Image registration methods for the synthesis and evaluation of anatomical population summaries

Comparison of morphology in populations of biological forms is complex due to the multitude of factors that govern shape and size, as well as the relative lack of validated robust quantitative tools to analyze structural data, especially 3D images. New image registration methods were developed for analysis of population morphology that reduce the large systematic errors inherent in earlier methods. High dimensional pairwise transforms of shape between individuals are generalized to represent group morphology. The validity and reproducibility of these transformations was established by simulation and experiment.; Image registration is a method for determining correspondences between intensities and edges optimized to match datasets pairwise according to a variety of criteria. In general, image registration methods are tailored to specific applications by the selection of the matching criteria and assumptions made in the mathematical formulation of the approach. The results obtained from registration of images with different methods vary dramatically, and many methods will produce large systematic errors. The relative importance of the non-unique image registration results depend on the intended application and data sets used. Among the most important systematic errors present in image registration is that due to the non-invertibility of shape transformations that encode the correspondences between individual data sets. In this dissertation, robust methods for consistent image registration were developed, tested and validated. Consistent refers to the incorporation of an inverse consistency constraint on the shape transformations. Consistent image registration was found to effectively suppress and virtually eliminate systematic registration errors.; Consistent image registration was applied to populations of skull and brain morphology based on CT and MRI data sets. In each instance, substantial reductions in registration errors were demonstrated with reduced variability in the results. The feasibility of population shape comparisons was established for infant skulls scanned with CT and adult brains scanned with MRI. The results indicate that consistent image registration is a robust and valid method for population comparisons of skull and brain shape based on CT and MRI scans. Future work is needed to establish these methods for multi-modality applications, extension to other body regions, and efficient statistical comparisons of population shape differences.