Lesion Explorer: A comprehensive segmentation and parcellation package to obtain regional volumetrics for subcortical hyperintensities and intracranial tissue
Research highlights
►Lesion Explorer segmentation and parcellation of structural MRI show high reliability. ►LE allows for segmentation of periventricular, deep white, VR, and lacunar subtypes. ►LE provides comprehensive volumetric profiles to examine subcortical vasculopathy.
Introduction
Subcortical hyperintensities (SH) are frequently observed on T2-weighted MRI of the aging brain (Jack et al., 2001, Kertesz et al., 1988, Longstreth et al., 1996). Clinico-pathological correlations suggest vascular and degenerative origins including: ischemic tissue damage via arteriosclerosis (Babikian and Ropper, 1987, van Swieten et al., 1991); vasogenic edema from periventricular venous collagenosis (Black et al., 2009, Gao et al., 2008, Moody et al., 1995); multiple lacunar infarcts (Longstreth et al., 1996); état criblé or dilated perivascular spaces (Awad et al., 1986); demyelination and subependymal gliosis; amyloid angiopathy (Pantoni and Garcia, 1997); and clasmatodendrosis (Sahlas et al., 2002).
Although the pathophysiological origins are not fully understood, the current literature suggests that SH: are common after age 60 (Longstreth et al., 1996); share common cerebrovascular risk factors such as diabetes and hypertension (De Leeuw et al., 2001, Liao et al., 1996, Sachdev et al., 2008); and are associated with increased risk of cognitive decline, stroke, gait disorders and neuropsychiatric disorders (De Groot et al., 2001, Koga et al., 2009, Longstreth et al., 1996, Longstreth et al., 2005, Srikanth et al., 2009, Vermeer et al., 2003a, Vermeer et al., 2003b). To further assess the effects of vascular risk factors in overt and covert cerebrovascular disease and in dementia, consensus criteria were developed that underline the importance of accounting for SH in studies on aging (Hachinski et al., 2006).
Although visual scales can provide quick ratings of SH severity on MRI (Bocti et al., 2005, Fazekas et al., 1987, Scheltens et al., 1993, Wahlund et al., 2001), inconsistencies in methodological properties (Mantyla et al., 1997) have led some researchers to apply intensity-based segmentation techniques that provide a more accurate estimation of SH burden — as well as quantify their extent and location. However, typical T1-based tissue segmentation techniques that do not explicitly segment T2 hyperintensities can inflate other tissue volumes. Depending on the segmentation technique used, gray matter volumes can be overestimated by failing to segment the hyperintensities (Levy-Cooperman et al., 2008), see Fig. 7 for example.
Quantitative segmentation approaches have been applied to capture SH on T2, proton density (PD) and fluid-attenuated inversion recovery (FLAIR) images. Some of the recent approaches in the last decade include: fuzzy clustering models that include a lesion class (Admiraal-Behloul et al., 2005, Gosche et al., 1999); Gaussian curve fitting to determine lesion intensity cut-off points (Decarli et al., 2005a); modal intensity cut-offs applied to slice-by-slice intensity histograms (Jack et al., 2001); k-Nearest Neighbor (kNN) algorithmic combination approaches (Anbeek et al., 2004, Seghier et al., 2008, Swartz et al., 2002, Wen et al., 2008); and coregistration to normal templates comparing the voxel-wise SH probabilities from FLAIR images to a white matter probability map using a weighting function (Burton et al., 2004, Wen and Sachdev, 2004). These approaches range from fully automated to semi-automated labor intensive processes.
Fully automated techniques offer the advantage of high reliability and are preferable for processing large scale studies. They typically require FLAIR imaging, where SH are often more conspicuous relative to standard PD/T2 images. However, FLAIR imaging is less sensitive in detecting focal thalamic lesions (Bastos Leite et al., 2004, Jack et al., 2001) and was not included in the multi-center Alzheimer's Disease Neuroimaging Initiative (ADNI) acquisitions (Jack et al., 2008). Furthermore, FLAIR imaging alone cannot disambiguate the possible heterogeneity that is implicated in SH pathology.
In an attempt to address the hypothesis of pathological heterogeneity within SH, various subtypes have been suggested for further segmentation. One common, albeit controversial, distinction is between periventricular (pvSH) and deep white (dwSH) subcortical hyperintensities (Decarli et al., 2005a, Sachdev and Wen, 2005). Although dwSH and pvSH volumes are correlated (Decarli et al., 2005a), some studies have shown pvSH and dwSH to be differentially associated with: gray matter atrophy; ventricular dilatation; and cognitive, behavioral and motor/gait performance (Sachdev and Wen, 2005, Sachdev et al., 2005, Silbert et al., 2008).
