Evaluating Multi-Atlas Segmentation for Structural MRI

We evaluated the proposed multi-atlas segmentation within AutoSeg with a dataset of 35 defaced T₁-weighted structural MRI scans. Fifteen scans (5 males and 10 females with an age range of 19–34) were used as the multi-atlas population and the remaining 20 scans (8 males and 12 females with an age range of 18–90) were used for testing. Thus, the 20 testing MRI scans were segmented one-by-one using the 15 atlases. These MRI scans were selected from the Open Access Series of Imaging Studies (OASIS) database (http://www.oasis-brains.org) (Asman and Landman, 2013 (link)). This dataset has been used in the MICCAI 2012 Multi-Atlas Labeling challenge, URL: https://masi.vuse.vanderbilt.edu/workshop2012/. This dataset was expertly labeled courtesy of Neuromorphometrics, Inc. (Somerville, MA) and provided under a non-disclosure agreement of the Creative Commons Attribution-NonCommercial (CC BY-NC). For each atlas, a collection of 28 labels of subcortical structures were used (Asman and Landman, 2013 (link)): 3rd ventricle, 4th ventricle, brain stem, left/right hemispheric accumbens, cerebral White Matter (WM), cerebellar WM, caudate, amygdala, hippocampus, lateral ventricle, pallidum, putamen, thalamus, and ventral diencephalon (DC), as well as cerebellar vermal lobules I-V, VI-VII, and VIII-X. All images are 1 mm isotropic resolution.