Recognition of white matter bundles using local and global streamline-based registration and clustering

Abstract

Virtual dissection of diffusion MRI tractograms is cumbersome and needs extensive knowledge of white matter anatomy. This virtual dissection often requires several inclusion and exclusion regions-of-interest that make it a process that is very hard to reproduce across experts. Having automated tools that can extract white matter bundles for tract-based studies of large numbers of people is of great interest for neuroscience and neurosurgical planning. The purpose of our proposed method, named RecoBundles, is to segment white matter bundles and make virtual dissection easier to perform. This can help explore large tractograms from multiple persons directly in their native space. RecoBundles leverages latest state-of-the-art streamline-based registration and clustering to recognize and extract bundles using prior bundle models. RecoBundles uses bundle models as shape priors for detecting similar streamlines and bundles in tractograms. RecoBundles is 100% streamline-based, is efficient to work with millions of streamlines and, most importantly, is robust and adaptive to incomplete data and bundles with missing components. It is also robust to pathological brains with tumors and deformations. We evaluated our results using multiple bundles and showed that RecoBundles is in good agreement with the neuroanatomical experts and generally produced more dense bundles. Across all the different experiments reported in this paper, RecoBundles was able to identify the core parts of the bundles, independently from tractography type (deterministic or probabilistic) or size. Thus, RecoBundles can be a valuable method for exploring tractograms and facilitating tractometry studies.

Introduction

Current tractography algorithms produce a large number of streamlines (up to millions) that form a tractogram. Recently, it has been reported that tractography can generate a great number of invalid streamlines and white matter bundles (Côté et al., 2013, Maier-Hein et al., 2016, Jones et al., 2013, Thomas et al., 2014), which can create biases in tractometry analysis (Dayan et al., 2016, Bells et al., 2011, Yeatman et al., 2012, O'Donnell et al., 2009, Renauld et al., 2016) and connectivity analysis (Hagmann et al., 2008). A partial solution to this problem is to inject strong anatomical priors into the tractography and track-dissection processes (Girard et al., 2014, Chamberland et al., 2017). The term virtual dissection (Catani et al., 2002a) is often used when an interactive visualization tool (Wang et al., 2007a, Chamberland et al., 2014) is used to segment a bundle of interest while removing all streamlines and bundles that do not correspond to the anatomical definition of this bundle. For this reason, many groups still manually segment hundreds of bundles using visual tools (Chamberland et al., 2014) with inclusion/exclusion regions-of-interests (ROIs) (Wang et al., 2007b). This manual process is very time consuming, is strongly user-dependent and adds biases to tract-based analyses. It can take hundreds of hours for trainees and neuroanatomical experts to manually segment bundles over multiple participants. There is thus an important need for automatic tools that can facilitate the process and provide new databases of white matter bundles.

We can divide the existing automated tools into two categories: a) connectivity-based and b) streamline-based. In (a), the streamlines of the tractograms are filtered in relation to their connection to specific ROIs, often pre-defined in a template. These approaches require image-based multi-modal nonlinear registration, for example by registering a T1 to diffusion data (or the opposite) so that the streamlines and the ROIs are in the same space. In (b), an atlas of known bundles is given, and the streamlines of the atlas are used directly to perform the segmentations of the given tractograms.

Most published methods belong in category (a). For example, one of the most recent methods was introduced by Wassermann et al. (2016) where the authors created rules describing how known anatomical bundles connect different regions of the brain (parcellated using Freesurfer), and then used those rules to extract bundles from tractograms. Another method which is based on Freesurfer parcellations is Tracula (Yendiki et al., 2011, Siless et al., 2016), which also requires registration of an individual T1 to a common template. A similar approach had been proposed earlier by Zhang et al. (2010) and later by Tunç et al. (2014). These approaches all need to have multiple modalities in a common space (Yeatman et al., 2012, Jin et al., 2014) using image-based registration, a possible parcellation, and filtering of streamlines connecting different ROIs. Obtaining such a perfect alignment between streamlines and ROIs can be a difficult task especially when processing pathological brains. The present paper belongs in category (b). Here, most of the existing approaches are quite recent (O'Donnell and Westin, 2007, Guevara et al., 2012, Garyfallidis et al., 2015a, O'Donnell et al., 2017) and do need some form of initial registration (Olivetti et al., 2016, Durrleman et al., 2011). However, they mostly use distances between streamlines and clustering of streamlines to simplify the segmentation process. Furthermore, multiple bundle atlases can be used simultaneously to increase the accuracy of segmentation as shown by Yoo et al. (2015) and Labra et al., 2013, Labra et al., 2016.

In our novel approach, the need for fast clustering and simplification of large tractograms using streamline distances is addressed using QuickBundles (Garyfallidis et al., 2012) and the need for fast linear registration of streamlines is addressed using Streamline-based Linear Registration (SLR) by Garyfallidis et al. (2015b). We show how by combining these two methods at a global and local level we can create a powerful streamline-based segmentation method. The new method is called RecoBundles, which stands for recognition of bundles. RecoBundles is 100% streamline-based and is efficient to work with millions of streamlines. Moreover, we show that RecoBundles is robust and adaptive to incomplete data and bundles with missing components. It is also robust to pathological brains with tumors which deform or reduce the size of bundles. We quantitatively evaluated our results in 4 bundles manually dissected by 2 experts in 60 individuals and showed that RecoBundles is in overall good agreement with experts having high sensitivity and specificity. We also qualitatively show that RecoBundles discovers the core parts of the bundles and, overall, produces more complete (dense) bundles. The amount of filling can be adjusted by a free parameter (λ) which is an advantageous feature for neuroanatomists wanting to explore the extent of specific tracts.

RecoBundles facilitates the testing of hypotheses about the existence of different bundles and running them across many participants. It opens novel perspectives for bundle-extraction for neurosurgical planning and tractometry. The algorithmic aspects of RecoBundles are explained in detail in Section 2 and the method is validated with simulations in Section 3.1. Furthermore, we show quantitative results with real datasets in Section 3.2.1 and qualitative results with healthy and patient datasets in Sections 3.2.2 and 3.3.