Magnetic resonance imaging (MRI) provides spatially accurate, three dimensional structural images of the human brain in a non-invasive way. This allows us to study the structure and function of the brain by analysing the shapes and sizes of different brain structures in an MRI image. Morphometric changes in different brain structures are associated with many neurological and psychiatric disorders, for example Alzheimer's disease. Tracking these changes automatically using automated segmentation methods would aid in diagnosing a particular brain disease and follow its progression. In this thesis we present a method for automatic segmentation of MRI brain scans using parametric generative models and Bayesian inference. Our method segments a given MRI scan to 41 different structures including for example hippocampus, thalamus and ventricles. In contrast to the current state-of-the-art methods in whole-brain segmentation, our method does not pose any constraints on the MRI scanning protocol used to acquire the images. Our model is based on two parts: the first part is a labeling model that models the anatomy of the brain and the second part is an imaging model that relates the label images to intensity images. Using these models and Bayesian inference we can find the most probable segmentation of a given MRI scan. We show how to train the labeling model using manual segmentations performed by experts and how to find optimal imaging model parameters using expectation-maximization (EM) optimizer.
We compare our automated segmentations against expert segmentations by means of Dice scores and point out places for improvement. We then extend the labeling and imaging models and show, using a database consisting of MRI scans of 30 subjects, that the new models improve the segmentations compared to the original models. Finally we compare our method against the current state-of-the-art segmentation methods.
The results show that the new models are an improvement over the old ones, and compare fairly well against other automated segmentation methods. This is encouraging, because there is still room for improvement in our models. The labeling model was trained using only nine expert segmentations, which is quite a small amount, and the automated segmentations should improve as the number of training samples grows. The upside of our method is that it is fast and generalizes straightforwardly to MRI images with varying contrast properties.
