Unsupervised automated segmentation of cerebral magnetic resonance images using multi-layer self-organizing feature maps

This dissertation presents an unsupervised approach for segmentation using a multi-layer self-organizing feature map. Tissue classification of brain magnetic resonance images for quantitative analysis is a challenging problem. Manual classification suffers from inter and intra-expert variability. This research achieves the major goal of reducing expert variability and training data required for supervised segmentation and classification. The performance of a segmentation and classification system is presented for clinical data and synthetic data generated via a computer model of the brain. The results are also analyzed in the context of information theory. Mutual information is calculated and used to compare classification results of different approaches or system configurations. Networks that generate greater values of mutual information yield better classification results. For the interactive experts in this study, the multi-layer approach completely eliminated expert variability in 80% of the combined clinical and synthetic database while eliminating training data and expert interaction required for supervised learning.