Adverse Effects of Template-based Warping on Spatial fMRI Analysis

Bernard Ng*a, Rafeef Abugharbieha, and Martin J. McKeownb

aBiomedical Signal and Image Computing Lab, Department of Electrical and Computer Engineering bPacific Parkinson’s Research Center, Department of Medicine (Neurology)

The University of British Columbia, Vancouver, BC, Canada V6T 1Z4

ABSTRACT

Conventional voxel-based group analysis of functional magnetic resonance imaging (fMRI) data typically requires warping each subject’s brain images onto a common template to create an assumed voxel correspondence. The implicit assumption is that aligning the anatomical structures would correspondingly align the functional regions of the subjects. However, due to anatomical and functional inter-subject variability, mis-registration often occurs. Moreover, whole-brain warping is likely to distort the spatial patterns of activation, which have been shown to be important markers of task-related activation. To reduce the amount of mis-registration and distortions, warping at the brain region level has recently been proposed. In this paper, we investigate the effects of both whole-brain and region-level warping on the spatial patterns of activation statistics within certain regions of interests (ROIs). We have chosen to examine the bilateral thalami and cerebellar hemispheres during a bulb-squeezing experiment, as these regions are expected to incur task-related activation changes. Furthermore, the appreciable size difference between the thalamus and cerebellum allows for exploring the effects of warping on various ROI sizes. By applying our recently proposed 3D moment-based invariant spatial features to characterize the spatial pattern of fMRI activation statistics, we demonstrate that whole-brain warping generally reduced discriminability of task-related activation differences. Applying the same spatial analysis to ROIs warped at the region level showed some improvements over whole-brain warping, but warp-free analysis resulted in the best performance. We hence suggest that spatial analysis of fMRI data that includes spatial warping to a common space must be interpreted with caution.

Keywords: functional neuroimaging, functional magnetic resonance imaging (fMRI), spatial fMRI analysis, region of interest (ROI) characterization, template-based spatial warping

1. INTRODUCTION In conventional functional magnetic resonance imaging (fMRI) analysis, each voxel is typically analyzed in isolation, which neglects the information encoded by spatial patterns of activation [1,2]. Making group inference under such voxel-based approach often entails whole-brain spatial normalization (hereafter referred to as ‘warping’ to avoid confusion with other normalization procedures) of each subject’s brain images to a common template to generate voxel correspondence. This approach assumes that aligning anatomical structures would correspondingly align the functional regions across subjects. However, due to the large anatomical and functional inter-subject variability, spatial warping is prone to mis-registration errors. For instance, as discussed in a review by Uylings et al. [3], the location of Brodmann areas can vary considerably across normal individuals even after non-linear elastic registration. This issue of anatomical variability is further magnified in studies investigating the effects of age [4,5] and neurological diseases [6], where the shape and size of the subjects’ brains may significantly depart from the commonly used templates [7,8]. To circumvent this issue, Brett et al. proposed to mask out regions of the brain that differ from the template so that these problematic regions would not affect the registration process [6]. However, manual delineation of such regions is quite laborious and automating the procedure is a highly challenging problem on its own. Moreover, further complicating the mis-registration problem, the location of the functional regions can still vary substantially across subjects even if the anatomical structures can be perfectly aligned [9]. In fact, in a comparative study, Crivello et al. showed that functional inter-subject variability is much larger than anatomical variability and the precise alignment of the anatomical structures appears to have little influence on the amount of functional overlap across subjects [10].

*bernardn@ece.ubc.ca; phone 1 604 822-8851; fax 1 604 822-5949; bisicl.ece.ubc.ca

Medical Imaging 2009: Biomedical Applications in Molecular, Structural, and Functional Imaging, edited by Xiaoping P. Hu,Anne V. Clough, Proc. of SPIE Vol. 7262, 72621Y · © 2009 SPIE · CCC code: 1605-7422/09/$18 · doi: 10.1117/12.811422

Proc. of SPIE Vol. 7262 72621Y-1

To reduce the amount of mis-registration, region-level warping has been recently proposed and shown to provide higher sensitivity to activation detection compared to whole-brain warping [11]–[13]. Under this approach, a brain region of interest (ROI) is first segmented for each subject in its native space. The segmented ROIs are then aligned and averaged across subjects to generate an ROI template. The ROI of each subject is then warped onto the template to generate voxel correspondence. This region-level warping approach is especially important for analyzing smaller brain regions such as the thalamus and other subcortical structures, which tend to be mis-aligned in whole-brain warping. The reason is that smaller regions contribute considerably less to the objective function in driving the whole-brain registration process compared to larger brain structures. Hence, adopting a region-level warping approach, where each ROI is separately aligned, is likely to reduce anatomical mis-registration. However, functional mis-alignment remains an issue. Also, any spatial warping will certainly distort the pattern of activation, which may affect the outcomes of spatial fMRI analysis.

