Multistructure Large Deformation Diffeomorphic Brain Registration
Khan, A. R.; Wang, L.; Beg, M. F.
Volume: 60, Issue:2, Publication Year: 2013, Page(s): 544 – 553
Whole brain MRI image registration, the task of spatially warping one anatomical image to match another, has many useful applications in group analysis and morphometry in the study of neurological diseases. Registration is a difficult task considering the anatomical variability and structural complexity present in the human brain, and thus accurate registration across neuropathological groups remains challenging. Structure-specific information, such as labels of individual brain structures, can be used to initialize and constrain registration to improve accuracy and robustness. Our work describes and validates a novel multistructure registration approach that uses shape matching of automatically segmented brains to aid the overall MRI image registration. Validation experiments carried out on openly-available datasets demonstrate comparable or improved alignment of subcortical and cortical brain structures over leading brain registration algorithms. We also detail how a group-wise average atlas could be built with multistructure registration and applied it to an openly available database of 150 MRI images from elderly demented and non-demented patients, showing that the multistructure registration accounts for greater inter-subject variability leading to greater sharpness and anatomical detail in the resulting average atlas. Utilizing this technique to study shape differences between the non-demented and demented patient populations revealed our proposed multistructure approach is more sensitive in detecting these shape differences and could thus lead to a more sensitive biomarker of disease progression. This work markedly impacts biomedical research in brain imaging and image analysis as it significantly improves upon the fundamental technique of brain registration, allowing for more powerful analyses of shape and structure, shown to be important biomarkers of disease progression. Furthermore, this work provides a general framework for multistructure image registration that can be extended to include additional imaging modalities or higher level image features to further advance the field.