Segmentação do miocárdio em imagens de MRI cardíaca utilizando redes neurais convolutivas

Abstract

Cardiovascular diseases are the leading cause of death worldwide. Noninvasive cardiac imaging technologies, such as magnetic resonance, are essential tools to support the diagnosis and monitoring of various pathologies. The previous step for the extraction of cardiac function indicators is the endocardium and epicardium contours segmentation in the left ventricular cavity. This process often is performed manually by the specialists, which requires a lot of time and effort, and is prone to intra and inter-observer errors. This dissertation develops an automatic methodology based on a fully convolutional neural network to segment the myocardium in short axis cardiac magnetic resonance images. The database used is divided into 10 sets for training and testing purposes. Six optimization methods are evaluated: stochastic gradient descend, Nesterov accelerated gradient, RMSProp, Adam, AdaDelta and AdaGrad. The best results were achieved with the stochastic gradient descend and RMSProp. With the former, a Dice coefficient of 0.9055 and 0.9146, Hausdorff distance of 10.5244 and 10.7240, sensitivity of 0.9263 and 0.9135, specificity of 0.9985 and 0.9986 were obtained for endocardium and epicardium, respectively. With RMSProp, a Dice coefficient of 0.9098 and 0.9167, Hausdorff distance of 9.0421 and 9.7663, sensitivity of 0.9200 and 0.9116, specificity of 0.9988 and 0.9987 were obtained for endocardium and epicardium, respectively.