An alternative approach for establishing a correspondence across subjects without spatial warping is to specify ROIs for each subject and examine the statistical properties of regional activation. Although this ROI-based analysis approach does not localize the active voxels in the brain as in voxel-based approach, it addresses a more hypothesis-driven question of whether a particular brain region is activated. The advantage of this approach is that no spatial distortions are introduced to the original pattern of activation. Furthermore, this approach was shown to enhance sensitivity to task-related activation detection compared to whole-brain warping [14]. However, designing methods that optimally combine voxel information within an ROI remains a challenge. Conventionally, mean activation statistics and percentage of activated voxels have been used [15,16], which implicitly assume that only the amplitude of the fMRI signals is modulated by task. Yet, in earlier work [17,18], we have demonstrated that using invariant spatial features to characterize the spatial pattern of activation within an ROI can in fact discriminate task-related activation differences. Hence, employing invariant spatial features under an ROI-based approach provides a means to analyze the spatial distribution of activation without the need to functionally align the subjects. In this paper, we examine the effects of warping on spatial fMRI analysis by comparing the task discriminability of an invariant spatial feature for ROIs extracted 1) from the subject’s native space, 2) using whole-brain warping, and 3) using region-level warping. Qualitative analysis is presented through 3D visualization of the ROI activation maps to highlight the effects of warping.

2. EXPERIMENTAL SETUP AND MATERIALS After obtaining informed consent, fMRI data were collected from ten healthy subjects (3 men, 7 women, mean age 57.4 ± 14 years) during a visual tracking motor task. Each subject was required to use their right hand to squeeze a bulb with sufficient pressure such that a black horizontal bar shown on a screen was kept within an undulating pathway. The pathway remained straight during baseline periods, which required a constant pressure to be applied, and became sinusoidal during the time of stimulus at a frequency of 0.25 Hz (slow), 0.5 Hz (medium) or 0.75 Hz (fast) presented in a pseudo-random order. Each run lasted 260 s, alternating between rest and stimulus of 20 s duration. The data were collected as part of a larger experiment for exploring the effects of the rate of change in force production on older subjects and subjects with Parkinson’s disease. For the purposes of the current work, only the fast and slow conditions were compared, where changes in brain activation patterns were expected.

2.1 fMRI data acquisition

Functional MRI was performed on a Philips Gyroscan Intera 3.0 T scanner (Philips, Best, the Netherlands) equipped with a head-coil. T2*-weighted images with blood oxygen level dependent (BOLD) contrast were acquired using an echo-planar (EPI) sequence with a repetition time of 1985 ms, echo time of 3.7 ms, flip angle of 90°, field of view (FOV) of 216×143×240 mm, in plane resolution of 128×128 pixels, and pixel size of 1.9×1.9 mm. Each functional run lasted 4 minute, where 36 axial slices of 3 mm thickness were collected in each volume with a gap thickness of 1 mm. We selected slices to cover the dorsal surface of the brain and included the cerebellum ventrally. A high resolution 3D T1-weighted image consisting of 170 axial slices was acquired of the whole brain to facilitate anatomical localization of activation for each subject.

2.2 fMRI preprocessing

The fMRI data was pre-processed for each subject using Brain Voyager’s (Brain Innovation B.V.) trilinear interpolation for 3D motion correction and sinc interpolation for slice time correction. Further motion correction was performed using motion corrected independent component analysis (MCICA) [19]. No temporal or spatial smoothing was performed on the data. To co-register the anatomical and functional images, the Brain Extraction Tool (BET) in MRIcro [20] was used to strip the skull off of the anatomical and first functional image from each run to enable a more accurate alignment of

Proc. of SPIE Vol. 7262 72621Y-2

the functional and anatomical scans. Custom scripts in Amira (Mercury Computer Systems, San Diego, USA) were then used to co-register the anatomical and functional images. To generate the activation statistics map for each subject, a hybrid Independent Component Analysis (ICA)-General Linear Model (GLM) scheme [21] was employed to contrast each of the three frequency blocks with the constant force periods. The underlying idea of this scheme was that the GLM regressors (reference signal and confounds) were derived using ICA, which would be subject-specific.

