This thesis is focused on automatic detection of white matter lesions (WML) in Fluid Attenuation Inversion Recovery (FLAIR) Magnetic Resonance Images (MRI) of the brain. There is growing interest within the medical community regarding WML, since the total WML volume per patient (lesion load) was shown to be related to future stroke as well as carotid disease. Manual segmentation of WML is time consuming, labourious, observer-dependent and error prone. Automatic WML segmentation algorithms can be used instead since they give way to lesion load computation in a quantitative, efficient, reproducible and reliable manner. FLAIR MRI are affected by at least two types of degradations, including additive noise and the partial volume averaging (PVA) artifact, which affect the accuracy of automated algorithms. Model-based methods that rely on Gaussian distributions have been extensively used to handle these two distortions, but are not applicable to FLAIR with WML. The distribution of noise in multicoil FLAIR MRI is non-Gaussian and the presence of WML modifies tissue distributions in a manner that is difficult to model. To this end, the current thesis presents a novel way to model PVA artifacts in the presence of noise. The method is a generalized and adaptive approach, that was applied to a variety of MRI weightings (with and without pathology) for robust PVA quantification and tissue segmentation. No a priori assumptions are needed regarding class distributions and no training samples or initialization parameters are required. Segmentation experiments were completed using simulated and real FLAIR MRI. Simulated images were generated with noise and PVA distortions using realistic brain and pathology models. Real images were obtained from Sunnybrook Health Sciences Centre and WML ground truth was generated through a manual segmentation experiment. The average DSC was found to be 0.99 and 0.83 for simulated and real images, respectively. A lesion load study was performed that examined interhemispheric WML volume for each patient. To show the generalized nature of the approach, the proposed technique was also employed on pathology-free T1 and T2 MRI. Validation studies show the proposed framework is classifying PVA robustly and tissue classes are segmented with good results.

This thesis is focused on automatic detection of white matter lesions (WML) in Fluid Attenuation Inversion Recovery (FLAIR) Magnetic Resonance Images (MRI) of the brain. There is growing interest within the medical community regarding WML, since the total WML volume per patient (lesion load) was shown to be related to future stroke as well as carotid disease. Manual segmentation of WML is time consuming, labourious, observer-dependent and error prone. Automatic WML segmentation algorithms can be used instead since they give way to lesion load computation in a quantitative, efficient, reproducible and reliable manner. FLAIR MRI are affected by at least two types of degradations, including additive noise and the partial volume averaging (PVA) artifact, which affect the accuracy of automated algorithms. Model-based methods that rely on Gaussian distributions have been extensively used to handle these two distortions, but are not applicable to FLAIR with WML. The distribution of noise in multicoil FLAIR MRI is non-Gaussian and the presence of WML modifies tissue distributions in a manner that is difficult to model. To this end, the current thesis presents a novel way to model PVA artifacts in the presence of noise. The method is a generalized and adaptive approach, that was applied to a variety of MRI weightings (with and without pathology) for robust PVA quantification and tissue segmentation. No a priori assumptions are needed regarding class distributions and no training samples or initialization parameters are required. Segmentation experiments were completed using simulated and real FLAIR MRI. Simulated images were generated with noise and PVA distortions using realistic brain and pathology models. Real images were obtained from Sunnybrook Health Sciences Centre and WML ground truth was generated through a manual segmentation experiment. The average DSC was found to be 0.99 and 0.83 for simulated and real images, respectively. A lesion load study was performed that examined interhemispheric WML volume for each patient. To show the generalized nature of the approach, the proposed technique was also employed on pathology-free T1 and T2 MRI. Validation studies show the proposed framework is classifying PVA robustly and tissue classes are segmented with good results.

[pt] As lesões da substância branca são lesões cerebrais não estáticas que têm uma taxa de prevalência de até 98 por cento na população idosa, embora também esteja presente na população jovem. Uma vez que elas podem estar associadas a várias doenças cerebrais, é importante detectá-las o mais cedo possível. A ressonância magnética fornece dados tridimensionais para visualização e análise de tecidos moles, pois contém informações ricas sobre sua anatomia. No entanto, a quantidade de dados adquiridos para essas imagens pode ser excessiva para análise / interpretação manual, representando uma tarefa difícil e demorada para especialistas. Portanto, esta tese de doutorado apresenta quatro novos métodos computacionais para detectar automaticamente lesões de substância branca em imagens de ressonância magnética, baseadas principalmente nos algoritmos SLIC0 e Convolutional Neural Networks. Nosso principal objetivo é fornecer as ferramentas necessárias para que os especialistas acelerem seus trabalhos e sugiram uma segunda opinião. Dos quatro métodos propostos, o que obteve melhores resultados foi aplicado em 91 imagens de ressonância magnética, e obteve uma precisão de 97,93 por cento, especificidade de 98,02 por cento e sensibilidade de 90,12 por cento, sem utilizar nenhuma técnica de redução de candidatos. / [en] White matter lesions are non-static brain lesions that have a prevalence rate up to 98 percent in the elder population, although it is also present in the young population. Because it may be associated with several brain diseases, it is important to detect them as early as possible. Magnetic resonance imaging provides threedimensional data for visualization and analysis of soft tissues as it contains rich information about their anatomy. However, the amount of data acquired for these images may be too much for manual analysis/interpretation alone, representing a difficult and time-consuming task for specialists. Therefore, this doctoral thesis presents four new computational methods to automatically detect white matter lesions in magnetic resonance images, based mainly on algorithms SLIC0 and Convolutional Neural Networks. Our primary objective is to provide the necessary tools for specialists to accelerate their works and suggest a second opinion. From the four proposed methods, the one that achieved best results was applied on 91 magnetic resonance images, and achieved an accuracy of 97.93 percent, specificity of 98,02 percent and sensitivity of 90,12 percent, without using any candidate reduction techniques.