| An important goal in the analysis of magnetic resonance images is to automatically determine the image regions that correspond to particular tissue classes. Algorithms for solving this problem, called image segmentation algorithms, play a critical role in applications such as presurgical planning, computer integrated surgery, computer aided diagnosis, and the study of anatomical structure. However, methods for segmenting MR images are hindered by the presence of artifacts such as noise, intensity inhomogeneities, and partial volume effects. The research presented in this dissertation focuses on the development of methods for improving and computing segmentations of MR images. First, we present a technique for statistically characterizing magnetic resonance tissue properties. The results of this technique provides information that allows optimization of the MR acquisition for obtaining accurate segmentations. Next, we present AFCM, an algorithm that segments MR images while being robust to intensity inhomogeneities. This method also has the advantage of yielding a soft segmentation that provides additional information about where partial volume effects are present in the image. We also propose another algorithm, AGEM, which is based on a statistical model of an MR image. AGEM provides accurate segmentations not only in the presence of intensity inhomogeneities and but also in excessive noise. Finally, we present some preliminary research directed towards isolating the response of the MR acquisition to each tissue class in the image. This results in an image representation that retains nearly all information from the original image while localizing the anatomical structure. |
