Machine Learning Department
School of Computer Science, Carnegie Mellon University


Feature-based cs. Intensity-based Brain Image Registration:
Voxel Level and Structure Level Performance Evaluation

Leonid A. Teverovskiy, Owen T. Carmichael*,
Howard J. Aizenstein**, Nicole A. Lazar***, Yanxi Liu+

November 2006


Keywords: Deformable registration evaluation, mutual information, false discovery rate

The power and validity of voxel based and tensor based morphometry methods depend on the accuracy of the brain image registration algorithms they employ. We propose a mutual information based quantitative evaluation method to compare the performance of two publicly available deformable registration packages: HAMMER and algorithms in the ITK package (FEM-Demons) The advantage of our approach is that registration algorithms are quantitatively compared at both global and local levels, thus enabling our method to pinpoint areas of the brain where one algorithm performs significantly better or worse than the others. The brain image dataset used for evaluation consists of a total of 59 images: 20 MR images of Alzheimer's (AD) patients, 19 MR images of people with mild cognitive impairment (MCI) and 20 MR images of normal (CTL) subjects. Global and localized mutual information scores are used to evaluate the quality of registration, and paired t-tests are used to determine the statistical significance of registration quality differences between the methods at three levels: global, voxel-wise and anatomical structures. We threshold the resulting p-value maps using a false discovery rate control method in order to correct for multiple comparisons. Our results show that both HAMMER and FEM-Demons algorithms do significantly better than an affine registration algorithm, FLIRT, at all three levels for all three subject groups. Comparison between the HAMMER and FEM-Demons algorithms shows that at the global level there is no significant difference in performance between the two algorithms on controls, and FEM-Demons outperforms HAMMER on Alzheimers patients (p-value 0.0416) and MCI patients (p-value 0.0055). At the local and anatomical levels, FEM-Demons and HAMMER dominate each other on different brain regions. Our results indicate that the choice between the HAMMER and the FEM-Demons algorithms should depend on the region of interest of a study.

*Department of Neurology, University of California, Davis, CA.
**Department of Psychiatry, University of Pittsburgh, Pittsburgh, PA
***Department of Statistics, University of Georgia, Athens, GA
+Department of Psychiatry, University of Pittsburgh; Department of Computer Science and Engineering, Pennsylvania State University, University Park, PA; Robotics Institute, Carnegie Mellon University

36 pages

SCS Technical Report Collection
School of Computer Science homepage

This page maintained by