Volumetric vs. surface-based alignment for localization of auditory cortex activation

Abstract

The high degree of intersubject structural variability in the human brain is an obstacle in combining data across subjects in functional neuroimaging experiments. A common method for aligning individual data is normalization into standard 3D stereotaxic space. Since the inherent geometry of the cortex is that of a 2D sheet, higher precision can potentially be achieved if the intersubject alignment is based on landmarks in this 2D space. To examine the potential advantage of surface-based alignment for localization of auditory cortex activation, and to obtain high-resolution maps of areas activated by speech sounds, fMRI data were analyzed from the left hemisphere of subjects tested with phoneme and tone discrimination tasks. We compared Talairach stereotaxic normalization with two surface-based methods: Landmark Based Warping, in which landmarks in the auditory cortex were chosen manually, and Automated Spherical Warping, in which hemispheres were aligned automatically based on spherical representations of individual and average brains. Examination of group maps generated with these alignment methods revealed superiority of the surface-based alignment in providing precise localization of functional foci and in avoiding mis-registration due to intersubject anatomical variability. Human left hemisphere cortical areas engaged in complex auditory perception appear to lie on the superior temporal gyrus, the dorsal bank of the superior temporal sulcus, and the lateral third of Heschl's gyrus.

Introduction

A common technique in functional magnetic resonance imaging (fMRI) studies is to compare the location of functional activations under varying conditions in a group of subjects, and to display these activations in the form of a group statistical map. The anatomy of the human brain is highly variable, however, and brain structures vary in their size, shape, position, and relative orientation. This variation is even more significant for pathological brains. This makes the problem of pooling data across different anatomies non-trivial (Rademacher et al., 1993, Roland and Zilles, 1994). In order to compare and overlay anatomies and functional activations, a one-to-one mapping needs to be specified so that each location in one brain corresponds to a unique location in another brain.

A common method for specifying this registration between brains is based on the 3D normalization described by Talairach and Tournoux (1988). This method accounts for differences in brain size and changes in head position inside the scanner across subjects. It is simple to use and is applicable to both cortical and subcortical structures. However, it does not afford a high degree of anatomical accuracy. Several studies have shown that the 3D distance between anatomical landmarks in different brains after Talairach normalization can be over 10 mm (Steinmetz et al., 1990, Thompson and Toga, 1996, Van Essen and Drury, 1997). Thus, precise localization of functional areas with respect to gyral and sulcal landmarks is not possible. Furthermore, due to the highly folded nature of the cortical sheet, the distance between two points in the 3D space is often not representative of the true distance between these points on the 2D cortical sheet. A small inaccuracy resulting from the Talairach normalization may translate into a large inaccuracy in terms of the true distances on the 2D sheet, for example, for points located on the opposite banks of a sulcus.

The basic reason for the inadequacy of the Talairach normalization is that it is based only on linear transformations of scaling, translation, and rotation. If nonlinear transformations such as local dilation, contraction, or shearing can be applied to deform the images, registration can be significantly improved. Considerable research has been directed towards development of nonlinear methods in the last decade (Maintz and Viergever, 1998, Thompson and Toga, 2000, Toga, 1999). High-dimensional morphing methods have been suggested that can morph an entire 3D volume to match the intensity values with a canonical anatomy (Christensen et al., 1996, Christensen et al., 1997, Evans et al., 1994, Joshi et al., 1997, Woods et al., 1998). However, the sulcal and gyral landmarks, which are often correlated with function, are a property of the 2D cortical sheet. Intensity-driven high-dimensional morphing does not guarantee the alignment of sulcal and gyral patterns; an explicit representation of the surface is needed. Furthermore, anatomical landmarks do not necessarily have a fixed relationship with functional areas. In some cases, it may be desirable to identify the functional areas in individual brains and align them explicitly. Talairach normalization and automated 3D morphing are not suitable for this purpose.

Several surface-based methods allow alignment based explicitly on surface landmarks. These methods optimally transform the cortical sheet to a mathematically simpler shape such as a 2D sheet (Drury et al., 1996, Drury et al., 1997, Drury et al., 1999, Van Essen et al., 1998), an ellipsoid (Sereno et al., 1996), or a sphere (Davatzikos, 1996, Fischl et al., 1999, Thompson and Toga, 1996, Thompson and Toga, 1998, Van Essen, 2004).

Van Essen et al. (1998) modeled the flattened surface as a viscoelastic fluid sheet (Joshi and Miller, 2000). In this approach, landmarks are manually identified on the surface, and the surface is deformed so as to align them closely with the corresponding landmarks on the target surface. The viscoelastic properties of the sheet reduce the distortions in the deformed surface and allow a satisfactory compromise between the objectives of bringing the landmarks into register as closely as possible and minimizing the distortions in the surface. Furthermore, working with 2D surfaces makes the morphing process computationally more tractable than 3D deformation of the entire volume. We refer to this approach as the Landmark Based Warping (LBW) method.

Fischl et al. (1999) provided an automated method for registration based on spherical representation of the cortex. Spherical representation provides a mathematically simple surface suitable for deformation. The surfaces of 40 individual brains were transformed to spherical forms while minimizing metric distortions using an energy function based on convexity. These individual spherical maps were combined to construct an average map of the large-scale folding patterns on a unit sphere. An automated method then non-rigidly aligns the surface of any individual brain, converted to the spherical representation, to this average sphere. We refer to this as the Automated Spherical Warping (ASW) approach.

While there is considerable ongoing research on registration methods, the current practice of fMRI group analysis lags behind to some extent. Our goals in this study were to (1) empirically evaluate and compare the registration based on Talairach transformation with the registration based on two commonly available surface-based methods, LBW and ASW, specifically with respect to auditory cortex activation; and (2) obtain high-resolution maps of the areas activated by speech sounds compared to non-speech sounds. Experiments in the monkey suggest that the auditory cortex may contain 12 or more subdivisions in a relatively small region (Kaas et al., 1999). Thus, intersubject alignment is potentially critical in elucidating the precise topographical arrangement of different auditory fields.

We designed an fMRI experiment to generate blood oxygenation (BOLD) signals in temporal areas involved in speech perception (Binder et al., 2000, Hall et al., 2003, Liebenthal et al., 2003). Subjects performed a two-alternative forced-choice discrimination task with tokens consisting of phonetic sounds (/ba/ and /da/) or tones. By comparing the BOLD signal elicited by the phonetic and tone conditions to a baseline of silence, one can identify areas involved in the perception of speech and complex non-speech sounds, respectively. The phonemes > tones contrast can identify areas that are activated more exclusively by speech sounds.

Note that there are three sources of variance when combining data across subjects: (1) the variance in the location of the anatomical landmarks, (2) the variation in the location of cortical fields with respect to the anatomical landmarks, and (3) the variation in the activation of various cortical fields given a task, and in the location of activation within a cortical field. Alignment methods based on anatomical landmarks (either manually or automatically chosen) can only reduce variance due to source (1). If the variance due to sources (2) and (3) is high, it is difficult to assess the anatomical accuracy of alignment based on functional data (Kang et al., 2004). To better assess and compare alignment accuracy with different methods based on anatomical criteria, we also calculated dispersion indices of several anatomical fiducial points after alignment with different methods.