Automatic fiber bundle segmentation in massive tractography datasets using a multi-subject bundle atlas
Graphical abstract
Highlights
► We propose an automatic and robust method for fiber bundle segmentation in massive tractography datasets. ► The method is based on a novel HARDI multi-subject human brain fiber bundle atlas, composed of 36 known deep white matter bundles. ► The atlas also contains 47 superficial white matter bundles in each hemisphere, included in a multisubject bundle atlas for the first time. ► The method considers the fiber shape, position and length information in the segmentation, leading to better results than ROI-based approaches. ► Results can be used for population studies where each generic bundle is analyzed separately.
Introduction
Human brain white matter (WM) structure and organization are not yet completely known. Diffusion-weighted magnetic resonance imaging (dMRI) offers a unique approach to study in vivo the structure of brain tissues through the measurement of the diffusion of water molecules (Basser and Pierpaoli, 1995, Le Bihan et al., 2001). The local fiber orientation distribution can be inferred from this data and fiber trajectories can be reconstructed using tractography algorithms (Mori and van Zijl, 2002).
The 3D curves stemming from tractography are commonly referred to as “fibers” though they do not represent individual fibers or axons. Instead, these curves represent an estimate of the trajectory of some large white matter fiber tracts. In this paper, we will use the terms “fiber”, “tract” or “fiber tract” to refer to the trajectories obtained with tractography algorithms. Consequently, a “fiber bundle” will be a bundle of tractography-based curves and not a real anatomical bundle of neural fibers.
Analysis of white matter organization from the results of tractography methods is a delicate task. The “spaghetti plate” made up by the tracts resulting from the current methods, indeed, is far from being a perfect representation of white matter structure. The poor spatial resolution of diffusion acquisitions puts strong limitations on the diameter of the bundles that can be mapped. Moreover, the difficulties raised by the numerous fiber crossings and white matter bottlenecks result in many spurious bundles. Nowadays, the recent dMRI techniques with high angular resolution (HARDI) have largely improved the quality of tractography relative to standard diffusion tensor imaging (DTI), but, despite this progress, it is not devoid of artifacts.
The resulting tractography datasets are highly complex and include millions of fibers which requires a new generation of analysis methods.
Unlike a simplistic bundle model, where known white matter tracts are represented by a relatively small number of fibers with the same shape, current tractography datasets reconstruct WM tracts represented by thousands of fibers, composed of various fiber fascicles of different shapes and lengths. Several examples of decomposition of major WM tracts were already proposed (Lawes et al., 2008, Catani et al., 2005). Also, these complex tractography datasets present a large amount of short association bundles, that have been rarely studied until now. The segmentation of WM fiber bundles is therefore a complex and not completely solved problem. Furthermore, clinical studies require the segmentation of white matter bundles in order to perform comparisons between populations. Our goal is to infer an atlas of the fiber bundles of the human brain and a method mapping this atlas to any new brain.
The segmentation of anatomical bundles requires the inclusion of anatomical information, in a more or less interactive manner, depending on the approach. The usual strategies proposed for the reconstruction of fiber bundles follow two complementary ideas. The first approach is based on regions of interest (ROI) used to select or exclude tracts. For the segmentation of new tractography data, the ROIs can be defined manually (Catani et al., 2002, Mori et al., 2005, Wakana et al., 2007, Catani and Thiebaut de Schotten, 2008), or using an ROI atlas after the application of affine (Oishi et al., 2008) or non-linear (Zhang et al., 2010) normalization. These automatic ROI-based approaches have shown to be very powerful but present a big dependence on the normalization quality. Furthermore, these methods do not use fiber shape to detect the bundles. The second strategy is based on tract clustering using pairwise similarity measures (Zhang et al., 2008, O'Donnell et al., 2006, Visser et al., 2011). This last approach is potentially less intensive in terms of user interaction and can also embed predefined knowledge represented by a bundle template (Maddah et al., 2005, O'Donnell and Westin, 2007). For example, O'Donnell and Westin (2007) created a “high-dimensional” WM atlas containing a representation of the known anatomical deep WM 3D tracts from ten different DTI tractography datasets, in an embedded space. The atlas was then used to automatically segment the most known 3D fiber bundles from five other subjects. Indeed, the anatomical information embedded in manually labeled clusters can be used as prior data for the clustering/classification of new tractography datasets (Wang et al., 2011). Other recent hybrid approaches extract the most known WM tracts by the combination of a priori information given by a GM/WM atlas and a fiber clustering based on a similarity measure (Wassermann et al., 2010, Li et al., 2010).
