Objective: Segmentation of MRI brain volumes into grey matter (GM), white matter (WM) and cerebrospinal fluid (CSF) is an integral part of many analysis techniques including cortical surface extraction, the determination of cortical thickness and substructure volumes, inter-subject and cross-modality registration, and longitudinal studies of cerebral atrophy.

Historically quantitative measurements of the grey-white ratio (GWR) employed weighing paper cutouts of tissue boundaries [1] and in vivo xenon washout [2]. Given the importance of the GWR, we evaluated the performance of six brain-tissue segmentation algorithms with respect to the absolute value and reproducibility of the computed whole-brain GWR. We also examined the differences between "hard" segmentations (tissue labels assigned to each voxel) and "soft" segmentations (percentage of each tissue component defined at each voxel).

Methods: Six repeat T1-weighted MRI brain scans were acquired from six normal subjects on a 4T Siemens scanner. During the first scanning session three sequential 1xNEX scans at 1.0 mm3 resolution and one 1xNEX scan at 1.5 mm3 resolution were obtained; a single 1xNEX scan was acquired during subsequent sessions for a total of eight 1.0 mm3 scans. For each subject the three 1xNEX scans acquired during the first scanning session were aligned and combined into three 2xNEX volumes and a single 3xNEX volume. Each scan was corrected for intensity nonuniformity using N3 [3], and non-brain tissues were removed using BET2 [4]. Six brain-tissue segmentation algorithms were applied to each MRI volume; four produced hard segmentations (FAST [5], INSECT [6], SEGM [7], and FANTASM [8]), and four produced soft segmentations (FAST, PVS [9], SPM [10], and FANTASM). For hard tissue segmentations GWRs were computed as the ratio of GM-to-WM voxels; for soft segmentations the GWR was defined as the ratio of the integral of fractional GM and WM volumes.

Results & Discussion: Figure 1 illustrates the mean and SD of the GWRs averaged across eleven volumes from each of six subjects (red). To minimize the effect of variations in brain masks (~1% variation between volumes) the GWR for seven aligned volumes (using a common brain mask) for each subject was computed (green). Two segmentation algorithms, INSECT and SPM, produced GWRs that were significantly different from the other algorithms and from historical measurements; these algorithms also produced the greatest variation in computed GWRs. For those techniques that created both hard and soft segmentations (FAST and FANTASM), the hard segmentation resulted in a larger GWR. Analyzing aligned volumes with a common brain mask resulted in slightly larger GWRs for all techniques.

Conclusions: Compared to other segmentation techniques and the historical literature, SPM and INSECT appear to significantly overestimate the GWR and produce variability in the computed GWR.

References & Acknowledgements:
1. Hennerberg R. J. Psychol. Neurol. 17:144, 1910.
2. Hoedt-Rasmussen K, Skinhoj E. Neurology. 16:515-520, 1966.
3. Sled JG, et al. IEEE TMI 17:87-97,1998.
4. Smith SM. HBM 17(3):143-155, 2002.
5. Zhang Y, et al. IEEE TMI 20(1):45-57, 2001.
6. Collins DL, et al. IPMI'99, LNCS 1613:210-23, 1999.
7. Gur RC, et al. J Neurosci, 19(10):4065-72, 1999.
8. Pham DL, Prince JL. IEEE TMI. 18(9):737-52, 1999.
9. Shattuck DW, et al. NeuroImage 13(5):856-876, 2001.
10. Ashburner J, Friston C. NeuroImage 11:805-821, 2000.

This work was supported by NIH grant P20 EB002013.