29 JanSfN2012 Poster

Bootstrapping Neuroanatomical Labeling in MRI

Andrew Worth¹, Jason Tourville²

¹ Neuromorphometrics, Inc., Somerville, MA
² Dept. of Speech, Language, and Hearing Sciences, Boston University, Boston, MA

Abstract (download poster, 3.7 MB)

The object of this work is to reach the next level in human MRI brain scan labeling by bootstrapping from previous results. By “next level” we mean more reproducible, accurate, and precise results, and also an increase in the number and specificity of labeled regions of interest (ROIs) as well as subdivisions of ROIs. “Bootstrapping“ refers to developing the system, applying it, examining results, and then using the experience and knowledge gained to improve the system by repeating the process.

Protocols were developed to explicitly define ROI boundaries and give rules of thumb on how to handle usual and unusual cases. The initial definitions were based on anatomical literature and information from atlases, but this is generally insufficient because protocol validity depends on the reproducibility of the results, and this cannot be known until after the protocol has been carried out for a large number of subjects.

Methods for localizing ROI boundaries and labeling individual voxels were developed along with the anatomical protocols. Boundaries are created by hand tracing, by using isointensity contours, histograms, or an edited combination of these. Voxel labeling is done slice-by-slice in 2D and also by using 3D surfaces to provide anatomical context.

After applying the protocols and methods to create a large number of manually labeled ROIs, this information was aggregated in order to examine the results. Normative statistics on volume, location, shape, and intensities were compiled and a probabilistic atlas was generated to understand the variation of the ROIs, and also to check, correct, and refine the boundaries and voxel labels. Regions of high variability were identified, localized and examined to determine if they were a result of errors, methodological issues, or normal anatomical variation.

The protocol and methods were then modified to eliminate or capture variability and to produce improved and more comprehensive results. Finally, the new results were used to re-generate normative statistics and atlases, and the process was repeated.

Using a database of 50 manually labeled T1-weighted MRI scans and protocols defining 157 ROIs, we illustrate the bootstrapping procedure by describing the types of corrections that were made to individual ROIs using the aggregated results. ROI volume, Dice overlap, and atlas mismatch were used to flush out labeling errors. We also show how the protocol can be updated by adding a new ROI and by parcellating an existing ROI.

The precise, comprehensive and reproducible labeling of human brains as seen in magnetic resonance imaging is a difficult, arduous task that can be greatly improved by bootstrapping.