Authors:
Maria C. Valdés Hernández
1
;
J. M. Wardlaw
1
and
Sean Murphy
2
Affiliations:
1
SFC Brain Imaging Research Centre, Division of Clinical Neurosciences, University of Edinburgh, United Kingdom
;
2
Institute for System Level Integration, United Kingdom
Keyword(s):
Unsupervised clustering, Segmentation, Multi-spectral MR, KMeans, EM, MeanShift, MVQ, Brain tissue, Grey matter, White matter, CSF.
Abstract:
The effects of atrophy and diffusion of the boundary between grey and white matter, common in elder individuals, represents a difficult problem for segmentation, not observed in healthy younger adults. The aim of this study is to evaluate four well-known unsupervised clustering algorithms in brain tissue segmentation using MR scans with atrophies and lesions. The brain is segmented into 3 different types: white matter, grey matter and CSF. We used four MR sequences: T1W, T2W, T2*W and FLAIR to classify each voxel in the image. No spatial information was used. The algorithms tested were k - means, EM (Gaussian mixture), MVQ (minimum variance quantisation) and Mean Shift. The datasets were acquired from an aged cohort (> 70 years). The resulting segmentations were quantitatively compared to expertly collected ground truth on 12 datasets, using the Dice coefficient as an overlap measure. The classification algorithms could be ranked in the following order: MVQ, k - means, EM and MeanShi
ft from best to worst. The MVQ algorithm did best of all with over a .9 Dice overlap on CSF, and over .8 on white matter.
(More)