Original contributionAn improved FSL-FIRST pipeline for subcortical gray matter segmentation to study abnormal brain anatomy using quantitative susceptibility mapping (QSM)
Introduction
Robust and accurate brain image segmentation is an essential prerequisite for a sensitive and unbiased identification of differences in MR image intensity or volume of brain structures between normal and pathologically affected brains. Clinical imaging studies encompassing large cohorts of patients and controls commonly apply automatic image segmentation tools. To date, several sophisticated software packages have been developed for segmenting subcortical gray matter (SGM) [1], [2], [3], [4], most of them relying on high-resolution T1-weighted (T1w) images. Among them, FMRIB's Integrated Registration and Segmentation Tool (FIRST) [3] of the FMRIB Software Library (FSL) is one of the most popular packages. FIRST has been applied in multiple brain imaging studies to investigate volume and shape changes of subcortical brain structures that may be associated with normal aging [5], [6] or neurodegenerative diseases, like Alzheimer's disease [7], [8], schizophrenia [9], multiple sclerosis [10], or epilepsy [11]. It has also been shown quite recently that FIRST performs similarly well regarding hippocampal segmentation to predict the progress in Alzheimer's Disease dementia when compared to other volumetric methods (e.g., Statistical Parametric Mapping) [12]. The scan-rescan reliability of SGM segmentation with FIRST on the same scanning platform [13] as well as between different scanning platforms [14] has been confirmed, which warrants the applicability of FIRST in large-scale longitudinal and multisite studies.
Brain atrophy is a hallmark of many neurological diseases, including stroke, traumatic brain injury or Alzheimer's disease [15], [16], [17], and has long been one of the primary targets of neuroimaging research. Specifically, it is correlated with disability in multiple sclerosis (MS) [18]. Consequently, if suboptimal and less accurate segmentation results occur with brain anatomies that show deviations from the anatomy of the training datasets, an analysis bias may consequently follow in clinical studies, which commonly compare patients with abnormal brains with matched controls with normal brains. This bias may not always be obvious and may even remain unidentified, especially in large-cohort studies. In this study, we therefore aim to improve the accuracy of FIRST-based segmentation of subjects with abnormal brain anatomy due to severe atrophy (see Fig. 1).
FIRST is a model-based subcortical brain segmentation tool, which models each subcortical structure as a surface mesh within a Bayesian framework, using shape and signal intensity information of manually segmented subcortical structures in 336 T1w datasets (including, but not limited to, normal brains and cases of schizophrenia and Alzheimer's diseases) as a prior [3]. The individual T1w images are linearly (affine) registered to the Montreal Neurological Institute (MNI) space using a two-stage registration with weighting of subcortical structures during the second stage. Subsequently, applying the inverse transformation brings the Bayesian model into the native space of the individual T1w images. Small inaccuracies resulting from the affine registration step are usually overcome by FIRST by applying sophisticated fine-tuning of the surface mesh. Because it may be difficult or even impossible to straighten out fully severe registration inaccuracies by the surface mesh optimization, there is consequently an intrinsic demand to transform as accurately as possible the individual subject data into the MNI space. While it is usually possible to transform brain images of healthy adults relatively accurately to the MNI space, affine registration of brain images of patients or individuals with substantial brain atrophy may produce residual misalignment in various brain structures (see Fig. 2). One reason for this is that the MNI template originates from a healthy adult population (age 18.5–43.5 years) [19], [20] and that affine registration is unable to deform and resolve local shape dissimilarities. Recently, Amann et al. [21] reported that linear registration to MNI space within the FIRST pipeline is prone to fail with T1w images of patients showing global brain atrophy, and suggested to replace the MNI template by a new template that includes advanced atrophy. However, while such a modification might result in a better matching of abnormal cases, it does not account for the need to affect FIRST's internal data modelling of the residual shape variance. The use of sub-group specific templates may also introduce further types of bias in the analysis of different subject groups. In addition, region-specific and/or pathology-related brain atrophy may occur [22], [23], [24], which cannot be covered by only a single template. Therefore, a more general approach is required to account for individual shape dissimilarities in cases of diseased brains.
