Appendix (E) A-1 Data Analysis Methods

Appendix (E) A-1 Data Analysis Methods

for the article

Changes in DWI and MRS associated with white matter hyperintensities in elderly subjects

Contact:

The basic steps of the analysis were

a) segmentation of the T1 weighted and FLAIR data to produce images showing the distribution of CSF, grey and white matter and WMH, and to calculate the total WMH volume.

b) resampling of the segmented images to the resolution of the spectroscopy data and linear regression between metabolite ratios and tissue content in the spectroscopic voxels to derive values for mean metabolite ratios in the grey matter, NAWM and WMH over the whole spectroscopic image.

c) use of segmented images to locate NAWM and WMH on the diffusion weighted image, and thus obtain values for the global median ADC for NAWM and WMH.

We now describe these steps in some more detail

Segmentation of grey and white matter.

We used SPM99 (http://www.fil.ion.ucl.ac.uk/spm/) to process the 1mm voxel T1 weighted data. The images were spatially normalized to the T1 weighted MNI (Montreal Neurological Institute, http://www.bic.mni.mcgill.ca ) template that approximates the space defined by Talairach and Tournoux 1. For spatial normalization, a 12-parameter affine transformation was used to match each image volume to the mean template image2. This was followed by nonlinear iterations using 7 x 8 x 7 basis functions to account for global nonlinear shape differences 3. Normalized images were then resampled by bilinear interpolation to a voxel size of 2 x 2 x 2 mm3

Following this normalization, images were automatically segmented, based on prior probability maps of the relative distribution of tissue types, into grey matter (GM), white matter (WM), cerebrospinal fluid (CSF), and a fourth partition comprising skull, fat muscle, and voxels which have a high degree of partial voluming and thus can not be included in one of the other three classes 4. The brain was then extracted by a fully automated procedure for removing nonbrain tissue and scalp from the segmented gray matter images. This procedure (a function of SPM99) involves a series of morphological operations involving a series of erosions and conditional dilations. We used the segmented FLAIR image (see next section) to remove any areas of WMH included in these segmented images.

White matter hyperintensity segmentation

We segmented the FLAIR images using code written in house using Matlab (The Mathworks, Inc, Massachusetts, USA), using some SPM99 routines. The first step was to remove all non-brain regions from the image. We used the extracted brain image from the T1 weighted segmentation described above. However, to ensure that the brain mask left only brain tissue, we eroded the outer surface further. To do this, we took the extracted brain image, and set pixels to a value of one if they were inside the brain. The pixels in the ventricles were also set to a value of one, using a prior probability map based on the SPM99 CSF map. An iterative process of convolving the extracted brain with a one pixel wide Gaussian, and thresholding to remove pixels of intensity less then 0.7 was repeated six times. This eroded brain mask was then transformed from MNI space to the orientation of the FLAIR image, using SPM’s deformation toolbox, utilizing information from the scanner about the relative orientation of the FLAIR slices.

This eroded brain mask was used to remove non brain regions from the FLAIR images. The images were then segmented on a slice by slice basis, by labeling pixels with an intensity greater than 1.4 times the modal pixel intensity as WMH. This threshold level was chosen empirically, as a value which provided good visual segmentation. Following the threshold operation, WMH pixels which were not directly connected to at least two other WMH pixels were removed. This step removed isolated bright pixels, which were mostly not actual WMH.

We created a region of interest template in MNI space, with regions for the major lobes of the brain. This template was transformed onto the FLAIR image for each subject. This template was used to eliminate areas of bright signal in regions of grey matter from the segmented images. It could also be used to quantify volumes of WMH in different lobes of the brain, though we do not consider that here.

The segmentation software and region of interest template may be obtained by request to the author (MJF).

We wished to validate the WMH segmentation process, so all the scans were separately segmented by an observer (TM) who was blinded to the automated segmentation results. Hard copy images of the T2 and FLAIR sequences were obtained. Lesions were initially identified on the hard copies and were considered present if visible as hyperintense on T2 weighted and FLAIR images. Lesion segmentation was performed on electronic data using these marked hard copies as references. The WMH volumes were measured from the FLAIR images using software developed in house, written in the platform independent java language (http://java.sun.com) and based on the dispimage program5. Using a mouse, the operator identified a point inside a lesion, and manually adjusted the contour threshold so that the outside of the lesion was best delineated. The same procedure was repeated for each lesion in each slice. Only a few lesions could not be delineated at all with the contour technique, and these were outlined manually.

Spectroscopy data

We quantified metabolite ratios using a fully automated spectral fitting software 6 to estimate the peak areas of NAA, choline and creatine from a Levenburg-Marquardt fitting of Gaussian curves to the magnitude spectra. Individual spectra were rejected if the fitted linewidth at half height was greater than 10 Hz. This software generated images of the NAA/Cre , NAA/Cho and Cho/Cre concentration ratios.