The fiber clustering approach has been successfully used to map the well-known fiber bundles of deep white matter (DWM). However, the clustering-based methods commonly present a limitation on the number of fibers that can be analyzed. The studies across a population of subjects are then limited in the number of tracts and the number of subjects that can be analyzed together. In spite of two recent works that describe the analysis of huge datasets (120,000 (Wang et al., 2011) and 480,000 fibers (Visser et al., 2011)), the segmentation of huge tractography datasets, presenting more than 1 million tracts, is still a challenge.
Furthermore, it would be very interesting to apply this kind of clustering to massive tractography datasets gathering a population of subjects in order to perform group analyses. This strategy could help to discover new reproducible bundles, in particular short association bundles. The limitation on the tractography dataset size could be one of the reasons why until now short fibers of superficial white matter (SWM) have been rarely considered. In fact, there does not exist detailed anatomical description in the literature for these bundles. Other issues that can also make more difficult this kind of study are the big inter-subject variability of these tracts as well as their important number. Only ROI-based approaches (Oishi et al., 2008, Zhang et al., 2010) have been used to study the structure of SWM. For example, (Zhang et al., 2010) used an atlas-based brain GM/WM segmentation, relying on non-linear normalization, to identify 29 short association bundles reproducible across subjects. All pairs of adjacent cortical regions were analyzed in order to find those that were connected by fibers in a population of 20 subjects.
As mentioned above, even though this kind of approach has been very powerful, it presents a big dependence on the normalization quality. Furthermore, no analysis was performed on the fibers' shape. The only condition used was the existence of fibers connecting both cortical regions (and not passing through deeper regions), which may lead to irregular and different bundles across subjects.
In this paper we present a method for the inference of a HARDI model of human brain WM bundles, based on massive tractography datasets, computed for a group of subjects. The model of the brain white matter organization is represented by an atlas made up of a set of generic fiber bundles present in most of the population. In order to overcome the limitation induced by the size and complexity of the tractography datasets, we propose a two-level strategy, chaining intra- and inter-subject fiber clustering. The first level, an intra-subject clustering, is composed of several steps performing a robust hierarchical clustering of a fiber tractography dataset that can deal with millions of diffusion-based tracts (Guevara et al., 2011a, Guevara et al., 2011b). The end result of this compression is a set of a few thousand homogeneous bundles representing the whole structure of the tractography dataset. The second level, an inter-subject clustering, gathers the bundles obtained in the first level for a population of subjects and performs a clustering after spatial normalization. It produces as output a model composed of a list of generic fiber bundles with consistent shape and localization in a normalized space, that can be detected in most of the population (Guevara et al., 2010). A validation with simulated datasets is designed in order to study the behavior of the inter-subject clustering over a population of subjects aligned only by affine registration. The whole method was applied to the tracts computed from HARDI data obtained for twelve adult brains. The inter-subject clusters have been manually labeled by an expert in anatomy in order to build the atlas. An atlas bundle can correspond to several inter-subject clusters to account for subdivisions of the underlying pathways often presenting large variability across subjects.
This multi-subject strategy, embedding the shape and localization variability, has been shown recently to be more efficient than the usual single template approach for brain structure recognition because of weaknesses of the spatial normalization paradigm (Lyu et al., 2010).
A novel HARDI multi-subject bundle atlas, representing the variability of the bundle shape and position across subjects was thus inferred. The atlas includes two main sets of bundles: LNAO-DWM12, consisting of 36 deep WM bundles, some of these representing a few subdivisions of known WM tracts, and LNAO-SWM12, consisting of 94 short association bundles of superficial WM. Finally, we propose an automatic segmentation method mapping this atlas to any new subject. New tractography datasets are first compressed with the same intra-subject clustering. The resulting clusters are then labeled using pairwise distances to the centroids representing the multi-subject atlas bundles. The segmentation method is tested in 20 subjects from another HARDI database and compared with a well-known ROI-based tractography segmentation method (Zhang et al., 2010).
