Elsevier

Medical Image Analysis

Volume 12, Issue 6, December 2008, Pages 713-726
Medical Image Analysis

Towards a validation of atlas warping techniques

https://doi.org/10.1016/j.media.2008.04.003Get rights and content

Abstract

Pre-operative magnetic resonance (MRI) and computed tomography (CT) image volumes are often used for planning and guidance during functional neurosurgical procedures. These operations can include the creation of lesions in the thalamus (thalamotomy) or the globus pallidus (pallidotomy), or the insertion of deep brain stimulation (DBS) electrodes in the subcortical nuclei. These subcortical targets are often difficult to localize in pre-operative imaging data due to the limited resolution and contrast available in standard MRI or CT techniques. To address this problem, digital atlases of subcortical nuclei are often used to accurately identify surgical targets since they can be warped to fit each patient’s unique anatomy. Targeting accuracy thus depends on the quality of the atlas-to-patient warp.

In this paper, three atlas-to-patient warping techniques are compared. Two methods rely on an MRI template as an intermediary to estimate a nonlinear atlas-to-patient transformation. The third is novel, and uses a pseudo-MRI derived from an atlas of the basal ganglia and thalamus to estimate the nonlinear atlas-to-patient transformation directly. The methods are compared using (1) manual segmentations of subcortical nuclei and (2) functional data from intra-operative thalamic stimulation. The results demonstrate that the template-based atlas-to-patient warping technique is the best of the three for customizing the atlas onto patient data.

Introduction

Movement disorders are often treated with functional neurosurgical procedures. These procedures can include the creation of lesions in the thalamus (thalamotomy) (Atkinson et al., 2002, Duval et al., 2005, Lenz et al., 1995, Otsuki et al., 1994) the globus pallidus (pallidotomy) (Cohn et al., 1998, Gross et al., 1999, Lombardi et al., 2000, Starr et al., 1999) as well the introduction deep brain stimulation (DBS) electrodes in the thalamus, globus pallidus, or the subthalamic nucleus (STN) (Eskandar et al., 2001, Krause et al., 2001, Sanchez Castro et al., 2006b, Starr et al., 1999). These procedures require accurate target localization of these subcortical structures using pre-operative magnetic resonance (MRI) or computed tomography (CT) image volumes. Despite recent advances in medical imaging techniques which allow improved visualization of the thalamus (Behrens et al., 2003, Deoni et al., 2005, Johansen-Berg et al., 2005), most clinical MRI volumes lack the contrast and resolution required to properly visualize the subcortical nuclei. These pre-operative scans are often taken with the body coil since a stereotactic head-frame is attached to the patient’s head to establish a co-ordinate system within the image volume. In addition, a variety of other factors can limit contrast and signal-to-noise ratio in awake patients undergoing stereotactic surgery, including motion artefacts related to movement disorders, and the limits of MRI acquisition time. These limitations can cause difficulty for surgeons when distinguishing between the borders of different subcortical nuclei used as surgical targets (see Fig. 1). Originally, print atlases were used in conjunction with anatomical landmarks to guide functional neurosurgical procedures (Schaltenbrand and Wahren, 1977, Talairach and Tournoux, 1988). However, digital atlases have proven to be useful in surgical planning and guidance as they can be customized to pre-operative patient data (Bertrand et al., 1973, Chakravarty et al., 2006a, Finnis et al., 2003, Ganser et al., 2004, Nowinski et al., 2000, St-Jean et al., 1998).

The use of computerized atlases to guide functional neurosurgery was pioneered at the Montreal Neurological Institute in work done by Bertrand, 1982, Bertrand et al., 1973, Bertrand et al., 1974. This work used a digitized version of the Schaltenbrand and Bailey atlas (Schaltenbrand and Bailey, 1959) and used an intra-operative ventriculogram as an imaging reference. Hardy, Bertrand and Thompson later used this atlas in a comprehensive analysis of the somatopy of fiber tracts and subcortical nuclei in the human brain in the late 1970s and early 1980s (Hardy et al., 1979a, Hardy et al., 1979b, Hardy et al., 1979c, Hardy et al., 1979d, Hardy et al., 1979e, Hardy et al., 1980a, Hardy et al., 1980b, Hardy et al., 1980c, Hardy et al., 1981).

