Voxel-Based Morphometric Approach to Modeling Stroke Lesions in Diffusion Weighted Images

(c) 2013 Moo K. Chung, Dong-Eog Kim

Department of Biostatistics and Medical Informatics
Waisman Laboratory for Brain Imaging and Behavior
University of Wisconsin-Madison

Dongguk University Ilsan Hospital, Korea

August 14, 2013

We present a streamlined pipeline for quantifying a collection of stroke lesion images in diffusion weighted images (DWI). Although Bernoulli models are often used in modeling a collection of binary images, the Bernoulli models actually break down for testing statistical significance of common overlaps. To remedy the limitation of the Bernoulli models, we propose to adapt the random field theory often used in voxel-based morphometry(VBM) to stroke lesion images. If you are using the Matlab codes/sample data below for your publication, please reference [4] in the Referenes. The codes have been tested under Matlab R2009b  64bit version in Macbook OS X 10.6.8.

For bug reports, email mkchung@wisc.edu

Stroke Lesions Binary Segmentaion Data
August 14, 2013

Stroke lesions are semiautomatically segmented and saved as binary images of size 370 by 301 in stroke.mat. The first group (group1) consists of 58 subjects and the second group (group2) consists of 23 subjects (Figure 1). For the detailed description of the groups, see [4]. The Matlab script for the codes below is given in stroke-v1.m

Figure 1. The first four columns are images of segmented infarct lesions in acute (less than 7 days) stroke patients whose dysphagia was improved after one month (n = 58). The last four columns are images of stroke patients whose dysphagia was not improved after one month (n = 23). We are only showing a subset of the whole data. We are interested in determining if there is any common region where stroke lesions consistently occur.

Voxel-Based Morphometry on Lesions Segmentaion
August 14, 2013

We perform the two sample test on the equality of the proportions of stroke lesions (Figure 2):

[pvalue Zvalue] = stat_proportion(group1, group2);

Figure 2. The statistical significance of the equality of the proportions of stroke lesions between the two groups.

The problem with the above approach is the lack of spatial smoothing that inflates a lot of small speckles of false positives that are caused by discretization errors. After smoothing with 10 pixelwide FWHM, we can signifcantly redue such discretizatione artifacts (Figure 2). Smoothed version is done as follows.

group1smooth= stat_smooth(group1, FWHM);
group2smooth=stat_smooth(group2, FWHM);
[pvalue Zvalue] = stat_proportion(group1smooth, group2smooth);

Figure 3.
The statistical significance of the equality of the proportions of stroke lesions between the two groups after Gaussian kernel smoowhing with 10 pixelwide FWHM.

However, after multiple comparsions correction based on the random field theory, the signal in Figure 3 wil be completely disappear (Figure 4). The random field theory based multiple comparions will be done using

pcorrected=stat_Zcorrected2D(Zvalue,  FWHM);
figure; imagesc(pcorrected); colormap('hot');
colorbar; caxis([0 0.1])

Figure 4. The statistical significance after the random field theory based multiple comparisons.

References on Voxel-Based Morphometry
August 14, 2013
  1. Chung, M.K., Dalton, K.M., Alexander, A.L., Davidson, R.J. 2004.  Less white matter concentration in autism: 2D voxel-based morphometry. NeuroImage 23:242-251.
  2. Chung, M.K., Shen, L., Dalton, K.M., Davidson, D.J. 2006. Multi-scale voxel-based morphometry via weighed spherical harmonic representation. International Workshop on Medical Imaging and Augmented Reality (MIAR). Lecture Notes in Computer Science (LNCS). 4091:36-43.
  3. Oakes, T.R., Fox, A.S., Johnstone, T., Chung, M.K., Kalin, N., Davidson, R.J. 2007. Integrating VBM into the general linear model with voxelwise anatomical covariates. NeuroImage 34:500-508.
  4. Chung, M.K., Woo, S.H., Lee, J. S., Kim, D.-E. 2013. Modeling stroke lesions in diffusion weighted images using 2D random field theory, SPIE Medical Imaging. submitted.