The creation of a brain atlas for image guided neurosurgery using serial histological data

Abstract

Digital and print brain atlases have been used with success to help in the planning of neurosurgical interventions. In this paper, a technique presented for the creation of a brain atlas of the basal ganglia and the thalamus derived from serial histological data. Photographs of coronal histological sections were digitized and anatomical structures were manually segmented. A slice-to-slice nonlinear registration technique was used to correct for spatial distortions introduced into the histological data set at the time of acquisition. Since the histological data were acquired without any anatomical reference (e.g., block-face imaging, post-mortem MRI), this registration technique was optimized to use an error metric which calculates a nonlinear transformation minimizing the mean distance between the segmented contours between adjacent pairs of slices in the data set. A voxel-by-voxel intensity correction field was also estimated for each slice to correct for lighting and staining inhomogeneity. The reconstructed three-dimensional (3D) histological volume can be viewed in transverse and sagittal directions in addition to the original coronal.

Nonlinear transformations used to correct for spatial distortions of the histological data were applied to the segmented structure contours. These contours were then tessellated to create three-dimensional geometric objects representing the different anatomic regions in register with the histological volumes. This yields two alternate representations (one histological and one geometric) of the atlas.

To register the atlas to a standard reference MR volume created from the average of 27 T1-weighted MR volumes, a pseudo-MRI was created by setting the intensity of each anatomical region defined in the geometric atlas to match the intensity of the corresponding region of the reference MR volume. This allowed the estimation of a 3D nonlinear transformation using a correlation based registration scheme to fit the atlas to the reference MRI.

The result of this procedure is a contiguous 3D histological volume, a set of 3D objects defining the basal ganglia and thalamus, both of which are registered to a standard MRI data set, for use for neurosurgical planning.

Introduction

Functional stereotactic neurosurgery is increasingly used for the treatment for movement disorders such as Parkinson's disease (Atkinson et al., 2002, Cohn et al., 1998, Cuny et al., 2002, Gross et al., 1999, Lenz et al., 1995, Lombardi et al., 2000, Samuel et al., 1998). While symptoms associated with Parkinson's disease (such as tremor, rigidity, bradykinesia, and impaired gait) can be treated pharmacologically, intractable cases require surgical intervention. For surgical candidates, this can include the introduction of recording or stimulating probes in deep cerebral structures and the creation of lesions in the thalamus (thalamotomy) (Atkinson et al., 2002, Gross et al., 1999, Lenz et al., 1995) or globus pallidus (pallidotomy) (Cohn et al., 1998, Lombardi et al., 2000, Samuel et al., 1998), or insertion of brain stimulation electrodes in the thalamus, globus pallidus, or subthalamic nucleus. Pre-surgical planning of these procedures requires a detailed analysis of the thalamus and the basal ganglia from pre-operative Computed Tomography (CT) and Magnetic Resonance Imaging (MR) volumes. However, cyto-architectonic boundaries between specific subcortical nuclei are often indistinguishable due to the limited contrast and resolution of these imaging modalities.

While new MR imaging and image processing techniques enable visualization of some nuclei (Deoni et al., 2005, Fujita et al., 2001, Starr et al., 1999), atlases are often used in conjunction with more standard imaging techniques to enhance the visualization of surgical targets for pre-operative planning and to accurately predict the optimal location of surgical targets in sub-cortical nuclei (Atkinson et al., 2002, Bertrand et al., 1973, Nowinski et al., 1997, Nowinski et al., 2000, St-Jean et al., 1998, Xu and Nowinski, 2001). Diffusion tensor imaging has also been used to map thalamo-cortical connections in order to derive probabilistic segmentations of the human thalamic nuclei (Behrens et al., 2003, Johansen-Berg et al., 2005). However, at this time, neither of these in vivo imaging techniques is able to provide the resolution required to accurately identify the subcortical nuclei targeted in functional neurosurgery, and therefore cannot provide the detailed segmentation that we present here.

Print atlases were the first visualization tools used to aid in the identification of anatomical structures for surgical planning (Afshar et al., 1978, Ono et al., 1990, Schaltenbrand and Bailey, 1959, Schaltenbrand and Wahren, 1977, Schnitzlein and Murtagh, 1980, Talairach and Tournoux, 1988, Talairach and Tournoux, 1993, Van Buren and Borke, 1972, Watkins, 1969). Typically, digital atlases are 3D visualizations of the data presented in these atlases. When digital atlases were first used, linear transformations matching the atlas to patient data were used to register an atlas to an individual patient scan (Bertrand et al., 1973, Nowinski et al., 1997, Otsuki et al., 1994). Our group was one of the first to estimate and apply nonlinear transformations to warp a digital atlas to fit pre-operative patient MR data in order to account for local variations in the anatomy (St-Jean et al., 1998).

