This paper evaluates the performance of ANTs, a software package for image registration, template building, and segmentation, focusing on the impact of similarity metrics on whole-head registration-based labeling. The authors discuss the importance of reproducible research and the limitations of existing evaluation studies, which often lack transparency in their methods and parameters. ANTs, built on the Insight Toolkit (ITK), offers a modular framework for evaluating different components of image processing, such as transformation models and similarity metrics. The paper provides an overview of ANTs' transformation models, including rigid, affine, and diffeomorphic transformations, and similarity metrics such as mean squared intensity difference (MSQ), cross-correlation (CC), and mutual information (MI). It details the implementation of these metrics and their derivatives, emphasizing the importance of efficient gradient calculations for computational efficiency. The evaluation uses the LPBA40 dataset, which contains 40 3D MRI images of healthy volunteers, manually labeled with 56 structures. The authors apply ANTs' methods to construct templates, label them, and map subjects to these templates. The evaluation pipeline includes constructing templates, labeling, and mapping subjects to the templates using different metric pairs (MSQ, MSQ), (CC, CC), (MI, MI), (MSQ, MI), and (CC, MI). Key findings include: - Template stability across metrics and populations is demonstrated, with overlap values exceeding human rater performance. - Mutual information-based affine registration provides the best initialization for deformable registration. - MI performs well in whole-head affine registration and is robust to scanner variations and pathomorphological changes. - Brain extraction performance is sensitive to affine initialization quality, with MI-based initialization showing superior results. - Cross-validation reduces bias and tests generalization to new data, highlighting the importance of quality affine initialization. The paper concludes by discussing the implications of these findings for future research and the need for further work to understand the nature and impact of residual differences in template shapes.This paper evaluates the performance of ANTs, a software package for image registration, template building, and segmentation, focusing on the impact of similarity metrics on whole-head registration-based labeling. The authors discuss the importance of reproducible research and the limitations of existing evaluation studies, which often lack transparency in their methods and parameters. ANTs, built on the Insight Toolkit (ITK), offers a modular framework for evaluating different components of image processing, such as transformation models and similarity metrics. The paper provides an overview of ANTs' transformation models, including rigid, affine, and diffeomorphic transformations, and similarity metrics such as mean squared intensity difference (MSQ), cross-correlation (CC), and mutual information (MI). It details the implementation of these metrics and their derivatives, emphasizing the importance of efficient gradient calculations for computational efficiency. The evaluation uses the LPBA40 dataset, which contains 40 3D MRI images of healthy volunteers, manually labeled with 56 structures. The authors apply ANTs' methods to construct templates, label them, and map subjects to these templates. The evaluation pipeline includes constructing templates, labeling, and mapping subjects to the templates using different metric pairs (MSQ, MSQ), (CC, CC), (MI, MI), (MSQ, MI), and (CC, MI). Key findings include: - Template stability across metrics and populations is demonstrated, with overlap values exceeding human rater performance. - Mutual information-based affine registration provides the best initialization for deformable registration. - MI performs well in whole-head affine registration and is robust to scanner variations and pathomorphological changes. - Brain extraction performance is sensitive to affine initialization quality, with MI-based initialization showing superior results. - Cross-validation reduces bias and tests generalization to new data, highlighting the importance of quality affine initialization. The paper concludes by discussing the implications of these findings for future research and the need for further work to understand the nature and impact of residual differences in template shapes.