Segmentation is an important step in the interpretation of Magnetic Resonance (MR) images of the human body. MRI reveals an unequaled view of the anatomy of the brain in terms of spatial and contrast resolution, and its multispectral nature has been exploited to obtain better performance in the segmentation process. This thesis presents new techniques based on artificial neural network (ANN) architectures for automatic segmentation and tissue classification of MR images of the human brain. Two different methodologies were adapted for supervised and unsupervised segmentation.

The Learning Vector Quantization (LVQ) ANN is utilized for multispectral supervised classification of MR images. The original LVQ was modified for better and more accurate classification. LVQ ANN segmentation results are compared to those achieved with a back propogation ANN and a conventional Maximum Likelihood Classifier (MLC).

In the second scheme a fully automated technique was developed for segmentation. The scheme utilizes the Self Organizing Feature Map (SOFM) ANN for feature mapping and generates a set of codebook vectors for each tissue class. An additional layer then completes the classification process. To minimize clustering artifacts, an algorithm has been developed for isolating the cerebrum prior to segmentation. The cerebrum is extracted by stripping away the skull pixels from the T2 weighted image. The network is tested for different sets of image slices from normal and abnormal brain studies. Images were selected from 54 axial images of the whole head.

Twenty nine brain studies were analyzed using the techniques developed in this thesis. Three tissue types of the brain are segmented:​ white matter, gray matter, and cerebrospinal fluid (CSF);​ in case of abnormality, the tumor or other unknown tissues were also segmented. From the evaluation of segmentation results, the advantages and disadvantages of each method are discussed.