A number of digital atlases of the human brain, based on print atlases, have been previously published. Nowinski et al. (1997) have developed an integrated digital atlas that incorporated data from three print atlases by Ono et al., 1990, Schaltenbrand and Wahren, 1977, and Talairach and Tournoux (1988). All three atlases are registered together using landmark based linear transformations needed to map the Ono et al. (1990) and Schaltenbrand and Wahren (1977) atlases into Talairach space (Talairach and Tournoux, 1988). In order to register the combined atlas to a subject or patient, a piece-wise linear approach is used to transform the atlas to the MR volume. The Talairach and Tournoux atlas was also used as the foundation for a digital atlas by Ganser et al. (2004). The original plates were scanned and reconstructed in three dimensions by calculating a Delauney tetrahedrization. The surfaces of anatomic structures in the Talairach atlas were reconstructed using the marching cubes algorithm (Lorensen and Cline, 1987). The resulting volume was then intersected at half the slice-to-slice distance to complete the interpolation.

The digital atlas used previously at the Montreal Neurological Institute (MNI) was developed by St-Jean et al. (St-Jean et al., 1998). It is based on a 3D reconstruction of the axial contour data from the Schaltenbrand and Wahren atlas (Schaltenbrand and Wahren, 1977). The digital atlas included 16 structures and has a varying slice thickness of 0.5 to 3mm due to the slice-to-slice distance of the original atlas. The reconstructed data set was interpolated with a Hermite polynomial (Foley et al., 1990) to achieve a 0.5mm isotropic resolution. Slice-to-slice spatial inconsistencies in structure contours were considered to be small, and thus not accounted for. The interpolated digital atlas was warped in 3D to fit a high resolution, high signal-to-noise ratio standard reference volume that is the result of the average of 27 MRI scans of the same subject (Holmes et al., 1998), known as the Colin27 MRI average. The warping was achieved via a thin-plate-spline (TPS) (Bookstein, 1989) transformation based on 250 homologous landmarks manually identified by a neuroanatomist on both volumes. The result is a set of anatomical labels defined in the region of the basal ganglia and thalamus that are aligned with the Colin27 MRI average. In order to customize the digital atlas to patient MRI data, a nonlinear transform between the Colin27 MRI average and a patient's MRI is estimated automatically (Collins and Evans, 1997, Collins et al., 1995). This transformation is applied to the digital atlas to map it onto patient's pre-operative MRI to facilitate surgical planning.

While this atlas has proven very useful (Atkinson et al., 2002), it has limited inherent resolution in the slice direction (0.5 mm), contains a limited number of structures, and contains some small misregistrations between the digital atlas and the Colin27 MRI average that are propagated to patient MRI data during the atlas customization procedure.

In this manuscript, these limitations are addressed. Techniques are developed for the creation of a new and improved atlas for stereotactic neurosurgery. This atlas contains both histological and geometric (i.e., structural anatomical) data and is registered to a MRI reference volume. Preliminary work on this new atlas was presented in (Chakravarty et al., 2003). The new digital atlas is derived from a single set of high-resolution, thin-slice histological data of the region of basal ganglia and thalamus. The atlas contains 105 anatomical structures that were manually delineated by a neuroanatomical expert on the histological data using sources for the gross anatomy (Schaltenbrand and Wahren, 1977), for the temporal lobe (Gloor, 1997) and for the thalamus (Hirai and Jones, 1989). The histology was parcelated three times according to these sources. To reconstruct the histological and geometric data in three dimensions (3D), the structure contours were used in the development of an optimization procedure for slice-to-slice registration and intensity correction of the histological data. These reconstructions were registered to the Colin27 high resolution reference MRI (Holmes et al., 1998), using a novel atlas-to-MRI matching technique. Atlas customization (to any subject scan) can be achieved through a flexible nonlinear atlas-to-subject registration technique.

Our long-term goal is to use this refined atlas to improve pre-operative planning and thereby positively affect the outcome for patients undergoing surgeries for movement disorders. Since this atlas contains a detailed segmentation and classification of subcortical nuclei, it can also be used in post-operative follow-up and in other applications requiring a detailed analysis of the basal ganglia and thalamus.

In this paper, we will discuss the optimization of a slice-to-slice histological data registration technique (Chakravarty et al., 2003) used to minimize morphological misalignment throughout the histological volume. The parameters are optimized based on the minimization of error between the segmented contours of seven pairs of adjacent slices which span the dataset. These parameters are then used to register all consecutive pairs of slices of the histological data. In addition, an improved intensity inhomogeneity correction technique based on previous work (Chakravarty et al., 2003), the creation of a 3D geometric atlas, and an atlas-to-template warping technique are also presented. Since the reconstruction and intensity correction of histological data plays an integral role in this work, the next section will review previous techniques used in these domains.