Another important issue regarding FIRST-based SGM segmentation is the observation that, depending on T1w sequence parameters of the clinical imaging protocols, SGM nuclei may exhibit deviating or even poor contrast on T1w images compared to that in FIRST's training datasets, which may further affect the segmentation accuracy [25]. While the MNI template shows detailed delineation of SGM structures with excellent contrast, a similar contrast is desirable to have available with FIRST to improve inferior segmentation results in cases of insufficient contrast on T1w images. Against this background, we aim to enhance the contrast of T1w images to render their appearance more similar to the MNI template by combining T1w images with a second MRI dataset that displays SGM nuclei with distinct contrast. Quantitative susceptibility mapping (QSM) is indeed able to provide such a contrast [26], [27], [28], [29], [30], [31].
In this contribution, we thus propose and evaluate an improved FIRST segmentation pipeline that replaces the affine transformation to the MNI space by a nonlinear deformation registration to overcome individual brain tissue variations and utilizes composite images with increased SGM contrast created from T1w images and quantitative susceptibility maps.
Section snippets
Data acquisition
We retrospectively identified eighty-two subjects with particularly small brain volumes from our research image database at the Buffalo Neuroimaging Analysis Center with > 2000 subjects who had undergone both T1w imaging and QSM. All subjects had been involved in studies approved by the Institutional Review Board of the University at Buffalo. Informed written consent was available from each individual. We determined the global brain volume for all subjects in the database from the T1w images
HC weighting coefficients and similarity to MNI template
The calculated weighting coefficients for creating the HC images in the two training groups (n = 4 each) are summarized in Table 1. Mean values and standards deviations were w1 = 1.48 ± 0.04 and w2 = (− 89.11 ± 22.81) ppm− 1 for MS patients, and w1 = 1.57 ± 0.03 and w2 = (− 72.42 ± 25.57) ppm− 1 for normal control subjects. The mean coefficients for the two groups were used to calculate HC images for all remaining 74 subjects not belonging to one of the two training datasets.
Fig. 5 illustrates the contrast
Discussion
In our study, we demonstrated that the widely applied FSL FIRST pipeline tends to produce inaccurate subcortical segmentations (or may even fail) in individuals with abnormal brain anatomy. To address this issue, we propose a modification of the FIRST pipeline (HC-nlFIRST) by incorporating nonlinear registration and using a dedicated hybrid image contrast created by combining standard T1w images and quantitative susceptibility maps. By investigating Dice coefficients and FNRs with respect to
Conclusion
We presented an extension to the established FIRST pipeline, referred to as HC-nlFIRST, for automated subcortical segmentation that incorporates non-linear registration and dedicated image contrast combination of T1w images and quantitative susceptibility maps. This method overcomes insufficiencies of the linear registration in the default FIRST pipeline, particularly in the case of abnormal brain anatomy. HC-nlFIRST does not require any modifications of the underlying FIRST algorithm or its
Acknowledgements
The study was supported by the German Research Foundation (DFG, RE 1123/9-2, DE 2516/1-1), by a stipend (Landesgraduiertenstipendium) of the Graduate Academy of the Friedrich Schiller University Jena awarded to XF, a seed grant awarded to A.D. by the Interdisciplinary Center for Clinical Research (IZKF) in Jena, Germany, and seed grants awarded to F.S. by the International Society for Magnetic Resonance in Medicine (ISMRM) and the Friedrich Schiller University Jena. It was also supported by the
References (58)
- et al.
Whole brain segmentation: automated labeling of neuroanatomical structures in the human brain
Neuron
(2002) - et al.
A Bayesian model of shape and appearance for subcortical brain segmentation
Neuroimage
(2011) - et al.
Multi-atlas segmentation of subcortical brain structures via the AutoSeg software pipeline
Front Neuroinform
(2014) - et al.
Regional changes in thalamic shape and volume with increasing age
Neuroimage
(2012) - et al.
Abnormal subcortical deep-gray matter susceptibility-weighted imaging filtered phase measurements in patients with multiple sclerosis: a case-control study
Neuroimage
(2012) - et al.
Quantification of multiple-sclerosis-related brain atrophy in two heterogeneous MRI datasets using mixed-effects modeling
Neuroimage Clin
(2013) - et al.
A probabilistic atlas of the human brain: theory and rationale for its development
Neuroimage
(1995) - et al.
Unbiased average age-appropriate atlases for pediatric studies
Neuroimage
(2011) - et al.
Subcortical brain segmentation of two dimensional T1-weighted data sets with FMRIB's Integrated Registration and Segmentation Tool (FIRST)
NeuroImage Clin
(2015) - et al.
The measurement and clinical relevance of brain atrophy in multiple sclerosis
Lancet Neurol
(2006)