3. METHODS The goal of this work was to analyze the effects of different warping approaches on the spatial analysis of fMRI data. Four motor-related ROIs, namely the bilateral thalami and the bilateral cerebellar hemispheres, were chosen for this analysis due to their appreciable size differences and the fact that these ROIs were expected to incur task-related activation changes. To quantitatively assess the effects of warping, we extracted ROI activation statistics from the whole-brain activation maps using three different procedures discussed in Section 3.1. We then applied our previously proposed spatial analysis method [17,18] to the ROI activation statistics as discussed in Section 3.2. To visualize the effects of warping, we also generated 3D rendering of the ROI activation maps.

3.1 ROI activation statistics extraction

ROI activation statistics were extracted from each subject’s whole-brain activation statistics map as per the following procedures (Fig. 1).

Warp-free: Each ROI was manually segmented from the whole-brain anatomical image by an expert based on anatomical landmarks and guided by a neurological atlas [7]. The segmented ROI was then used to extract the corresponding ROI activation statistics from the whole-brain activation statistics map without warping to any template.

Whole-brain warping: The whole-brain anatomical image was warped to the Montreal Neurological Institutes (MNI) template using SPM 5 [22]. The same warp was then applied to the whole-brain activation statistics map with each ROI’s activation statistics extracted using predefined labels associated with the MNI template.

Region-level warping: The segmented anatomical ROIs from the warp-free procedure were rigidly aligned and averaged across subjects to generate an ROI template for each brain structure to be analyzed. This method of generating an ROI template is similar to the approach proposed by Stark and Okado [11], which presumably provides a template that more closely matches the ROI shapes of the subjects under study. The anatomical ROI of each subject was then warped onto the ROI template with the same warp applied to the ROI activation statistics.

3.2 Invariant spatial features

For each experimental condition and each subject’s ROI (e.g. slow, left thalamus without warping), we first scaled the spatial coordinates (x,y,z) to account for ROI size differences across subjects [23]:

∑∈

−+−+−====ROIi

iiisss zzyyxxsszzsyysxx ))()()((,/,/,/ 222 , (1)

where x , y , and z are the centroid coordinates of the ROI and (xs,ys,zs) are the scaled spatial coordinates. We then calculated an invariant spatial feature that characterized the spatial extent of ROI activation [17,18]:

0020202001 μμμ ++=J , (2)

∫∫∫ −−−=ROI

ssssssr

ssq

ssp

sspgq dzdydxzyxzzyyxx ),,()()()( ρμ , (3)

where n = p+q+r is the order of the centralized 3D moment, μpqr, ρ(xs,ys,zs) is the t-value of a voxel located at (xs,ys,zs), and sx , sy , and sz are the centroid coordinates of ρ(xs,ys,zs). Only positive t-values were used due to the presently unclear interpretation of negative t-values [24], and the t-values were normalized to the range of [0,1] to decouple the effects of amplitude from spatial changes [17]. We note that the resulting spatial feature, J1, would be invariant to scaling (1), rotation (2), and translation (3), which accounts for differences in ROI size and subject orientation in the scanner.

Proc. of SPIE Vol. 7262 72621Y-3

ahi s*d's nS. FaA*

1.0

D*i. NOIs hi sit4scl'snSa e'

2. Ap ROl naSa to udrSNOl aan '-'Y'o.

'I

I I I

Fig. 1. Extraction of ROI activation statistics. (a) Under the warp-free scheme, ROI activation statistics were extracted for

each subject by masking out the appropriate regions from the whole-brain activation map using manually defined labels. (b) The whole-brain warping scheme involved warping the anatomical brain image of each subject to a template, applying the same warp to the whole-brain activation map, and extracting the ROI activation statistics using the pre-defined labels of the template. (c) The region-level warping scheme required creating an ROI template and warping of each subject’s ROI to this template with the same warp applied the corresponding activation statistics.

