Optimizing automated preprocessing streams for brain morphometric comparisons across multiple primate species

Abstract

AbstractIntroduction: MR techniques have delivered images of a wide array of species, ranging from invertebrates to birds to elephants and whales. However, their potential to serve as a basis for comparative brain morphometric investigations has rarely been tapped so far (Christidis and Cox, 2006; Van Essen & Dierker, 2007). In order to address one of the potential reasons behind this, we tested various automated segmentation streams (using CARET, FSL and lab-written code) that have been mainly designed for applications in humans, with the aim of optimizing them for comparisons across multiple species. Methods: T1-weighted MR images from individuals representing 11 primate species have been acquired by Rilling and Insel (1998) on the same scanner, using size-dependent protocols. These data are freely available via "http://www.fmridc.org/f/fmridc/77.html":http://www.fmridc.org/f/fmridc/77.html . While newer multi-species datasets acquired on one scanner are not currently available, our experiments with this decade-old dataset can serve to highlight the lower boundary of the current possibilities of automated processing pipelines. After manual orientation into Talairach space, an automated bias correction was performed using CARET (Van Essen et al., 2001) before the brains were extracted with FSL BET (Smith, 2002) and either smoothed by an isotropic Gaussian Kernel, FSL SUSAN (Smith, 1996), an anisotropic diffusion filter (Perona & Malik, 1990), or not at all. Segmentation of the brains was performed separately by either FSL FAST (Zhang, 2001) without atlas priors, or using an Adaptive Maximum A Posteriori Approach (Rajapakse et al., 1997). Finally, the white matter surface was extracted with CARET, and inspected for anatomical and topological correctness. Results: As shown in Figure 1, the images contained in this dataset had a low SNR by current standards. Bias correction by means of CARET did generally not have significanteffects on the quality of the segmentation, though it interfered with brain extraction in some cases. Figure 1 also shows that, as previously documented for humans, brain extraction by means of BET worked well in general, though the processed image was often missing parts of the brain tissue, or contained larger amounts of non-brain tissue (example in Figure 4). The reasons for this are that some of the key assumptions of the algorithm were only partially applicable to non-human species, namely that neither size nor centre of gravity differ greatly between head and brain. Additionally, the histograms vary across species, thus affecting the thresholding used to seed the surface growing algorithm. Figure 2 shows that noise reduction was generally necessary but could be achieved by simple isotropic filtering, with anisotropic filtering (SUSAN, diffusion filter) providing no or little improvement. While none of the tested segmentation methods performed uniformly well in all 11 species, the best cross-species segmentations were achieved by iteratively applying FSL FAST with different bias fields (Figure 3). Nonetheless, all of the non-human segmentations required topology correction and - often considerable - manual cleanup (Figure 4). Conclusions: Automated processing pipelines for surface-based morphometry still require considerable adaptations to reach optimal performance across brains of multiple species.Until these are available, time-consuming manual interventions are indispensable. Considering, however, that newer imaging datasets generally have a better signal-to-noise ratio than those used in this study and that scanner-induced variability was found, in our species, to be below group differences between healthy subjects and patients (e.g. Stonnington et al., 2008), comparative evolutionary analyses of cortical parameters like gyrification or cortical thickness are now in sight. While algorithmic improvements are essential in reaching this goal, progress along these lines will also depend on the development and adoption of cross-species standards for imaging protocols, processing pipelines and databasing.