A hybrid approach to automatic clustering of white matter fibers
Introduction
Diffusion tensor imaging (DTI) allows in vivo measurement of the diffusivity of water molecules in living tissues (Le Bihan et al., 2001, Basser et al., 1994, Basser and Jones, 2002). Although the diffusivity of water molecules is generally represented as a Brownian motion, the microstructure of living tissues imposes certain constraints on this motion, which results in an anisotropic diffusion measured by DTI (Le Bihan et al., 2001, Basser et al., 1994). The measured diffusion can be approximated by an anisotropic Gaussian model, which is parameterized by the diffusion tensor in each voxel (Basser et al., 1994) to create the tensor field. Diffusion tensor measure provides a rich data set from which a measurement of diffusion anisotropy can be obtained through the application of mathematical formulas and calculation of the underlying Eigenvalues (Moseley et al., 1990; Le Bihan et al., 2001; Basser and Jones, 2002). The recent review article Mori and Zhang (2006) provided an excellent tutorial on the principles of DTI and its applications to neuroscience.
It is widely believed that DTI provides insights into the nature and degree of white matter injury that occurs in neurological diseases and sheds light on early detection and diagnosis of devastating neurological diseases. DTI has been widely used in the investigation of WM abnormality associated with various progressive neuropathologies (Werring et al., 1999, Bozzali et al., 2002, Horsfield and Jones, 2002, Moseley, 2002, Stahl et al., 2003, Sundgren et al., 2004, Park et al., 2004, Schocke et al., 2004, Eluvathingal et al., 2006), since it yields quantitative measures reflecting the integrity of WM fiber tracts. These DTI studies on WM are very useful in the investigation of the abnormality that occurs on fiber pathways connecting remote computation centers of various gray matter (GM) regions. From computational quantitation perspective, most of previous clinical WM DTI studies were based on region of interests (ROI) analysis (Alexander and Lee, 2007, Chang et al., 2007) or voxel-based morphometry (VBM) analysis (Barnea-Goraly et al., 2004). Notably, ROI-based methods are time-consuming and their reproducibility is limited. VBM-based methods add uncertainty into the analysis since they need non-linear warping of the tensor field and re-orientation of the tensors (Ruiz-Alzola et al., 2000, Alexander et al., 2001.)
Recently, a new methodology called tract-based analysis has been investigated by a variety of research groups (Shimony et al., 2002, Fillard and Gerig, 2003, Gerig et al., 2004, Brun et al., 2004, O'Donnell and Westin, 2005, Smith et al., 2006, Maddah et al., 2008). The basic idea of this methodology is to cluster the WM fibers into anatomically meaningful tracts or bundles, and then perform quantitative measurements on these clustered fiber tracts. The major advantages of this methodology include its better biological meanings and facilitation of tract-based comparisons across different subject groups. The tract-based analysis of WM fibers has raised interests from the neurology and clinical neuroscience community, e.g., Goldberg-Zimring et al., 2005, since this methodology provides direct quantification of the properties of the specific WM bundles rather than the individual image voxels or the entire human brain.
However, clustering WM fibers into meaningful bundles is nontrivial due to the following reasons. First, although the conceptual definition of meaningful WM bundles is quite clear in neuroanatomy (Mori et al., 2005), their accurate quantitative definitions are largely unknown, e.g., where the boundaries of different WM bundles are. Second, the human brain architecture and its connectivity pattern is quite complex. In typical DTI tractography results, it is quite common that there are many thousands of tracked WM fibers. How to automatically assign meaningful anatomic labels to them is quite challenging. Finally, the white matter architecture of the human brain across individuals is considerably variable in terms of its geometric properties and connectivity strengths and patterns. Development of a WM fiber clustering method that could reliably and robustly work on different human brain DTI data is not easy. Due to above reasons, many existing WM fiber clustering methods need manual guidance in the clustering procedure, so that expert neuroscience knowledge could be integrated into the decision-making process of WM fiber clustering. A couple of approaches have been proposed to incorporate prior knowledge, e.g., Xia et al. (2005) and O'Donnell and Westin (2007). One popular methodology to incorporate neuroscience prior knowledge to WM fiber clustering is to use an atlas of fiber tracts and perform atlas-based warping, e.g., Maddah et al. (2005) and O'Donnell and Westin (2007). For this methodology, it needs to build up a WM atlas first, and then perform non-linear registration of DTI data across different individual brains (Yang et al., 2008). Another approach is to use anatomically defined gray matter regions to guide the clustering of WM fibers (Xia et al., 2005). In this method, connections between anatomically delineated brain regions are identified by first clustering fibers based on their terminations in anatomically defined gray matter regions, and then these connections are refined based on geometric similarity criteria.
In this paper, we propose a new computational framework to automatically cluster whole brain WM fibers into biologically meaningful neuro-tracts. This framework explicitly divides 19 major WM fiber bundles into two groups based on neuroanatomical knowledge and our experimental observations. The proposed framework is a hybrid approach, that is, the bundles in the first group are consecutively clustered via top-down brain anatomy guidance, while the bundles in the second group are clustered via a similarity-based bottom-up clustering method. The major advantages of this framework are its intuitiveness, effectiveness, and biological soundness. Our experimental results demonstrate that this hybrid approach can deal with the complexity and variability of white matter architecture of human brain.
Section snippets
Overview and pre-processing
A schematic diagram of our computational framework for fiber clustering and tract-based fiber analysis is illustrated in Fig. 1a. Totally, there are six steps in this computational framework. This framework uses FSL FDT for eddy current correction and uses DTIStudio for tensor calculation and channel image generation (Fig. 1b). It then employs the multi-channel DTI segmentation method in Li et al. (2006, 2007) to perform tissue segmentation based on DTI data (Fig. 1c). The basic idea is to
Results
In this section, we present experimental results of the evaluation of the proposed computational framework of fiber bundle clustering and labeling.
Discussions and Conclusion
In this paper, we proposed a hybrid approach to WM fiber clustering and labeling. The top-down anatomy-guided fiber clustering step in this computational framework is based on the brain parcellation results obtained by an atlas-based warping algorithm (Liu et al., 2004), which means the WM fiber clustering result is dependent on the performance of the brain parcellation algorithm. Although the hybrid volumetric and surface warping method (Liu et al., 2004) used in this paper has reasonably good
Acknowledgments
This research work was supported by The Methodist Hospital Research Institute and the NIH Grant 5G08LM008937 (STCW). The authors would like to thank Kaiming Li from UGA for proof reading of the manuscript.
References (49)
- et al.
Diffusion tensor imaging of the corpus callosum in autism
Neuroimage
(2007) - et al.
White matter structure in autism: preliminary evidence from diffusion tensor imaging
Biol. Psychiatry
(2004) - et al.
MR diffusion tensor spectroscopy and imaging
Biophys. J
(1994) - et al.
Fiber tract-oriented statistics for quantitative diffusion tensor MRI analysis
Med. Image Anal.
(2006) - et al.
Automatic cortical sulcal parcellation based on surface principal direction flow field tracking
NeuroImage.
(2009) - et al.
Deformable registration of cortical structures via hybrid volumetric and surface warping
NeuroImage
(2004) - et al.
76-Space analysis of grey matter diffusivity: Methods and applications
NeuroImage
(2006) - et al.
Brain tissue segmentation based on DTI data
Neuroimage
(2007) - et al.
A unified framework for clustering and quantitative analysis of white matter fiber tracts
Medical Image Analysis
(2008) - et al.
Principles of diffusion tensor imaging and its applications to basic neuroscience research
Neuron
(2006)