Validation of FS+LDDMM by automatic segmentation of caudate nucleus in brain MRI

For over last three decades, numerous automatic brain segmentation techniques in magnetic resonance (MR) images have been proposed. These techniques, however, need to be validated comprehensively. In this study, FS+LDDMM, a recently proposed fully automatic template-based brain segmentation technique, is validated. The validation method uses novel approach in which dependency of FS+LDDMM on initial segmentation parameters is evaluated. These segmentation parameters include choice of template, gross alignment, cropping size and initialization schemes. A database of 46 MR images from young ADHD subjects of an average age of 10.6 years is employed to segment caudate nucleus in subcortical region. The accuracy of the segmentation is computed by comparing FS+LDDMM segmentation with gold standard manual segmentation using metrics, such as, percent volume error, dice coefficient, L1 error and intraclass correlation coefficient (ICC). The FS+LDDMM shows robustness to all these parameters and outperforms FreeSurfer (FS) segmentation. To generalize the performance of FS+LDDMM, however, more experiments need to be conducted for various subcortical objects.