Singapore. The image protocols
Discussion
We present a multi-stage automated segmentation framework for delineating the white matter hyperintensity, cortical and lacunar infarcts from the T1-, T2-weighted, and FLAIR images. This segmentation algorithm contains several key components, including hyperintensity region initialization and refinement based on two-stage Gaussian mixture models, as well as the classification of brain abnormalities. To our knowledge, this is the first paper on the automated segmentation of the cortical infarct.
Acknowledgments
The work was supported by grants A*STAR SICS-09/1/1/001, a center grant from the National Medical Research Council (NMRC/CG/NUHS/2010), the Young Investigator Award at the National University of Singapore (NUSYIA FY10 P07), and the National University of Singapore MOE AcRF grants.
References (49)
- et al.
Fully automatic segmentation of white matter hyperintensities in MR images of the elderly
Neuroimage
(2005) - et al.
Probabilistic segmentation of white matter lesions in MR imaging
Neuroimage
(2004) - et al.
Segmentation of age-related white matter changes in a clinical multi-center study
Neuroimage
(2008) - et al.
Whole brain segmentation: automated labeling of neuroanatomical structures in the human brain
Neuron
(2002) - et al.
Improved optimization for the robust and accurate linear registration and motion correction of brain images
Neuroimage
(2002) - et al.
Brain segmentation and the generation of cortical surfaces
Neuroimage
(1999) - et al.
Lesion Explorer: a comprehensive segmentation and parcellation package to obtain regional volumetrics for subcortical hyperintensities and intracranial tissue
Neuroimage
(2011) - et al.
A topology-preserving approach to the segmentation of brain images with multiple sclerosis lesions
Neuroimage
(2010) - et al.
Automatic morphology-based brain segmentation (MBRASE) from MRI-T1 data
Neuroimage
(2000) - et al.
Adaptive, template moderated, spatially varying statistical classification
Med. Image Anal.
(2000)
The topography of white matter hyperintensities on brain MRI in healthy 60- to 64-year-old individuals
Neuroimage
Automated segmentation of multiple sclerosis lesion subtypes with multichannel MRI
Neuroimage
User-guided 3D active contour segmentation of anatomical structures: significantly improved efficiency and reliability
Neuroimage
Incidental subcortical lesions identified on magnetic resonance imaging in the elderly. II. Postmortem pathological correlations
Stroke
Mild cognitive impairment is related to Alzheimer disease pathology and cerebral infarctions
Neurology
Subcortical lacunes are associated with executive dysfunction in cognitively normal elderly
Stroke
Longitudinal changes in white matter disease and cognition in the first year of the Alzheimer disease neuroimaging initiative
Arch. Neurol.
The clinical importance of white matter hyperintensities on brain magnetic resonance imaging: systematic review and meta-analysis
BMJ (Clinical Research Ed.)
Maximum likelihood from incomplete data via the EM algorithm
J. Royal Stat. Soc. B
Measures of the amount of ecologic association between species
Ecology
Computer-mediated measurement and subjective ratings of white matter hyperintensities in vascular dementia: relationships to neuropsychological performance
Clin. Neuropsychol.
Automatic segmentation of white matter hyperintensities in the elderly using FLAIR images at 3T
J. Magn. Reson. Imaging
Brain structure and cerebrovascular risk in cognitively impaired patients: Shanghai Community Brain Health Initiative—pilot phase
Arch. Neurol.
FLAIR histogram segmentation for measurement of leukoaraiosis volume
J. Magn. Reson. Imaging
Cited by (54)
Subjective vs informant-reported cognitive complaints have differential clinical significance in covert cerebral small vessel disease
2023, Cerebral Circulation - Cognition and BehaviorComputer-aided extraction of select MRI markers of cerebral small vessel disease: A systematic review
2022, NeuroImageCitation Excerpt :We identified 2 published studies applying deep learning algorithms (Al-Masni et al., 2021; Ghafoorian et al., 2017), 2 using image processing techniques (Wang et al., 2012; Yokoyama et al., 2007), and a body of 6 studies by Uchiyama et al. combining image processing with machine learning methods (Uchiyama et al., 2015, 2009, 2012; Uchiyama et al., 2008, 2007a, b), in the extraction of lacunes and separate lacunes from other imaging biomarkers of CSVD. Study samples – While some studies did not specify details of included samples, others have used community-based cohorts (Wang et al., 2012), as well as patients with CSVD, transient ischaemic attack, and/or ischaemic/haemorrhagic stroke (Ghafoorian et al., 2017). The median sample size in studies of lacunes was 132 (range, 100–1075), but due to the low prevalence of lacunes in the general population, the number of lacunes is generally low in studies (usually < 100).
Automatic quantification of white matter hyperintensities on T2-weighted fluid attenuated inversion recovery magnetic resonance imaging
2022, Magnetic Resonance ImagingCitation Excerpt :Kawata et al. [20], introduced a region growing method for adaptive selection of segmentation by using a support vector machine (SVM) with image features extracted from initially identified WMH candidates. Fully automated algorithms [3,10–13,21–25] have the advantage of reducing operator bias, facilitating replicable delineation of WMH, and allowing for higher throughput processing of larger datasets. The Lesion Segmentation Tool Lesion Growing Algorithm (LGA) from the Statistical Parametric Mapping (SPM) toolbox in MATLAB [13] uses an a priori WMH map and multimodal data (T1-weighted and T2-weighted FLAIR images), and grows the initial WMH map along adjacent hyperintense voxels.
Comparison of domain adaptation techniques for white matter hyperintensity segmentation in brain MR images
2021, Medical Image AnalysisCitation Excerpt :Various methods using hand-crafted features have been used for WMH segmentation (Caligiuri et al., 2015), and in recent years, deep learning (DL) models are being increasingly used and have been shown to outperform traditional methods (Rachmadi et al., 2018; Kuijf et al., 2019). Many of the existing methods (using either hand-crafted features or DL models) were trained with a large amount of manual labels (Wang et al., 2012; Admiraal-Behloul et al., 2005; Ghafoorian et al., 2016) and/or evaluated on specific population group (Wang et al., 2012; Gibson et al., 2010; De Boer et al., 2009; Steenwijk et al., 2013; Jeon et al., 2011; Hong et al., 2020; Park et al., 2018), acquired with the same scanner/protocol or validated on isotropic or axial acquisition images (Ghafoorian et al., 2017a; Kuijf et al., 2019). However, in the real-world scenario, most of the clinical datasets are small in size, acquired using various protocols and scanners, and from people with diverse demographic and pathological characteristics.
Triplanar ensemble U-Net model for white matter hyperintensities segmentation on MR images
2021, Medical Image Analysis