Segmentation of White Matter Lesions – Using Multispectral MRI and Cascade of Support Vector Machines with Active Learning.

Segmentation of White Matter LesionsUsing Multispectral MRI and Cascade of Support Vector Machines with Active LearningSoheil DamangirDecember 2011The areas in cerebral white matter that appear hyperintense on T2-weighted (T2) magnetic resonance image (MRI) and hypointense on computed tomography (CT) are commonly referred as white matter lesions (WML). WML have been consistently related to vascular risk factors and they could have predictive value for future stroke, dementia and functional decline in activities of daily living.Although, there is an important body of evidence regarding the importance of WML, the segmentation of these changes on different MRI modalities is not yet, validated and standardized.The aim of this thesis was to segment the WML using multispectral MRIs based on a fully automated method.A new method using a cascade of support vector machine (SVM) classifiers with active learning is proposed for the segmentation of the WML. It has been shown that this training method not only increase the speed of classification, but also helps to boost the accuracy of classification in biased datasets.The classification method is put together with preprocessing and post-processing to form a general segmentation framework. To validate the method, a model was trained using two subjects and tested against the remaining 100 subjects. Results were validated against manually outlined WML which is the method that could provide replicable results between studies so far. Also, comparisons with other automatic segmentation procedures were performed. Results of this study proved to be comparable with the results from manual outlining in terms of both sensitivity and specificity