| Finding subcortical structures accurately in 3D brain images is critical for disease analysis and verification of treatment. Most studies, however, rely on manual segmentation by human experts, which is both time consuming and produces inter-rater reliability issues. We introduced two automatic subcortical brain image segmentation algorithms, Ada-SVM and the auto context model, based on fast and accurate machine learning principles. These methods expand upon a highly effective machine learning technique called AdaBoost. Our algorithms require a training set of hand labeled data. From this training set, we learn a model and use that model to segment new data. After algorithm development, we validated both methods by using metrics that compare automated and manual segmentations, and we examined how well the methods detected disease related differences in the brain. We used standard volumetric and distance based metrics to compare our automated segmentations to manual segmentations. In addition to agreement with manual gold standard data, we showed that our algorithm predicted cognitive decline known to be associated with hippocampal degeneration in Alzheimer's disease (AD). This was validated by showing both a decline in hippocampal volume from normal elderly controls to patients with mild cognitive impairment (a transitional state carrying increased risk of conversion to AD) to AD patients and by visualizing the 3D pattern of deficits. To do this, we used a radial atrophy mapping technique that assigns a p-value to each point on the hippocampal surface (p-maps). Next, we used our algorithm to segment hippocampi for a practical study of AD. As a related novel project, we developed an online learning algorithm that losslessly combines weak learners to form a decision rule (called Lossless Online Ensemble Learning, or LOEL). We compared our online learning algorithm to other online learning methods---including both online AdaBoost and online bagging---and we showed that LOEL outperforms both on synthetic data and for hippocampal segmentation. Finally, we showed that the auto context model is effective at segmenting multiple sclerosis lesions from five-channel brain MRI data. Our results suggest that our approach is an effective MS lesion segmentation algorithm, even without domain-specific knowledge regarding MS lesions. |
