Chimeric image creation

The set of chimeric images (artificially lesioned brain images) was generated by replacing in each non-lesioned image and sequence (b0 and b1000) the signal intensities in the corresponding voxels within each of the lesions, resulting in 38 x 75 = 2850 chimeric images.

The point of this manoeuvre was to evaluate the algorithm’s ability to

distinguish lesioned from non-lesioned signal at the same location. Unless one is fortunate enough to have a pre- and post-lesion scan for several patients, which we are not, the only way of performing such an analysis is by using “chimeric” images of this kind (Nachev et al., 2008). If we had no grounds to suspect that lesion discriminability varied with location an analysis of this kind would have been unnecessary. However, it is clear that this cannot be assumed to be so, especially with DWI where some locations (e.g., frontal and temporal poles) are much more prone to artifactually abnormal signal than others.

For the chimeric images to be realistic the non-lesioned tissue in each donor/ recipient pair must be broadly similar. Since, unlike CT scans, MRI images do not have a standardized intensity scale (Bergeest and Jager, 2008; Nyul and Udupa, 1999), each transplantation step must be preceded by equating the global intensity of the donor and recipient images. This was done by calculating the mean whole brain signal, excluding voxels falling within the lesion, of the lesioned, donor image and the mean whole brain signal of the recipient image, and multiplying the signal value at each voxel within the latter image by the ratio of the two means (Brett et al., 2001). This procedure ensured that the transplanted signal was as close as possible to what

abnormal signal might have looked like had it been present in the recipient; without it, unrealistic global variations in the signal would have clouded our assessment of the algorithm’s performance. We confirmed that this procedure did not artificially introduce an artefactual difference between the lesion signal

and the signal in the immediate vicinity of the lesion by comparing the ratios between intralesion (5 voxels inside the lesion boundary) and perilesion (5 voxels outside the lesion boundary) signal for all original lesion images and their corresponding sets of chimerics (paired t test p value < 0.01 for each set). Figure 2.2 is a flow diagram illustrating the process of creating these images.

b-sequence- Recipient

nb-sequence-

Recipient nb-sequence-Donor nb1000-Mask

nb-sequence- Organ nb-sequence- Chimeric b-sequence- Donor

Figure 2.2 - Chimeric image creation flow diagram.

Both the recipient and donor images (b0 and b1000 sequences) were

spatially normalised into standard MNI space using the unified segmentation- normalisation routine in SPM5 to create nb-sequence-Recipient and

nb-sequence-Donor. The default settings for the unified segmentation- normalisation routine were used (appendix B), except for the interpolation being set to 7 in the normalise estimate and write parameters. The lesion

within the donor b1000 sequence was then manually segmented. Whilst in MNI space, the binary mask was applied to the normalised b sequences (b0 and b1000) donor brain (nb sequence-Donor) to extract the lesioned voxels (nb- sequence-Organ), which was subsequently transplanted into the corresponding normalised b sequence recipient brain (nb-sequence-Recipient) to create a chimeric brain (nb-sequence-Chimeric).

Abbreviation Description

T1-Recipient Unlesioned T1 image in native T1 space nT1-Recipient Unlesioned T1 image in MNI space

b1000-Donor Lesioned b1000 image in native b1000 space

nb1000-Donor Lesioned b1000 image in MNI space

nb1000-Mask Binary lesion mask in MNI space

nb1000-Organ Extracted lesioned voxels in MNI space nT1nb1000-Chimeric Chimeric image in MNI space

Figure 2.3 - Table of abbreviations for figure 2.2.

The various images are named with the volume space first and their role second. All images are diffusion weighted scans.

112

2.2.3

Zeta (z) anomaly score

We have seen that the gamma (g) score of a given point, x, is the average distance to its k nearest neighbours (nn)

The zeta score is the difference between gamma, an index of the anomaly of the point in relation to its neighbours, and the average inner-clique distance of its neighbours, an index of the density of the neighbourhood clique (Rieck and Laskov, 2006).

This was done by calculating the mean whole brain signal, excluding voxels falling within the lesion, of the lesioned, donor image and the mean whole brain signal of the recipient image, and multiplying the signal value at each voxel within the latter image by the ratio of the two means (Brett et al., 2001). This procedure ensured that the transplanted signal was as close as possible to what abnormal signal might have looked like had it been present in the recipient; without it, unrealistic global variations in the signal would have clouded our assessment of the algorithm’s performance. We confirmed that this procedure did not artificially introduce an artefactual difference between the lesion signal and the signal in the immediate vicinity of the lesion by comparing the ratios between intralesion (5 voxels inside the lesion boundary) and perilesion (5 voxels outside the lesion boundary) signal for all original lesion images and their corresponding sets of chimerics (paired t test p value <.01 for each set). Figure XX is a flow diagram illustrating the process of creating these images Figure XX : flow diagram illustrating the process of creating the chimeric images.

Both the recipient and donor images (b0 and b1000 sequences) were spatially normalised into standard MNI space using the unified segmentation-normalisation routine in SPM5 to create n-Recipient and n-Donor. The default settings for the unified segmentation-

normalisation routine were used (appendix A), except for the interpolation being set to 6 in the normalise estimate and write parameters. The lesion within the donor b1000 sequence was then manually segmented (see above). Whilst in MNI space, the binary mask was applied to the normalised b1000 donor brian (n-Donor) to extract the lesioned voxels (n- Organ), which was subsequently transplanted into the normalised b1000-recipient brain (n- Recipient) to create a chimeric brain (n-Chimeric).

figure - Table of abbreviations

The various images are named with the volume space first and their role second. All recipient images are T1 scans, whilst donor images are b1000 scans.