More recently, several atlases of the human brain have been used to guide stereotaxic procedures. Several of these atlases rely on digitized versions of the print atlases which are commonly available. Nowinski et al. (1997) have developed a digital atlas that incorporated data from three print atlases including Ono et al., 1990, Schaltenbrand and Wahren, 1977, Talairach and Tournoux, 1988. In order to register the atlas to a subject or patient, a piece-wise linear approach is used to transform the atlas to the MR volume. Later, the atlas-to-patient warping approach was refined by estimating a nonlinear transformation using a finite element approach (Xu and Nowinski, 2001). 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 estimating a Delaunay tetrahedrization on each structure of the atlas. Surfaces of these structures were extracted using the marching cubes algorithm (Lorensen and Cline, 1987). Their atlas warping technique improves on Talairach and Tournoux’s (1988) so-called “proportional grid system” by using a set of radial basis functions to define a deformation field.

Many groups use template based procedures to warp atlases to pre-operative patient data. These procedures warp an atlas to a template MRI. After atlas-to-template alignment, the atlas warping problem is narrowed down to a standard MRI-to-MRI registration problem. Once the template-to-patient transformation is estimated it is applied to the atlas to customize it to patient data. For these template-based techniques, the atlas-to-patient and template-to-patient warps are equivalent.

In the original work of St-Jean et al. (1998) the Colin27 MRI template (Holmes et al., 1998) was used to estimate atlas-to-patient transformations. This template is the average of 27 T1-weighted MRI volumes of the same subject which results in a volume with enhanced signal-to-noise ratio and improved contrast. A 3D version of the Schaltenbrand and Wahren atlas was warped to the template using a thin-plate-spline (TPS) warp (Bookstein, 1989). The TPS warp is a landmark-based technique which aligns landmarks exactly and will interpolate the data between these landmarks by constraining them to the properties of a thin metal plate. As a result, it is possible that there are some misregistrations in the initial atlas-to-template customization, and these errors will be propagated throughout the atlas-to-patient warping process. This is the atlas previously used at the Montréal Neurological Institute.

The Colin27 high-resolution MRI template (Holmes et al., 1998) has also been used to estimate nonlinear atlas-to-patient transformations by Finnis et al. (2003). In their work, a probabilistic functional atlas was developed by warping a digitized version of the Schaltenbrand and Wahren atlas (Schaltenbrand and Wahren, 1977). Co-ordinates representing successful intra-operative stimulation points were clustered into the common atlas space using nonlinear registration methods. To customize the atlas to patient data, the nonlinear transformation was then applied to the digitized atlas and the point clusters. Recent work has expanded the amount of intra-operative electro-physiological data that has been included in their electro-physiological database (Guo et al., 2005).

D’Haese et al. have developed an atlas (D’Haese et al., 2005a) which uses electro-physiological intra-operative recordings from the STN registered to a template created from the average of pre-operative data. This choice of template was later refined (D’Haese et al., 2005b) to a single MRI template, selected based on the best prediction of the final DBS location when using radial basis functions to estimate the nonlinear template-to-patient transformation (Rohde et al., 2003).

Bardinet et al. (2005) reconstructed a set of serial histological data and warped this data set to T1 and T2 reference MRIs using only linear transformations. The transformations are optimized at the level of the basal ganglia and thalamus. Contours of structures in the basal ganglia and the mesencephalon were manually traced on the histology to allow for increased visualization of the anatomy. The final atlas-to-patient transformation was estimated using an affine transformation estimated in a hierarchical fashion and is used to predict the location for the STN DBS stimulator. This position was then correlated with intra-operative electro-physiological findings. The methods used for the development of this atlas were recently detailed in Yelnik et al. (2007).

