Segmentação de gorduras cardíacas utilizando uma rede generativa adversária condicional

Abstract

In recent years, studies have shown that the increase in the amount of fat surrounding the heart was associated with a higher risk of triggering some cardiovascular diseases, such as atrial brillation and coronary heart disease. Manual segmentation of these fats has not been widely implemented in clinical practice due to the human workload required and the high cost of physicians and technicians. Therefore, the need to perform quantitative analysis more accurately and faster in exams with a lot of information has driven the development of new computational methods for organ segmentation. In this work, a new uni ed method for the autonomous segmentation and quanti cation of two types of cardiac fats was proposed, using the pix2pix network, a generative conditional adversary network ideally created to perform image-to-image translation. Segmented fats are called epicardium and mediastinal and are separated from each other by the pericardium. Different experiments were carried out to obtain satisfactory results regarding segmentation, explore the most suitable approaches when working with this network and evaluate the performance of a new, powerful and versatile architecture. Experimental results of the proposed methodology showed that the mean accuracy in relation to epicardial and mediastinal fats is 99.08%, with a true positive mean rate of 99.34%. The similarity indices were, on average, 99.28% and 98.59%, for the F1 score and IoU, respectively.