Lack of a standardized methodology for the definition of pvSH may be the cause of inconsistent reports in the literature. The typical method to distinguish pvSH from dwSH is to create an arbitrary two-dimensional cut-off line lateral to the ventricles on axial slices in a slice-by-slice manner. This arbitrary line is sometimes calculated as a proportional distance from the ventricular border to the dura mater (Decarli et al., 2005a), or set using an arbitrary voxel distance from the ventricle out into the centrum semiovale. Some reviewers have suggested that there may be some neuroanatomic justification for classifying SH within a 13 mm rim around the ventricle as pvSH (Mayer and Kier, 1991, Sachdev et al., 2008). Various other methods arbitrarily delineate pvSH from dwSH using a linear distance calculation. However, a standardized, unbiased method that recognizes the 3-dimensional nature of SH would be preferable.
Other subtypes of SH include perivascular (Virchow–Robin) spaces and cystic fluid-filled lacunar-like infarcts. Virchow–Robin (VR) spaces are CSF-filled extensions of the subarachnoid space in the sheath surrounding blood vessels. They appear as hyperintense dots or lines on T2 images, are isointense on PD and typically 1 mm or less in diameter (Awad et al., 1986). Their size, shape, and differential appearance on T2 and PD allow them to be distinguished from other SH subtypes — including lacunes.
Automatic segmentation of lacunes is less common since this requires coregistration of T2/PD/FLAIR images to a T1-based segmentation, in order to identify CSF intensity within SH. Lacunes are associated with aging, hypertension, increased risk of stroke, and are found in 11–28% of elderly (Longstreth et al., 1998, Vermeer et al., 2003b, Vermeer et al., 2007). These so-called silent strokes or covert infarcts, are usually defined as 3–15 mm in diameter, are hypointense on T1, and hyperintense on both T2 and PD images. Their presence is associated with increased risk of dementia and have been correlated with decreased frontal lobe glucose metabolism with positron emission tomography (PET) imaging (Reed et al., 2004). However, without a coregistered T1-segmentation and PD-T2 contrast for comparison, lacunes and VR spaces are difficult to quantify through volumetric segmentation with FLAIR alone.
An additional benefit of a T1-based tissue segmentation combined with PD-T2-based SH segmentation is that it allows for relative volumetric tissue comparisons for gray matter (GM), white matter (WM), ventricular-CSF (vCSF) and sulcal-CSF (sCSF). However, whole brain global volumetrics alone provide limited information, and regionalized quantification, whether ROI-based or template-based, has become a standard expectation for any MRI-based segmentation procedure.
Hence, the need for a comprehensive, individualized MRI processing pipeline that reliably segments the brain into regionalized tissue compartments and includes the various SH subtypes has become increasingly important. Lesion Explorer (LE) is the final component of an MRI-based processing pipeline that was developed with these considerations. It was built upon updated versions of two previously published components: an automated T1-based tissue segmentation protocol (Kovacevic et al., 2002); and the Semi-Automated Brain Region Extraction (SABRE) parcellation procedure (Dade et al., 2004). The LE pipeline makes use of 3 processing components that effectively allow comprehensive analysis of individual brains through the segmentation of 8 tissue classes: GM, WM, sCSF, vCSF, lacunar and non-lacunar pvSH and dwSH — tissue volumes are then parcellated into 26 SABRE brain regions. Inter-rater and inter-method reliability data is presented with validation against an alternative kNN segmentation approach and 2 different visual rating scales.
Section snippets
Subjects
Images used for this study (n = 20) were selected from participants in the Sunnybrook Dementia Study — a large ongoing longitudinal observational study conducted in the LC Campbell Cognitive Neurology Research Unit and the Heart & Stroke Foundation Centre for Stroke Recovery (http://www.heartandstroke-centrestrokerecovery.ca) at Sunnybrook Health Sciences Centre, a University of Toronto academic healthcare institution. See Table 1 for additional details. Exclusion criteria for this sub-study
Whole brain results
Whole brain mean absolute volume differences between the two raters was 230 mm3 with the following results: ICC = .99, p < .0001, and mean SI = .97, indicating excellent inter-rater reliability for whole brain volumetric and pixel-wise spatial agreement (see Table 1 for raw data). Compared to volumes from a previously published semi-automated segmentation using a kNN algorithm, high reliability was also demonstrated (ICC = .97, p < .0001). Whole brain segmentation of pvSH volumes also yielded high
Discussion
Lesion Explorer is the final component of a comprehensive segmentation and parcellation package that provides an individualized volumetric profile from standard structural MRI. The overall MRI volumetrics package is a reliable application that may be used with confidence in aging populations for both cross-sectional, and longitudinal studies with a standard structural acquisition protocol.
The brain-extraction component, Brain-Sizer, provides an accurate measure of an individual's total
Acknowledgments
The authors gratefully acknowledge financial support from the following sources. The development and testing of this image analysis pipeline was supported by several grants, most notably from the Canadian Institutes of Health Research (MT#13129), the Alzheimer Society of Canada and Alzheimer Association (US), the Heart and Stroke Foundation Centre for Stroke Recovery, and the LC Campbell Foundation. JR receives salary support from the Alzheimer Society of Canada; BL from CIHR; NLC from U of T
References (63)
- et al.
