Segmentation of brain MRI during early childhood

The objective of this thesis is the development of automatic methods to measure the changes in volume and growth of brain structures in prematurely born infants. Automatic tools for accurate tissue quantification from magnetic resonance images can provide means for understanding how the neurodevelopmental effects of the premature birth, such as cognitive, neurological or behavioural impairment, are related to underlying changes in brain anatomy. Understanding these changes forms a basis for development of suitable treatments to improve the outcomes of premature birth. In this thesis we focus on the segmentation of brain structures from magnetic resonance images during early childhood. Most of the current brain segmentation techniques have been focused on the segmentation of adult or neonatal brains. As a result of rapid development, the brain anatomy during early childhood differs from anatomy of both adult and neonatal brains and therefore requires adaptations of available techniques to produce good results. To address the issue of anatomical differences of the brain during early childhood compared to other age-groups, population-specific deformable and probabilistic atlases are introduced. A method for generation of population-specific prior information in form of a probabilistic atlas is proposed and used to enhance existing segmentation algorithms. The evaluation of registration-based and intensity-based approaches shows the techniques to be complementary in the quality of automatic segmentation in different parts of the brain. We propose a novel robust segmentation method combining the advantages of both approaches. The method is based on multiple label propagation using B-spline non-rigid registration followed by EM segmentation. Intensity inhomogeneity is a shading artefact resulting from the acquisition process, which significantly affects modern high resolution MR data acquired at higher magnetic field strengths. A novel template based method focused on correcting the intensity inhomogeneity in data acquired at higher magnetic field strengths is therefore proposed. The proposed segmentation method combined with proposed intensity inhomogeneity correction method offers a robust tool for quantification of volumes and growth of brain structures during early childhood. The tool have been applied to 67 T1-weigted images of subject at one and two years of age