Segmentação Automática de Lesões de Esclerose Múltipla em Imagens de Ressonância Magnética

• Autor(a) principal: Rafaela Inês Pires Pinto
• Data de Publicação: 2017
• Tipo de documento: Dissertação
• Idioma: por
• Título da fonte: Repositório Científico de Acesso Aberto de Portugal (Repositórios Cientìficos)
• Texto Completo: https://hdl.handle.net/10216/105286
• topic: Ciências médicas e da saúde; Medical and Health sciences

Multiple sclerosis is the most commonly diagnosed neurological disorder in young adults with unexplained causes and major repercussions in the lives of patients, urging researchers to actively search for answers. Although the disease cannot be cured or prevented, the available treatments nowadays reduce its severity and delay its progression. It is becoming increasingly necessary to use imaging techniques and also image processing and analysis techniques, to help doctors perform an early diagnosis and start appropriate treatment in order to provide a better quality of life for the patient. Several approaches based on automatic segmentation of multiple sclerosis lesions have been extensively investigated in recent years for this purpose. This project was developed, firstly, with the recognition of the steps necessary to implement and optimize an image processing and analysis methodology for automatic segmentation of MS lesions, and secondly, by the exploration of pre-processing, segmentation and classification techniques for objective and quantitative characterization of the lesions. This work will also be discussed basic concepts of multiple sclerosis disease and magnetic resonance imaging (MRI), as well as the bibliographical study of some of the currently existing methodologies. The methodology developed in this dissertation was based on the implementation of several pre-processing algorithms for noise smoothing and removal, non-cerebral tissue removal, contrast correction and normalization of images intensity. For lesion segmentation was applied to the study of neural networks, a very promising and current approach to the proposed problem, and to classify were extracted and analyzed some characteristics of the lesions through shape and size. It is intended that this new methodology is flexible and allow the testing and analysis of the results. The results obtained demonstrate that pre-processing techniques are essential to the subsequent steps allowing better image quality. Segmentation of lesions through the use of neural networks proved to be appropriate for this study as shown by the metrics analyzed, with a structural similarity index very close to 1, mean absolute error rate of 3.8% and Dice coefficient of 0.58. Finally, the various practical applications performed was possible to demonstrate the usefulness and adequacy of image processing and analysis techniques in the study and detection of multiple sclerosis lesions through MR images.