The use of expert identification of surgical targets was used to create an atlas for STN DBS targeting by Sanchez Castro et al., 2005, Sanchez Castro et al., 2006a, Sanchez Castro et al., 2006b. This atlas is based on the manual identification of the DBS target on multiple patients. The target labeling is repeated on all subjects, by the same expert on five separate occasions to eliminate intra-rater error. The points from each subject were averaged and then warped to a reference pre-operative MRI in order to create an atlas. Four different warping techniques were evaluated (Schaltenbrand and Wahren (1977), affine transformations (Maes et al., 1997), Demons algorithm (Thirion et al., 1996, Thirion, 1998), and B-splines (Rueckert et al., 1999)) using correspondence to the final target point assessed in post-operative data as the evaluation criteria.

Each procedure described above is predicated on finding and optimizing techniques for atlas-to-patient warping. The goal is to integrate these atlases in the planning phase for functional neurosurgery. Thus the pre-operative planning is limited by the accuracy of the warping techniques used for atlas-to-template warping. However validation of any warping techniques is typically difficult due to the lack of an universally accepted “gold standard” and requires the development of novel techniques for evaluation. Many of the groups have used intra-operative functional recordings (Bardinet et al., 2005), post-operative data (D’Haese et al., 2005a), or STN DBS targets defined by surgical experts (Sanchez Castro et al., 2006a, Sanchez Castro et al., 2005). To our knowledge, groups have not used a volumetric representation of the anatomy for validation. Finnis et al. (2003) used intra-operative stimulation eliciting visual activations and compared the location of these stimulation points with regards to their position relative to the optic tract. As an extension to this original work, Guo et al. (2006) recently published a more comprehensive validation of the initial work from this group using final target locations from pallidotomies, thalamotomies, and thalamic and subthalamic nucleus DBS implantations. In the validation provided by Nowinski et al. (2000) for the warping techniques for their multiple brain atlas database, a comparison was done of subjects who were operated (for pallidotomy, thalamotomy, and thalamic DBS) and the final target location was compared with respect to the atlas-derived predicted target location. However, only visual validation results were presented when the atlas-to-patient and inter-atlas (i.e., Talairach to Schaltenbrand atlas) warping techniques was refined using an FEM method (Xu and Nowinski, 2001). In the atlas developed by Ganser et al. (2004) only the location of the frontal tip of the putamen was verified for validation purposes. However the accuracy of the identification of structures is critical as targets are often chosen relative to structures which are easily visible using pre-operative MRI volumes.

In other comparisons of nonlinear registration algorithms, the use of volumetric classifications and segmentations have been used optimize or compare different strategies. Robbins et al. (2004) used the minimization of entropy between segmented MRI volumes as a technique for the optimization of nonlinear registration parameters of the automatic nonlinear image matching and anatomical labeling (ANIMAL) algorithm of Collins and Evans, 1997, Collins et al., 1995. In a broader study Hellier et al. (2003) studied the accuracy of several commonly-used nonlinear registration techniques using several different criteria including: global volume, overlap of different segmented tissue classes, curvature of the iso-intensity surfaces, consistency of the nonlinear deformation, as well as quantitative and qualitative evaluation of sulcii after nonlinear warping.

In this paper we use an atlas created from reconstructed serial histological data (Chakravarty et al., 2006a) which has been warped to the Colin27 high resolution MRI template (Holmes et al., 1998) (based on an atlas-to-template warping procedure). This anatomical atlas is described in Section 2. The atlas is customized to a particular patient’s anatomy (using an atlas-to-patient warping procedure) with the ANIMAL nonlinear registration technique (Collins and Evans, 1997, Collins et al., 1995) (Section 2.2), using three different strategies: two depend on the Colin27 template and the third uses a “pseudo-MRI” (Chakravarty et al., 2006a). The work presented here is a continuation of our preliminary validation work (Chakravarty et al., 2005).