Proc. of SPIE Vol. 7262 72621Y-4

A paired t-test was then applied to the ROI features to contrast the effects of different frequencies (e.g. fast vs. slow, left thalamus without warping). Significance was declared at a family-wise critical p-value of 0.05 (with Bonferroni correction to account for the number of ROIs). The corrected critical p-value was 0.0127 and the corresponding t-value was 3.10. To demonstrate the presence of important spatial information in the ROI activation statistics, we also calculated the mean ROI activation statistics (without warping) for comparison.

4. RESULTS Employing the traditionally-used mean activation statistics detected no significant task-related activation differences in any of the ROIs (left thalamus t-value = 0.213, right thalamus t-value = 0.160, left cerebellum t-value = 1.613, and right cerebellum t-value = 0.356). In contrast, the invariant spatial feature, J1, without warping detected significant activation differences in all four ROIs as summarized in Fig. 2, thus suggesting that important task-related information was encoded in the spatial distribution of the ROI activation statistics. J1 with region-level warping detected the bilateral thalami and the left cerebellum but not the right cerebellum. J1 with whole-brain warping detected significance only in the left thalamus and at reduced discriminability compared to J1 without warping.

0

1

2

3

4

5

6

left thalamus right thalamus left cerebellum right cerebellumROIs

t-val

ue

Warp-free analysis Region-level warping Whole-brain warping

t-threshold at 3.10

Fig. 2. Discriminability comparisons between the contrasted procedures for extracting ROI activation statistics. The

horizontally (blue), diagonally (magenta), and vertically (yellow) hatched-bars correspond to warp-free analysis, region-level warping, and whole-brain warping, respectively. The green dash line is the t-threshold for a family-wise p-value of 0.05 with Bonferroni correction (corrected t-value = 3.10). Significant activation differences were detected in all ROIs for J1 in the warp-free case. Region-level warping resulted in reduced discriminability (except in the left thalamus) and a loss of significance in the right cerebellar hemisphere. Whole-brain warping led to further loss of significance in the right thalamus and the bilateral cerebellar hemispheres.

5. DISCUSSION The quantitative results in Fig. 2 suggest that the invariant spatial feature, J1, was by far more discriminant to task-related spatial changes under the warp-free settings than after whole-brain warping. J1 without warping also appeared more discriminant than after region-level warping but to a lesser extent. To determine the reasons for the differences in discriminability, we generated 3D renderings of the ROI activation statistics under each of the three ROI activation statistics extraction schemes to visualize the effects of warping.

5.1 Effects of warping on smaller ROIs

The top two rows of Fig. 3 show the 3D rendering of the warp-free ROI activation statistics within the right thalamus under the slow condition for all ten subjects. As apparent from the figure, the locations of the more highly activated voxels varied substantially across subjects and similar level of functional inter-subject variability was also observed for the fast condition (Fig. 4). Hence as an aside, even if the ROIs were anatomically aligned, group activation differences might be difficult to detect using conventional voxel-based approach [22] due to the lack of functional overlap across

Proc. of SPIE Vol. 7262 72621Y-5

subjects. Instead, comparing Figs. 3 and 4, a wider area of the right thalamus appeared to be consistently recruited in all subjects under the fast condition. Thus, measuring the spread of ROI activation in a pose invariant manner as we previously proposed in [17,18] might provide higher sensitivity than voxel-based approach for the type of activation changes observed.

One might argue that the lack of functional overlap was due to ROI mis-alignments, and conventional whole-brain warping should be performed before any conclusions could be drawn. The middle two row of Fig. 3 show the activation maps of the right thalamus extracted using the whole-brain warping scheme described in Section 3.1. Consistent with Crivello et al.’s findings [10], the amount of functional overlap did not seem to increase after whole-brain warping. Worse yet, whole-brain warping appeared to have pooled neighboring regions as part of the right thalamus. For instance, the more activated voxels were originally concentrated on the left side of the right thalamus in Fig. 3 (f), but an extra cluster emerged in the lower right region after whole-brain warping in Fig. 3 (p). This mis-registration problem was more pronounced in Fig. 3 (t), where the entire upper region of the right thalamus appeared to be active when only a few voxels may actually be activated (Fig. 3 (j)). Moreover, as evident by comparing Fig. 3 (c) and (m), whole-brain warping might have so heavily distorted the ROI activation patterns that the patterns no longer resemble the original data. Since both problems of mis-registration and spatial distortions were inherent in the whole-brain warping process, similar artifacts were found in the ROI activation maps for the fast condition (middle two rows of Fig. 4) and in the left thalamus (Fig. 7). Thus, although significant activation differences were detected in the left thalamus under whole-brain warping scheme, results should be interpreted with caution since the detected differences could have partly arise from the pooling of neighboring active regions as opposed to the true differences observed in the warp-free data.