To estimate metabolite ratios in grey and white matter, we followed a similar procedure to other researchers 7, 8. The GM, WM, CSF and WMH segmented images were spatially transformed from MNI space to the orientation of the spectroscopy data, using information from the scanner on the relative orientation of the different scans. The segmented images were resampled at the same resolution and point spread function as the spectroscopic images by Fourier transforming the segmented images, averaging over the spectroscopic image slice thickness, resampling the data on a 24x24 grid, and performing an inverse Fourier transform to generate a 24x24 pixel image. For each spectroscopic voxel, we then calculated the volume of grey matter (Vg), white matter (Vw), CSF (Vc) and WMH (Vh)

To estimate the metabolite values in pure grey and white matter, we assumed that the metabolite ratio for each voxel could be considered as a sum of the contribution from the different tissue types in the voxel. Assuming that the CSF does not contribute to the metabolite signal, the metabolite ratio Mv in the voxel can be written

[1]

where Mw , Mg , Mh are the average metabolite ratios for white matter, grey matter and WMH, and fw , fg and fh are the fractional volume of white matter, grey matter and WMH per brain tissue in the voxel , ie

[2]

For each of the two CSI slices, we performed a multiple linear regression with equation [1] to determine the mean metabolite ratios for grey matter, white matter and WMH. We did not include voxels which contained more than 50% CSF in the regression. We used routines written in IDL (Research Systems, Boulder, Colorado, USA ) to calculate the regressions.

The mean metabolite ratio values for grey and white matter were corrected for T2 decay using figures for T2 from 9 (Grey matter NAA 388, Cr 207, Cho 421; White matter NAA 515 Cr 214, Cho 340 ms), and using an echo time of 272 ms.

Diffusion images

We calculated the mean apparent diffusion coefficient using the equation10

[3]

where S(Bz) is the image intensity with the diffusion gradient in the z direction, and S(Bo) is the image intensity with no diffusion weighting.

As is common with EPI images, there was distortion present in the diffusion images. In order to estimate the distortion, we used the nonlinear transformations in SPM. For each patient, we spatially normalized the T2 weighted image to the T2 weighted MNI template. The ADC image for that patient was then spatially normalized to the normalized T2 image for that patient.

We segmented the T2 image using SPM99 (as described above). Because CSF and WMH are well defined on T2 images, the WM segmentation from this was free of both CSF and WMH.

The transformation matrix for the ADC image was then used to transform the segmented WMH and NAWM from the T2 segmentation to the space of the diffusion image.

These segmentations were then used to generate a masked ADC image which consisted of only normal appearing white matter, and we calculated the median value of the apparent diffusion coefficient. We also calculated the median ADC value in the areas of segmented WMH.

References

1. Talairach J, Tournoux P. Co-planar Stereotactic Atlas of the Human Brain. Stuttgart: Thieme,1988.

2. Ashburner J, Neelin P, Collins DL, Evans A, Friston K. Incorporating prior knowledge into image registration. Neuroimage 1997; 6: 344-352.

3. Ashburner J, Friston KJ. Nonlinear spatial normalization using basis functions. Hum Bra Mapp 1999; 7: 254-266.

4. Ashburner J, Friston KJ. Voxel-based morphometry— The methods. Neuroimage 2000; 11: 805-821.

5. Plummer DL. DispImage: A display and analysis tool for medical images. Rivista di Neuroradiologia 1992; 5: 489-495.

6. Soher BJ, Young K, Govindaraju V, Maudsley AA. Automated spectral analysis III: Application to in vivo proton spectroscopy and spectroscopic imaging. Magn Reson Med 1998; 40: 822-831.

7. Schuff N, Ezekiel F, Gamst AC, et al. Region and tissue differences of metabolites in normally aged brain using multislice 1H magnetic resonance spectroscopic imaging. Magn Reson Med 2001; 45: 899-907.

8. Weber-Fahr W, Ende G, Braus DF, et al. A fully automated method for tissue segmentation and CSF correction of proton MRSI metabolites corroborates abnormal hippocampal NAA in schizophrenia. Neuroimage 2002; 16: 49-60.

9. Kreis R, Ernst T, Ross BD. Absolute quantitation of water and metabolites in the human brain. II metabolite concentrations. J Magn Reson B 1993; 102: 9-19.

10. Thomas DL, Lythgoe MF, Pell GS, Calamante F, Ordidge RJ. The measurement of diffusion and perfusion in biological systems using magnetic resonance imaging. Phys Med Biol 2000; 45: R97-138.