The main goal of this paper is to compare three atlas-to-patient warping strategies to determine which is best for surgical planning. To achieve this goal, three subgoals must be met: (1) validate the initial atlas-to-template warping procedure, since any errors in atlas–template alignment will be propagated in the two template-based atlas-to-patient mappings; (2) develop metrics for minimally-biased anatomical evaluation of the atlas-to-patient warping techniques; (3) develop metrics for electro-physiological evaluation of the atlas-to-patient warping techniques. In addition to validating the ANIMAL algorithm for atlas customization in the context of targeting in functional neurosurgical procedures used in movement disorders, the contributions of this paper include the development and evaluation of a pseudo-MRI-based nonlinear registration procedure, the development of manually-derived silver-standard labels for anatomical validation, and the development of new metrics based on electro-physiological recordings for further warping validation.

The paper is organized in the following manner: Section 2 presents the methods, describing both the atlas (Section 2.1) and the nonlinear registration algorithm (Section 2.2) used. Section 2.3 describes the atlas-to-template warping strategy, while Sections 2.4 Template based atlas-to-patient warping: Method A, 2.5 Template based atlas-to-patient warping: Method B, 2.6 Pseudo-MRI based atlas-to-patient warping: Method C describe the three atlas-to-patient warping strategies. A series of experiments are described in Section 3. The creation and evaluation of the manual label-based silver standard is described in Section 3.2. Once the quality of the silver standard has been characterized (Section 3.2.3), it is used to evaluate the atlas-to-template warping (Section 3.2.4) and the three atlas-to-patient warping strategies (Section 3.2.5). Section 3.3 describes the electro-physiological metrics and validation. Section 4 presents the experimental results, Section 5 includes a summary and discussion and Section 6 ends with conclusions and suggestions for future work.

Section snippets

Methods

In this section the atlas of the basal ganglia and thalamus developed for planning functional neurosurgical procedures and the techniques used to warp this atlas to a high resolution MRI template are discussed in Section 2.1. The ANIMAL algorithm used for atlas-to-template and atlas-to-patient nonlinear warping is described in Section 2.2. The atlas-to-template warping procedure is described in Section 2.3. The three different atlas-to-patient warping techniques based on the ANIMAL algorithm

Atlas warping evaluation – experimental methods

The following sections present anatomical and functional validation for the atlas-to-patient transform using manual anatomical labels and intra-operative electro-physiological recordings. The manual labeling protocol is first presented followed by four different validation experiments. In the first experiment, the quality of the manual labels are evaluated to verify that they can be used to validate both the atlas-to-template and atlas-to-patient warps. The second experiment evaluates the

Quality of manual labeling

The results of the manual labeling are given in Table 3. A MANOVA using the kappas of the raters as the main effect and structure and hemisphere as covariates showed no significant effect due to hemisphere (p = 0.5591). Subsequently the covariate of hemisphere was eliminated from the analysis. An effect of structure (F = 106.9225, p < 0.0001, DF = 2) and rater (F = 30.4028, p < 0.0001, DF = 4), as well as interactions between structure and rater were observed (F = 14.0902, p < 0.001, DF = 9). A post hoc test

Summary

In this paper we developed techniques for the anatomical and electro-physiological validation of atlas-to-template and atlas-to-patient warping techniques using a digital atlas of the basal ganglia and thalamus previously developed in our group (Chakravarty et al., 2006a). The atlas was nonlinearly warped using the ANIMAL algorithm (Collins and Evans, 1997, Collins et al., 1995) to a high resolution MRI template (Holmes et al., 1998) via a direct atlas-to-template warping scheme that relies

Conclusions and future work

Based on the findings of this study it can be inferred that Method B is the most accurate method, of the three methods tested, to customize subcortical structures to a particular patients anatomy. The results are comparable to the variability of the manual raters. However this technique does not behave in a manner which is statistically better than Method A, therefore it is left to the user to decide between the trade-off for accuracy and time taken to develop the high-resolution vector field

References (72)

  • J. Atkinson et al.

