2009 Grace Tang & Kelly Hennigan

Back to Psych 204 Projects 2009

Introduction

Within neuroimaging literature, a prevalent procedure for coregistration is SPM's standard normalization procedure. For our project, we examined the spatial accuracy of this normalization method with respect to the nucleus accumbens (NAc), a structure that is frequently examined in reward, addiction and decision making studies. Masks of the left and right nucleus accumbens were manually segmented in native space for ten individuals and then normalized into a common space. Parameters of variance are reported and accuracy is assessed. Theoretically, once normalized, masks from distinct individuals should agree perfectly, with each other and with established templates such as those defined by the Montreal Neurological Institute. It is worthwhile to examine the accuracy of this assumption, and the extent of inter-individual variation, since templates such as these and group averaged images are used widely to define the location of the nucleus accumbens across subjects.

Methods

Image acquisition

Structural T1 images were obtained for ten individuals with a 3T Siemens scanner at Baylor College of Medicine in Texas.

Data Pre-processing stream

Manual Segmentation

Manual segmentation was conducted using ITKSnap software (Yushkevich et al., 2006). Anatomical characteristics of the NAc were used to identify and mask the NAc in each subject's native T1 image individually. In the axial plane, the point at which the caudate and putamen elements converge was taken to be the dorsal boundary of the NAc. From posterior to anterior in the coronal plane, the NAc is defined as the nuclei that appears under the anterior limb of the internal capsule. Figure 1 shows a mask created for the right NAc of a sample subject in all three planes (Neto et al., 2008). Voxels within a mask were set to “1” and voxels outside a mask were set to “0”.

Normalization

Structural images were then normalized to the Montreal Neurological Institute (MNI) ICBM152 template (Talairach and Tournoux, 1988) by applying a 12-parameter affine transformation followed by non-linear warping using basis functions in SPM8 (Ashburner, 2009). This produced a transform matrix unique for each individual. We then applied the transform matrix to the appropriate NAc mask, effectively normalizing the NAc masks into a common space.

Figure 1

Fig 1a: Mask in axial plane in one subject

Fig 1b: Mask in coronal plane

Fig 1c: Mask in sagittal plane

Fig. 1: Screenshot of the normalized NAc mask in the axial, coronal and sagittal plane for one sample subject.

Data Analysis

Normalized NAc masks were then superimposed to determine the extent to which the ten individual masks overlap. Matrices of the individual masks were summed together using Matlab (R2009b). Since areas within each mask were coded as 1, voxels in the summed matrix with, for example, three overlapping masks would be represented with the value 3, giving an index of overlap.

Comparison with SPM template

To allow comparison with the single representative subject brain provided by SPM, T1 images and segmentation masks were renormalized in SPM with the bounding box specified to match that of the representative brain image. Voxel size was also changed to match the representative brain's voxel size of 2x2x2.

A segmentation mask was defined for the single representative subject brain provided by SPM using the same procedure as above. The mean index of overlap in the voxels within the representative mask volume was computed in order to give an idea of the extent to which the masks defined in our study's ten individual subjects agreed with the mask defined in the representative brain. The mean index of overlap outside the representative mask volume was computed as well to represent the extent to which the individually defined masks did not agree with the representative brain.

The number of voxels in the summed matrix (in which individual matrices from all ten subjects were added together) containing at least one subject's NAc (coded 1 in at least one subject, giving rise to a non-zero value in the summed matrix) within and outside the representative mask was also noted. We refer to voxels that are included in at least one subject's mask as NAc voxels.

Smoothing

Gaussian smoothing was performed on the masks in Matlab. The smoothed masks were then rethresholded such that post-smoothing values equal to or larger than 0.5 were coded as 1, and below 0.5 , 0. Individual smoothed and rethresholded masks were then summed together and viewed in ITKSnap.

This was repeated with three sizes of convolution kernel: 1x1x1, 3x3x3 and 5x5x5 (voxels were 1x1x1mm), and a lower threshold, 0.3.

Results

Location

The group mean center foci of the right and left NAc masks were identified as (17, 7, -13) and (-14, 7, -13) in MNI coordinates, which is comparable to other published ROI coordinates for the NAc in neuroimaging studies (Berns et al., 2001, Knutson, 2008). Figure 1 gives a visual representation for a sample subject at these coordinates.

Overlap Index

Masks of right and left NAc contained an average of 2,491 voxels/subject in normalized MNI space. Figure 2 shows the NAc masks in normalized space overlaid on a sample subject’s T1 structural image. The color coding represents the overlap index described in the methods section: distinct colors denote the number of subject NAc masks included in each voxel, e.g. the outermost purple colored band denotes voxels containing one subject's mask, the second dark blue band has an overlap index of 2, the cyan band has 3, and so on (see fig. 2a).

