Structural Fingerprinting of the Human Brain: How unique is tract shape to the individual?
Greg D Parker1, George J.A. Evans2, and Derek K Jones1,3

1CUBRIC, School of Psychology, Cardiff University, Cardiff, United Kingdom, 2School of Medicine, Newcastle University, Newcastle, United Kingdom, 3Neuroscience and Mental Health Research Institute (NMHRI), School of Medicine, Cardiff University, Cardiff, United Kingdom

Synopsis

Even amongst healthy subjects, brain function and structure is known to be highly variable across individuals1,2. Recently3 it was shown that inter-subject variation in functional connectivity is sufficient to allow robust and reliable identification of individuals across different sessions and tasks. Here we demonstrate for the first time that the same is true of white matter structure; using the shape of an individual's white matter tracts we generate fingerprints that uniquely identify individuals across different scan sessions.

Purpose

To demonstrate that white matter tract shape is uniquely identifiable information (i.e., serves as a “Fingerprint”).

Methods

Imaging: Two cohorts were recruited. Group 1: 6 subjects imaged each 5 times at 1 week intervals with a b = 1200s/mm2, 60 directions, 6 b0, 2.4mm isotropic protocol; Group 2: 248 unique subjects imaged once with a protocol matching Group 1.

Image Processing/Tractography: All images were corrected for motion, eddy currents and EPI distortions using ExploreDTI4 with in-house modifications to accommodate damped Richardson-Lucy based tractography5. Tracking parameters were a 1mm step size, 45o angular threshold, 0.05 fODF threshold, 2mm isotropic seeding and [30 300]mm length thresholds.

Shape Categorisation: 30 individuals' FA maps (selected at random from Group 2) were affinely co-registered to an MNI (2mm FA) template and the transformation matrices reused to spatially normalise the corresponding whole brain tractography result. Streamlines were then re-parameterised to 30 knot-points (b-spline interpolation), translated to the origin (centre-of-mass subtraction), vectorised (coordinate concatenation) and the resultant feature descriptors subjected to principal component analysis. The first seven eigenvectors (encompassing ~98% of total shape variance) were then selected to form a set of shape basis functions, combinations of which should be sufficient to approximate any plausible streamline shape. Finally, the streamline feature descriptors were then decomposed onto the shape basis functions and the resultant weight vectors clustered (k-means) into 800 classes of shape, examples of which are presented in Figure 1.

Fingerprinting: To obtain "fingerprints", subject tractographies were spatially normalised (eliminating rotation as a confound) and reduced to weighted combinations of shape basis functions. For each streamline, the Euclidean distance between its weight vector and each of the 800 cluster centroids was calculated, assigning the (arbitrary, but unique) identifying number of the closest cluster to that streamline. A histogram of streamline identification numbers – i.e., the histogram of shapes (Figure 2) – then forms the "fingerprint". Similarity between fingerprints is assessed by vectorising all histogram bin-counts and computing cross-correlations between them.

Results

Figure 3 shows the cross-correlation matrix of fingerprints within Group 1. Figure 4 displays a histogram of correlation coefficients resulting from the combination of Groups 1 and 2. Notice how the distribution of correlation coefficients derived from intra-individuals comparisons naturally separates from that arising from comparing different subjects. This allows same/different subject identification through simple thresholding of the correlation coefficient. Thresholding at 0.8 allowed, across the combined dataset (Group 1 & 2), identification of repeat-scan subject images (comparing each image in the combined cohorts, in turn, to all other images) with 93% precision.

Discussion/Conclusion

The work presented here indicates that tract shape does, indeed, carry uniquely identifiable information. Beyond subject identification – which is trivially achievable through other means – why is this important? Firstly, our technique opens up a new avenue for longitudinal study. Figure 3 demonstrates that, at least within the short term, streamline fingerprints are a highly consistent property. If this consistency can be shown to persist across longer periods, then sudden or progressive reductions in correlation coefficient (with respect to the subject's baseline) might be useful as a marker indicative of, for example, the appearance or progress of tumours, lesions or atrophy that are known to alter/disrupt white matter on a global scale. Secondly, our technique may serve as a new method for examining brain development. Many of the 800 shape parcellations here are specific to particular white matter structures but, importantly, individual clusters represent only one potential morphology and thus one fibre bundle is most likely represented by multiple clusters. By understanding the relationship between such clusters and studying the evolution of the fraction of assigned streamlines over time, brain fingerprints could be used to track, on both global and tract specific scales, the evolution of brain shape.

Acknowledgements

This work was supported through a Wellcome Trust New Investigator Award

References

1. Rypma B and D'Esposito M. The roles of prefrontal brain regions in components of working memory: effects of memory load and individual differences. Proc Natl Acad Sci USA. 1999;96:6558-6563

2. Bürgel U, et al. White matter fiber tracts of the human brain: three-dimensional mapping at a macroscopic resolution, topography and intersubject variability. Neuroimage. 2006;29:1092-1105

3. Finn ES, et al. Functional connectome fingerprinting: identifying individuals using patterns of brain connectivity. Nature Neuroscience. 2015;18:1664-1671

4. Leemans A, et al. ExploreDTI: a graphical toolbox for processing, analyzing, and visualizing diffusion MR data. Proc. ISMRM 17. 2009; abstract 3537

5. Dell'Acqua F, et al. A modified damped Richardson-Lucy algorithm to reduce isotropic background effects in spherical deconvolution. Neuroimage. 2009;49(2):1446-48

Figures

Example cluster contents. (A) Cluster 624 - shapes predominantly found in the left corticospinal tract. (B) Cluster 124 - shapes predominantly found in the body of the corpus callosum. (C) Cluster 243 - shapes found predominantly in the right inferior fronto-occipital fasciculus. Note different morphologies of these tracts are represented in different clusters.

Example fingerprint histograms. (A-B) Fingerprints of the same individual recorded at different time points (correlation coefficient=0.834). (C) Fingerprint of a unique individual (correlation with (A)=0.411 and correlation with (B)=0.451)

Fingerprint cross-correlation across six subjects (labelled 1-6) imaged over a period of five weeks. Note the distinct difference in correlation coefficient within individuals compared to across individuals.

Distribution of correlation coefficients across different individuals (blue) and same individuals (red). Note the distinct separation between these distributions. (Note: due to the differing numbers of subjects/comparisons within each group, areas under each distribution have been normalised).



Proc. Intl. Soc. Mag. Reson. Med. 24 (2016)
0122