ABSTRACT
Biomedical data is accumulating at an increasing rate. The storage requirements, particularly of imaging data, are becoming very large. Also, the analysis of this data is becoming increasingly complex. However, the data from large databases provides the potential to extract comprehensive conclusions that enable personalized patient treatment. The analysis of the data is very often complicated by its multi-contrast nature and its varying quality. Novel methods must be developed to improve the quality of the datasets and to make them comparable. In this study, the quality of the datasets of a patient is improved with non-parametric methodologies. They are first pre-processed for intensity correction and then for rigid co-registration. Both formulations are statistical and non-parametric. The rigid registration method developed in this work uses the Bayesian posterior expectation to provide an intermediate vector field. This field is processed with the Procrustes method to provide a rigid transformation as the maximum likelihood estimate. The registration method developed is multi-contrast, and hence it can accommodate different types of datasets. The methodology is validated with images from three different databases of brain datasets with existing and with simulated misregistrations. One database consists of images from the BrainWeb simulator, the other from the Human Connectome Project (HCP), and a third database of images of Parkinson’s disease patients. The validation demonstrates the ability of the method to compensate for very extensive rigid misregistrations.
