| HBM 2005 Abstracts: Previous | Next | Click here for a PDF of this poster. |
1Department of Neurology, University of Minnesota (UM), Minneapolis, USA, 2Department of Radiology, University of Minnesota (UM), Minneapolis, USA, 3McConnell Brain Imaging Centre, Montreal Neurological Institute (MNI), McGill University, Montreal, Quebec, Canada |
||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
Objective:
Accurate, reproducible landmarking of the cortical surface is difficult due to variability in indivdual patterns of sulcation and gyration, and no previous study has generated cortical landmarks with sufficient inter-rater reproducibility to be useful for the quantitative evaluation of intersubject registration algorithms. Two recent attempts at landmarking employed reconstructed cortical surfaces [1,2] with landmarks based on visible sulcal intersections or visually extrapolated intersections assumed to lie on the adjacent cortical surface. For the two raters in [1] the Euclidean distance between individual and mean landmark positions for 12 x 2 landmarks (12 landmarks per hemisphere) averaged 6.0 mm; intra-rater variability averaged 3.7 mm. Repeatability in [2] based on replicate measurements by a single rater ranged from 0.25 to 1.55 mm. In both studies [1,2] the distribution of landmarks included few landmarks on the parietal and occipital surfaces, and none on the temporal surfaces. In another study [3] 128 x 2 landmarks were defined on ten anatomic reference planes; intra-rater reliability of these points involving a single repeat labeling of each landmark on a single reference brain yielded an RMS error of 1.6 mm.
Methods: Four raters at UM and MNI independently identified 13 x 2 cortical and 10 sub-cortical point landmarks on 37 high-quality1.5T T1-weighted MRI volumes (voxel dimensions 1 x 1 x 1 mm) based on a detailed written protocol. Prior to landmarking MRI volumes were corrected for intensity non-uniformity, intensity normalized, and transformed into unscaled Talairach space [4]. Using REGISTER visualization software (MNI), landmarks were identified by 3D navigation within the MRI volumes and their (x, y, z) coordinates were recorded for subsequent statistical analysis.
To calculate inter-rater reproducibility, we computed the distance from the four raters' points to the centroid of those four points for each landmark/volume combination, yielding 5,328 distances. Protocols for 6 x 2 landmarks were revised, and the resulting 1,776 points were re-labeled. To minimize the effect of possible careless errors, an "outlier index" was computed for each landmark/volume combination: the three distances between a given rater's point and those of the other three raters were averaged and subtracted from the mean of the three distances between the other three raters. Possible outliers were identified based on an outlier index > 2 mm, and raters were asked to re-examine their own outlier points without being told how to adjust them. Of the 433 possible outliers identified, 233 were adjusted leading to a mean improvement of 10.5 mm with respect to the centroid of the other three raters' points.
Results & Discussion:
Preliminary reproducibility statistics for the 5,328 distances are listed in Tables 1 and 2. The median distance to centroid for cortical and non-cortical landmark points, respectively, was 1.24 and 0.68 mm.
Conclusions:
Our results suggest that a large number of distributed cortical and subcortical landmarks are highly reproducible across raters, sites, and subjects and can serve as well-characterized points of reference for future developmental, neuroanatomical, and disease-related studies as well as goodness-of-warp metrics.
References & Acknowledgements: | ||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||