Objective: Surface extraction techniques play an increasingly important role in mapping patterns of change in healthy and diseased brains. Surface-based characteristics such as area, volume, gyrification, or cortical thickness may help to quantify changes in the underlying parenchyma. Although previous studies have often employed statistical filtering to identify cortical areas that exhibit a significant amount of change, no systematic attempt has been made to account for the variability introduced by the surface extraction technique itself. This variability can be ignored when the number of subjects is large but poses significant problems for longitudinal studies of single subjects. In such cases large variability precludes the ability to draw conclusions concerning anatomical change. Based on repeat MRI brain scans from the same subject over a period of weeks to months we measured the variability of six cortical extraction techniques by computing the coefficient of variation for a set of surface metrics.

Methods: A set of MRI brain volumes was analyzed for each of two normal subjects. The first set consisted of six T1-weighted volumes acquired on a Phillips 1.5T scanner (1 mm³ voxels). The second consisted of eight T1-weighted MRI volumes acquired on a Siemens 4T scanner (also 1 mm³ voxels). All volumes were non-uniformity corrected using N3 [1], and stripped using BET2 [2]; to ensure consistency all required manual steps were performed by the same person. Surfaces were generated using six different packages: FreeSurfer [3], SurfRelax [4], CLASP [5], BrainVisa [6], BrainVoyager [7], and SureFit [8]. For each package we calculated surface area, surface volume, normalized shape index (SI, the deviation of an object from a sphere), and gyrification index (GI, the inner-to-outer cortical surface area ratio) [9]. Cortical thickness was approximated by the surface separation (SS, the mean of the closest distance from the white surface to the gray surface for each vertex).

Results & Discussion: The coefficients of variation for each surface metric are presented in Table 1, and Figs. 1 and 2. It should be noted that the metrics listed do not refer to the accuracy of a given method, only to its precision. Metrics involving gray surfaces were not computed for BrainVoyager and SureFit; BrainVoyager creates a gray surface for reference purposes only and SureFit only creates a layer 4 mid-cortical surface.

Conclusions: The performance of each surface extraction method varied within and between subjects, although packages tended to perform consistently across most metrics. FreeSurfer performed best overall but also required the greatest amount of manual interaction. BrainVisa tended to have the highest variability, but this may arise from the relatively low number of triangles in the final tessellations.

References & Acknowledgements:
1. Sled, JG. et al. (1998) TMI 17(1):87-97.
2. Smith, SM. (2002) HBM 17:143-155.
3. Dale AM. et al. (1999). NeuroImage 9:179-194.
4. Larsson J. (2001). PhD Thesis. Karolinska Institute.
5. Kim JS, et al. (2004). Submitted to NeuroImage.
6. Cointepas Y. et al. (2001). NeuroImage 13(6):S98.
7. Kriegeskorte N & Goebel R. (2001). NeuroImage 14:329-346.
8. Van Essen DC. et al. (2001). J Am Med Inform Assoc. 8:443-459.
9. Zilles, K. (1988) Anat Embryol 179(2):173-9.

Supported by NIH Grant P20 EB002013.