    Optimal location of thalamotomy lesions for tremor associated with Parkinson’s disease: a probabilistic analysis based on postoperative magnetic resonance imaging and integrated digital atlas

    Journal of Neurosurgery

    (2002)
  • E. Bardinet et al.

    Retrospective cross-evaluation of a histological and deformable 3D atlas of the basal ganglia on series on Parkinsonian patients treated by deep brain stimulation

  • T.E.J. Behrens et al.

    Non-invasive mapping of connections between human thalamus and cortex using diffusion imaging

    Nature Neuroscience

    (2003)
  • A.L. Benabid et al.

    Imaging of subthalamic nucleus and ventralis intermedius of the thalamus

    Movement Disorders

    (2002)
  • G. Bertrand

    Computers in functional neurosurgery

  • G. Bertrand et al.

    The computerized brain atlas: its use in stereotaxic surgery

    Transactions of the American Neurological Association

    (1973)
  • G. Bertrand et al.

    Computer display of stereotaxic brain maps and probe tracts

    Acta Neurochirurgica

    (1974)
  • P. Blomstedt et al.

    Thalamic deep brain stimulation in the treatment of essential tremor: a long-term follow-up

    British Journal of Neurosurgery

    (2007)
  • F.L. Bookstein

    Principal warps: thin-plate splines and the decomposition of deformations

    IEEE Transactions on Pattern Analysis and Machine Intelligence

    (1989)
  • M.M. Chakravarty et al.

    Anatomical and electrophysiological validation of an atlas for neurosurgical planning

  • M.M. Chakravarty et al.

    Towards a multi-modal atlas for neurosurgical planning

  • M.C. Cohn et al.

    Pre- and postoperative evaluation of stereotactic pallidotomy

    American Journal of Neuroradiology

    (1998)
  • D.L. Collins et al.

    ANIMAL: validation and application of non-linear registration-based segmentation

    International Journal of Pattern Recognition and Artificial Intelligence

    (1997)
  • D.L. Collins et al.

    Automatic 3-D model based neuroanatomical segmentation

    Human Brain Mapping

    (1995)
  • D.L. Collins et al.

    Automatic 3D intersubject registration of MR volumetric data in standardized Talairach space

    Journal of Computer Assisted Tomography

    (1994)
  • S.C.L. Deoni et al.

    High resolution T1 and T2 mapping of the brain in clinically acceptable time with DESPOT1 and DESPOT2

    Magnetic Resonance in Medicine

    (2005)
  • P.F. D’Haese et al.

    Computer-aided placement of deep brain stimulators: from planning to intraoperative guidance

    IEEE Transactions on Medical Imaging

    (2005)
  • P.F. D’Haese et al.

    Automatic selection of DBS target points using multiple electrophysiological atlases

  • E.G. Duerden et al.

    A method for analysis of electrophysiological responses obtained from the motor fibers of the human internal capsule

  • E.N. Eskandar et al.

    Surgical treatment of parkinson’s disease

    Journal of American Medicine

    (2001)
  • K.W. Finnis et al.

    Three-dimensional database of subcortical electrophysiology for image-guided stereotactic functional neurosurgery

    IEEE Transactions on Medical Imaging

    (2003)
  • P. Gloor

    The Temporal Lobe and Limbic System

    (1997)
  • R.E. Gross et al.

    Relationship of lesion location to clinical outcome following microelectrode-guided pallidotomy for Parkison’s disease

    Brain

    (1999)
  • T. Guo et al.

    Development and application of functional databases for planning deep-brain neurosurgical procedures

  • T. Guo et al.

    Visualization and navigation system development and application for stereotactic deep-brain neurosurgeries

    Computer Aided Surgery

    (2006)
  • T.L. Hardy et al.

    Organization and topography of sensory responses in the internal capsule and nucleus ventralis caudalis found during stereotactic surgery

    Applied Neurophysiology

    (1979)
  • Cited by (85)

    View all citing articles on Scopus
    View full text