Fully automatic segmentation of white matter hyperintensities in MR images of the elderly
Neuroimage
(2005) - et al.
Automatic segmentation of different-sized white matter lesions by voxel probability estimation
Med. Image Anal.
(2004) - et al.
Semi-automatic brain region extraction (SABRE) reveals superior cortical and deep gray matter atrophy in MS
Neuroimage
(2006) - et al.
Semiautomatic brain region extraction: a method of parcellating brain regions from structural magnetic resonance images1
Neuroimage
(2004) - et al.
Measures of brain morphology and infarction in the Framingham Heart Study: establishing what is normal
Neurobiol. Aging
(2005) - et al.
- et al.
Cognitive consequences of multiple lacunes and leukoaraiosis as vascular cognitive impairment in community-dwelling elderly individuals
J. Stroke Cerebrovasc. Dis.
(2009) - et al.
A robust method for extraction and automatic segmentation of brain images
Neuroimage
(2002) - et al.
A semiquantative rating scale for the assessment of signal hyperintensities on magnetic resonance imaging
J. Neurol. Sci.
(1993) - et al.
Lesion identification using unified segmentation–normalisation models and fuzzy clustering
Neuroimage
(2008)
Utility of simultaneous brain, CSF and hyperintensity quantification in dementia
Psychiatry Res.
Silent brain infarcts: a systematic review
Lancet Neurol.
Bayesian analysis of neuroimaging data in FSL
Neuroimage
User-guided 3D active contour segmentation of anatomical structures: significantly improved efficiency and reliability
Neuroimage
Diagnostic and Statistical Manual of Mental Disorders
Incidental subcortical lesions identified on magnetic resonance imaging in the elderly. II. Postmortem pathological correlations
Stroke
Binswanger's disease: a review 1
Stroke
Thalamic lesions in vascular dementia: low sensitivity of fluid-attenuated inversion recovery (FLAIR) imaging
Stroke
Medical image analysis with fuzzy models
Stat. Methods Med. Res.
Understanding white matter disease: imaging–pathological correlations in vascular cognitive impairment
Stroke
A new visual rating scale to assess strategic white matter hyperintensities within cholinergic pathways in dementia
Stroke
White matter hyperintensities are associated with impairment of memory, attention, and global cognitive performance in older stroke patients
Stroke
Comparison of manual and semi-automated delineation of regions of interest for radioligand PET imaging analysis
BMC Nucl. Med.
Overlap in frontotemporal atrophy between normal aging and patients with frontotemporal dementias
Alzheimer Dis. Assoc. Disord.
Magnetic resonance imaging in frontotemporal dementia shows subcortical atrophy
Dement. Geriatr. Cogn. Disord.
Cerebral white matter lesions and subjective cognitive dysfunction: the Rotterdam Scan Study
Neurology
Prevalence of cerebral white matter lesions in elderly people: a population based magnetic resonance imaging study. The Rotterdam Scan Study
J. Neurol. Neurosurg. Psychiatry
Anatomical mapping of white matter hyperintensities (WMH): exploring the relationships between periventricular WMH, deep WMH, and total WMH burden
Stroke
MR signal abnormalities at 1.5 T in Alzheimer's dementia and normal aging 6
AJR Am. J. Roentgenol.
Automated quantification of brain magnetic resonance image hyperintensities using hybrid clustering and knowledge-based methods
Int. J. Imag. Sci. Technol.
National Institute of Neurological Disorders and Stroke–Canadian Stroke Network vascular cognitive impairment harmonization standards
Stroke
Cited by (103)
Vascular Burden Moderates the Relationship Between ADHD and Cognition in Older Adults
2024, American Journal of Geriatric PsychiatrySecondary thalamic atrophy related to brain infarction may contribute to post-stroke cognitive impairment
2023, Journal of Stroke and Cerebrovascular DiseasesImaging perivascular space structure and function using brain MRI
2022, NeuroImageCitation Excerpt :They used a threshold-based semiautomatic method, originally developed for the quantification of brain lesions (Makale et al., 2002), where connectivity and threshold values were adjusted based on manually determined PVS (Wuerfel et al., 2008). Similar threshold-based approaches were adopted by Ramirez et al. (2011), who segmented PVS via a modified lesion extraction algorithm employing T1-weighted, T2-weighted, and proton-density (PD) weighted images, and by Wang et al. (2016), who were able to segment PVS on one single slice in centrum semiovale and in a manually-traced ovoid region of the basal ganglia after performing a linear intensity adaptive adjustment in 3 stages (normalization, gamma correction, and linear mapping). Once the PVS have been segmented, several metrics and morphological features of PVS can be computed from the segmentation masks (Fig. 1).
Neuropsychiatric symptoms as a sign of small vessel disease progression in cognitive impairment
2022, Cerebral Circulation - Cognition and Behavior