Segmentação automática de estruturas cardíacas em imagens de cine-ressonância magnética de eixo curto usando redes neurais convolucionais

Abstract

Cardiovascular diseases are responsible for millions of deaths every year. In this scenario, non-invasive exams such as cine-magnetic resonance imaging (cine-RM) have favored a better understanding of these pathologies, helping early diagnosis and previous treatments essential to improve the quality of life of individuals. Through this exam, specialists can obtain more accurate information about cardiac structures, including the myocardium (Myo), the left ventricular cavity (LVC) and the right ventricle (RV). Given this context, this work presents automatic methods for the segmentation of those cardiac structures in short-axis cine-MRI images. Two methods are proposed. The first, called Cascaded Segmentation with Reconstruction, is divided into three main steps. The first step consists in extracting a region of interest (ROI) to reduce the scope of processing. The second applies a fully convolutional network (FCN) proposed to generate the initial Myo, LVC and RV segmentations. These initial segmentations are passed on to the third step, called refinement, in which a mask reconstruction module based on U-Net is used to restore the generated segmentations. In addition, in this step some specific post-processing techniques are also applied for each structure of interest. The second method, called Specialized Segmentation by Context, is similarly divided into three steps, the first of which also focuses on extracting an ROI; the second step comprises a combination of FCNs for Myo and LCV segmentation; and the third step uses another proposed FCN, based on X-Net, for RV segmentation. The methods developed reach promising results in tests with the dataset made available by the ACDC challenge, both at the local level and in the evaluation made by the challenge’s online platform, in which the proposed methods present results with little difference to the best approaches. Expressive results are also obtained in tests with the M&Ms dataset, which contains more exams than the ACDC, which are acquired by scanners from different vendors, therefore having a greater variability of features, similar to the real application scenario.