A plausible explanation for the observed mis-registration and spatial distortions under the whole-brain warping scheme is that the thalamus is smaller than many other regions in the brain, hence may have little influence on the warping process. One way to reduce the amount of mis-registration is to perform region-level warping, which mitigates the problem of pooling neighboring regions. The bottom two rows of Figs. 3 and 4 show the activation statistics of the right thalamus after region-level warping. Compared to the original warp-free data (top two rows of Figs. 3 and 4), the overall spatial patterns appeared to be conserved although slightly distorted and shifted. The amount of functional inter-subject variability also seemed to have mildly decreased, but this gain might have been offset by the smearing and spatial distortions introduced, which led to reduced discriminability (Fig. 2). For the case of the left thalamus, the decrease in functional variability actually appeared to have outweighed the artifacts introduced, though this trend was not consistent for the other ROIs examined. Nevertheless, compared to whole-brain warping, a reduced amount of mis-registration and spatial distortions was definitely evident, thus the increase in discriminability.

5.2 Effects of warping on larger ROIs

The top two rows of Figs. 5 and 6 show the warp-free activation statistics of the right cerebellum for the slow and fast conditions, respectively. As can be seen from the figures, considerable functional inter-subject variability was present in the spatial activation pattern. Also, comparing Figs. 5 and 6, the spread of activation only mildly increased under the fast condition. For instance, only the middle region of the right cerebellum (Fig. 5 (e) and (j)) became more active (Fig. 6 (e) and (j)). Thus, any distortions introduced by warping might remove these subtle spatial differences.

Based on the inherent nature of whole-brain warping, the cerebella should theoretically be less mis-aligned than the thalami, since larger ROIs have more influence on the warping process. Examining the middle two rows of Fig. 5 supports this hypothesis. Although spatial distortion was apparent and neighboring regions might have been pooled into the border of the right cerebellum (e.g. top of right cerebellum in Fig. 5 (m)), correspondence in the overall spatial pattern could roughly be drawn between the warp-free and whole-brain warped activation maps.

Region-level warping again resulted in less spatial distortions than whole-brain warping as shown in the bottom two rows of Figs. 5 and 6. Correspondence in the overall spatial patterns could easily be identified between the warp-free and region-level warped cases (e.g. Fig. 5 (j) and (η)), but ample amount of smearing was clearly present (e.g. Fig. 5 (i) and (γ)). Since only mild spatial differences were observed between the slow and fast conditions in the original warp-free activation maps, this smearing effect seemed to have reduced task discriminability, leading to a loss of significance.

Proc. of SPIE Vol. 7262 72621Y-6

k1

I SI LI I

(a)

(b)

(c)

(d)

(e)

(f)

(g)

(h)

(i)

(j)

(k)

(l)

(m)

(n)

(o)

(p)

(q)

(r)

(s)

(t)

(u)

(v)

(w)

(x)

(y)

(z)

(α)

(β)

(γ)

(η)

Fig. 3. ROI activation statistics of the right thalamus under the slow condition. The top two rows correspond to warp-free ROI activation statistics of the ten subjects in this study. The middle and bottom two rows correspond to activation statistics extracted after whole-brain warping and region-level warping. Region-level warping appeared to have better preserved the spatial patterns with less mis-registration and spatial distortions compared to whole-brain warping.

Proc. of SPIE Vol. 7262 72621Y-7

IL]

S SI I Li

(a)

(b)

(c)

(d)

(e)

(f)

(g)

(h)

(i)

(j)

(k)

(l)

(m)

(n)

(o)

(p)

(q)

(r)

(s)

(t)

(u)

(v)

(w)

(x)

(y)

(z)

(α)

(β)

(γ)

(η)

Fig. 4. ROI activation statistics of the right thalamus under the fast condition. The top, middle, and bottom two rows correspond to ROI activation statistics extracted warp-free, after whole-brain warping, and after region-level warping, respectively. A wider area of the right thalamus seemed to be recruited during the fast condition as compared to Fig. 3.

