| This thesis addresses the task of automatically segmenting brain tumors and edema in magnetic resonance images. This is motivated by potential applications in assessing tumor growth, assessing treatment responses, enhancing computer-assisted surgery, planning radiation therapy, and constructing tumor growth models. The presented framework forms an image processing pipeline, consisting of noise reduction, spatial registration, intensity standardization, feature extraction, pixel classification, and label relaxation. The key advantage of this framework is the simultaneous use of features computed from the image intensity properties, and the locations of pixels within an aligned template brain. Automatically learning to combine these features allows recognition of tumors and edema that have relatively normal intensity properties. Our results on 11 patients with brain tumors show that the system achieves nearly perfect performance given patient-specific training, but also achieves accurate results in segmenting patients not used in training. |