Section snippets
Learning adult HARDI database (DB1)
The atlas inference was performed from twelve subjects of the NMR public database (Poupon et al., 2006). This database provides high quality T1-weighted images and diffusion data acquired with a GE Healthcare Signa 1.5 Tesla Excite II scanner. The diffusion data presents a high angular resolution (HARDI) based on 200 directions and a b-value of 3000 s/mm2 (voxel size of 1.875 × 1.875 × 2 mm).
Testing adult HARDI database (DB2)
Twenty subjects of another adult HARDI database, were used to test the segmentation method. This database
Results
A general problem for evaluating white matter bundle segmentation is the lack of a gold standard. This is even more complex for superficial white matter, which cartography is still largely unknown and to the best of our knowledge, only the shape of four SWM bundles has been described in the literature (Oishi et al., 2008). We evaluate our approach using the adult HARDI testing database (DB2): 20 adults for the segmentation of deep white matter bundles, and 10 adults for the segmentation of
Discussion and conclusion
Our results depend strongly on the quality of the tractography results: bundles that are not tracked in individuals cannot be segmented. Therefore, a future research program could be the use of the deep white matter atlas to add a priori knowledge in the tractography algorithm.
Nevertheless, the current method is already successful for the major tracts of deep and superficial WM. Thanks to the use of a novel multi-subject representation of bundles and shape information, the bundles are cleaner
Acknowledgments
NMR HARDI database is the property of CEA/I2BM/NeuroSpin and can be provided on demand to [email protected]. Data were acquired with NMR pulse sequences, reconstructed with NMR reconstructor package and postÂprocessed with BrainVisa/Connectomist software. The authors would like to thank Drs. Marion Leboyer and Josselin Houenou for providing the HARDI brain dataset DB2.
This project was supported in part by grants from Région Ile-de-France (France), the French Embassy in Chile, CONICYT Chile
References (43)
- et al.
A generic framework for parcellation of the cortical surface into gyri using geodesic Voronoï diagrams
Med. Image Anal.
(2003) - et al.
Virtual in vivo interactive dissection of white matter fasciculi in the human brain
NeuroImage
(2002) - et al.
A diffusion tensor imaging tractography atlas for virtual in vivo dissections
Cortex
(2008) - et al.
Multi-contrast large deformation diffeomorphic metric mapping for diffusion tensor imaging
NeuroImage
(2009) - et al.
Robust clustering of massive tractography datasets
NeuroImage
(2011) - et al.
Atlas-based segmentation of white matter tracts of the human brain using diffusion tensor tractography and comparison with classical dissection
NeuroImage
(2008) - et al.
A hybrid approach to automatic clustering of white matter fibers
NeuroImage
(2010) - et al.
Spectral-based automatic labeling and refining of human cortical sulcal curves using expert-provided examples
NeuroImage
(2010) - et al.
Tract-based morphometry for white matter group analysis
NeuroImage
(2009) - et al.
Atlas-based whole brain white matter analysis using large deformation diffeomorphic metric mapping: application to normal elderly and Alzheimer's disease participants
NeuroImage
(2009)
Human brain white matter atlas: identification and assignment of common anatomical structures in superficial white matter
NeuroImage
Tract-based spatial statistics: voxelwise analysis of multi-subject diffusion data
NeuroImage
Robust determination of the fibre orientation distribution in diffusion MRI: non-negativity constrained super-resolved spherical deconvolution
NeuroImage
Diffeomorphic demons: efficient non-parametric image registration
NeuroImage
Partition-based mass clustering of tractography streamlines
NeuroImage
Reproducibility of quantitative tractography methods applied to cerebral white matter
NeuroImage
Tractography segmentation using a hierarchical dirichlet processes mixture model
NeuroImage
Unsupervised white matter fiber clustering and tract probability map generation: applications of a Gaussian process framework for white matter fibers
NeuroImage
Structure-specific statistical mapping of white matter tracts
NeuroImage
Atlas-guided tract reconstruction for automated and comprehensive examination of the white matter anatomy
NeuroImage
Diffeomorphic brain registration under exhaustive sulcal constraints
IEEE Trans. Med. Imaging
Cited by (166)
Anat-SFSeg: Anatomically-guided superficial fiber segmentation with point-cloud deep learning
2024, Medical Image Analysis