Proc. of SPIE Vol. 7262 72621Y-8

C 0

(a)

(b)

(c)

(d)

(e)

(f)

(g)

(h)

(i)

(j)

(k)

(l)

(m)

(n)

(o)

(p)

(q)

(r)

(s)

(t)

(u)

(v)

(w)

(x)

(y)

(z)

(α)

(β)

(γ)

(η)

Fig. 5. ROI activation statistics of the right cerebellum under the slow condition. The top, middle, and bottom two rows correspond to ROI activation statistics extracted warp-free, after whole-brain warping, and after region-level warping, respectively. Spatial distortions were again introduced by whole-brain warping, but to a lesser extent compared to the case of the right thalamus, since larger ROIs have more influence on the warping process. Region-level warping appeared to have better preserved the overall spatial patterns, but a substantial amount of smearing was evident.

Proc. of SPIE Vol. 7262 72621Y-9

4

(a)

(b)

(c)

(d)

(e)

(f)

(g)

(h)

(i)

(j)

(k)

(l)

(m)

(n)

(o)

(p)

(q)

(r)

(s)

(t)

(u)

(v)

(w)

(x)

(y)

(z)

(α)

(β)

(γ)

(η)

Fig. 6. ROI activation statistics of the right cerebellum under the fast condition. The top, middle, and bottom two rows correspond to ROI activation statistics extracted warp-free, after whole-brain warping, and after region-level warping, respectively. A subtle increase in the spread of activation was observed under the fast condition.

Proc. of SPIE Vol. 7262 72621Y-10

*1

(a)

(b)

(c)

(d)

Fig. 7. ROI activation statistics of the left thalamus of an exemplar subject. (a) Warp-free ROI activation statistics under the slow condition. (b) Whole-brain warped ROI activation statistics under the slow condition. (c) Warp-free ROI activation statistics under the fast condition. (d) Whole-brain warped ROI activation statistics under the fast condition. Whole-brain warping appeared to have heavily smeared and distorted the spatial activation patterns.

6. CONCLUSIONS In this paper, we have quantitatively demonstrated that retaining the original spatial information can offer higher discriminability in ROI-based spatial fMRI analysis compared to performing whole-brain or region-level based warping. The intended purpose of whole-brain warping is to align the functional regions across subjects, which presumably reduces inter-subject variability. However, both our quantitative and qualitative analyses showed that whole-brain warping may in fact drastically reduce the discriminability of task-related activation differences. Due to the inherent nature of the whole-brain warping process, mis-registration and spatial distortions appeared to be more pronounced in smaller ROIs than larger ROIs. Region-level warping also distorts the spatial pattern of activation but to a much lesser degree. Also, region-level warping mitigates the problem of pooling neighboring regions, hence would be preferrable over whole-brain warping if spatial analysis is to be performed. Nevertheless, for ROI-based spatial characterization, warp-free analysis appears to provide the highest discriminability based on our results.

REFERENCES

[1] Norman, K. A., Polyn, S. M., Detre, G. J. and Haxby, J. V., "Beyond mind-reading: multi-voxel pattern analysis of fMRI data," Trends Cogn. Sci. 10, 424-430 (2006).

[2] Haynes, J. D. and Rees, G., “Decoding mental states from brain activity in humans,” Nat. Rev. Neurosci. 7(7), 523-534 (2006).

[3] Uylings, H. B. M., Rajkowska, G., Sanz-Arigita, E., Amunts, K. and Zilles, K., “Consequences of large interindividual variability for human brain atlases: converging macroscopical imaging and microscopical neuroanatomy,” Anat. Embryol. 210, 423-431 (2005).

[4] Samanez-Larkin, G. R. and D’Esposito, M., “Group comparisons: imaging the aging brain,” Soc. Cogn. Affect. Neurosci. 3(3), 290-297 (2008).

[5] Wilke, M., Schmithorst, V. J. and Holland, S. K., “Assessment of Spatial Normalization of Whole-Brain Magnetic Resonance Images in Children,” Hum. Brain Mapp. 17(1), 48-60 (2002).

[6] Brett, M., Leff, A. P., Roden, C. and Ashburner, J., “Spatial Normalization for Brain Images with Focal Lesions Using Cost Function Masking,” NeuroImage 14(2), 486-500 (2001).