Figure 2

Fig 2a: Overlap in the nucleus accumbens

Fig. 2b: overlap between individual masks in axial slices (moving from dorsal to ventral)

Figures 3 & 4

Figure 3 below show the total number of voxels (included in at least one mask) with corresponding overlap index in the x-axis. The majority of voxels contain only one mask, and the number of voxels containing a x number of overlapping masks decreases with x in an almost exponential manner. Figure 4 represents the overlap index as a cumulative histogram.

Figure 3

Figure 4

Data for Adjusted Voxel Size

To see if resolution would effect the overlap index, we also transformed the masks from native to normalized space with a voxel size of 2x2x2 mm. This lower resolution had no impact on the proportional number of voxels captured by the masks (they contained an average of 310 voxels/subject, which is proportionally equivalent to the average mask size with the smaller voxel size) nor the profile of overlap within the index. Figure 5 below shows relative histograms of the overlap index at both voxel sizes.

Figure 5

Masks for 1x1x1 voxel size

Masks for 2x2x2 voxel size

Fig 5 - images of all subjects NAc masks with regular and double voxel size in matched axial plane.

Figure 6

1x1x1 mm voxels

2x2x2 mm voxels

Comparison with SPM template

The number of non-zero voxels from the summed matrix that fell inside the representative mask volume was 266 (the size of the representative mask volume is 266 - every voxel within the representative mask overlapped with at least one individual subject brain), while the number of non-zero voxels that fell outside the representative mask was 848. While this at first glance appears to show poor agreement with the representative mask, closer inspection of the index of overlap within and outside the representative mask indicates otherwise. The mean index of overlap within the representative mask was 5.0, while the mean outside the mask was 2.088. An independent samples t-test showed that the index of overlap was significantly higher within the representative mask (p<0.01).

Smoothing

Mask size

An effect of the smoothing procedure was to increase the size of the individual masks. This increased 'spread' should increase the overlap between masks of individual subjects.

Smoothing with a kernel size of 1x1x1 made little difference compared to not smoothing at all, regardless of threshold. However, using a smoothing kernel of 3x3x3 at a threshold of 0.3 increased the number of voxels within the mask. The effect was not apparent when the same smoothing kernel was used with a threshold of 0.5. Changing the smoothing kernel from 3x3x3 to 5x5x5 did not make a big difference at either threshold.

[1 1 1], 0.3

[1 1 1], 0.5

[3 3 3], 0.3

[3 3 3], 0.5

[5 5 5], 0.3

[5 5 5], 0.5

Histograms

• • • • CHECK... [1 1 1] 05 is a placeholder

[1 1 1], 0.3

[1 1 1], 0.5

[3 3 3], 0.3

[3 3 3], 0.5

[5 5 5], 0.3

[5 5 5], 0.5

Relative Histograms

Relative histograms can be used to investigate changes in overlap while controlling for increases in mask size. A shift to the right (i.e. a higher proportion of voxels with higher overlap index) in relative histograms implies an improvement of overlap that is not due to an overall increase in number of voxels per mask.

With smoothing kernels of 3x3x3 and 5x5x5, there is a shift to the right when the threshold is lowered from 0.5 to 0.3, indicating better overlap. Changing the threshold does not make a difference when using a smoothing kernel of 1x1x1.

[1 1 1], 0.3

[1 1 1], 0.5

[3 3 3], 0.3

[3 3 3], 0.5

[5 5 5], 0.3

[5 5 5], 0.5

Summary

Though there was some variability in the position of NAc across subjects, as to be expected, we showed that there was a sizeable extent of agreement between subjects, both with each other and with the representative template brain. Smoothing also improved the overlap between masks. That said, it is worthwhile to note that there was no one voxel in which all individual masks overlapped, even with a relatively small sample size of ten, indicating a fairly large variance. This implies that while group averaged images or standard templates may be sufficient in representing the nucleus accumbens across subjects, defining the location of the nucleus accumbens in individual subjects' native space may improve spatial localization, and therefore the power of analyzing functional activity.

Citations

Ashburner, J. (2009). Computational anatomy with the SPM software. Magnetic Resonance Imaging (27)8, 1163-1174.

Berns, G.S., McClure, S.M., Pagnoni1, G., and Montague, P.R. (2001). Predictability Modulates Human Brain Response to Reward. The Journal of Neuroscience, 21(8):2793-2798.

Knutson, B., Wimmer, G., Rick, S., Hollon, N., Prelec, D., & Loewenstein, G. (2008). Neural antecedents of the endowment effect. Neuron, 58(5), 814-822.

Neto, L.L., Oliveira, E., Correia F., Ferreira, A.O. (2008). The Human Nucleus Accumbens: Where Is It? A Stereotactic, Anatomical and Magnetic Resonance Imaging Study. Neuromodulation(11)1, 13-22.

Yushkevich, P.A., Piven, J., Hazlett, H.C., Smith, R.G., Ho, S., Gee, J.C., and Gerig, G. (2006). User-guided 3D active contour segmentation of anatomical structures: Significantly improved efficiency and reliability. Neuroimage 31(3):1116-28.

Appendix

• Code
• More Code
• Smoothing Code