[7] Talairach, J. and Tournoux, M., [Co-Planar Stereotaxic Atlas of the Human Brain], Thieme Medical Publishers Inc., New York (1988).

[8] Evans, A. C., Collins, D. L. and Milner, B., “An MRI-based stereotactic atlas from 250 young normal subjects,” Soc. Neurosci. Abstr. 18, 408 (1992).

[9] Brett, M., Johnsrude, I. S. and Owen, A. M., “The problem of functional localization in the human brain,” Dementia 383, 707-710 (1996).

[10] Crivello, F., Schormann, T., Tzourio-Mazoyer, N., Roland, P. E., Zilles, K. and Mazoyer, B. M., “Comparison of spatial normalization procedures and their impact on functional maps,” Hum. Brain Mapp. 16, 228-250 (2002).

[11] Stark, C. E. and Okado, Y., “Making memories without trying: medial temporal lobe activity associated with incidental memory formation during recognition,” J. Neurosci. 23, 6748-6753 (2003).

Proc. of SPIE Vol. 7262 72621Y-11

[12] Miller, M. I., Beg, M. F., Ceritoglu, C. and Stark, C., “Increasing the power of functional maps of the medial temporal lobe by using large deformation diffeomorphic metric mapping,” in Proc. Natl. Acad. Sci. 102, 9685-9690 (2005).

[13] Yushkevich, P. A., Detre, J. A., Mechanic-Hamilton, D., Fernández-Seara, M. A., Tang, K. Z., Hoang, A., Korczykowski, M., Zhang, H. and Gee, J. C., “Hippocampus-specific fMRI group activation analysis using the continuous medial representation.” NeuroImage 35, 1516-1530 (2007).

[14] Nieto-Castanon, A., Ghosh, S. S., Tourville, J. A. and Guenther, F. H., “Region of interest based analysis of functional imaging data,” NeuroImage 19, 1303–1316 (2003).

[15] Constable, R. T., Skudlarski, P., Menci, E., Pugh, K. R., Fulbright, R. K., Lacadie, C., Shaywitz, S. E. and Shaywitz, B. A., “Quantifying and Comparing Region-of-Interest Activation Patterns in Functional Brain MR Imaging: Methodology Considerations,” Magn. Reson. Imaging 16(3), 289-300 (1998).

[16] Buck, R., Singhal, H., Arora, J., Schlitt, H. and Constable, R. T., “Detecting change in BOLD signal between sessions for atlas-based anatomical ROIs,” NeuroImage 40, 1157-1165 (2008).

[17] Ng, B., Abugharbieh, R., Huang, X. and McKeown, M. J., “Spatial Characterization of fMRI Activation Maps Using Invariant 3D Moment Descriptors,” IEEE Trans. Med. Imaging, (In press)

[18] Ng, B., Abugharbieh, R., Huang, X. and McKeown, M. J., “Characterizing fMRI Activations within Regions of Interest (ROIs) Using 3D Moment Invariants,” in Proc. IEEE Workshop on Mathematical Methods in Biomedical Image Analysis, New York, NY, 63 (2006).

[19] Liao, R., Krolik, J. L. and McKeown, M. J., “An information-theoretic criterion for intrasubject alignment of FMRI time series: motion corrected independent component analysis,” IEEE Trans. Med. Imaging 24(1), 29-44 (2005).

[20] Rorden, C. and Brett, M., “Stereotaxic display of brain lesions,” Behav. Neurol. 12, 191-200 (2000). [21] McKeown, M. J., “Detection of consistently task-related activations in fMRI data with hybrid independent

component analysis,” NeuroImage 11, 24-35 (2000). [22] Frackowiak, R. S. J., Friston, K. J., Frith, C., Dolan, R., Price, C. J., Zeki, S., Ashburner, J. and Penny, W. D.,

[Human Brain Function 2nd ed.] Academic Press (2003). [23] Goodall, C., “Procrustes methods in the statistical analysis of shape,” J. R. Statist. Soc. B 53, 285-339 (1991). [24] Harel, N., Lee, S. P., Nagaoka, T., Kim, D. S. and Kim, S. G., “Origin of Negative Blood Oxygenation Level-

Dependent fMRI Signals,” J. Cereb. Blood Flow Metab. 22(8), 908-917 (2002).

Proc. of SPIE Vol. 7262 72621Y-12