Segmentação automática do fígado e lesões hepáticas em imagens de tomografia computadorizada usando redes neurais convolucionais

Abstract

Liver cancer is one of the major causes of death by cancer. It is known that the early detection of lesions in the liver provides a better chance of treatment and cure of the disease. For this detection, computed tomography (CT) exams are commonly used, from which specialist doctors perform manual segmentation of the liver and lesions. However, this segmentation is time-consuming and prone to errors and variations between specialists. Due to this hard work, Computer-aided detection systems have been developed to assist specialists in the segmentation of the liver and in detecting and characterizing of liver lesions and thus reduce the required time for diagnosis. However, automatic segmentation of the liver is a complex task, as the liver has variability in shape, ill-defined borders and lesions can affect its appearance. The automatic segmentation of the lesions becomes more complex because they present variability in contrast, shape, size and location. This work proposes a method for liver segmentation and a method for liver lesions segmentation. The liver segmentation method, which is based on deep convolutional neural network models, consists of four main steps: (1) image pre-processing, (2) initial segmentation, (3) reconstruction and (4) final segmentation. The method for liver lesions segmentation, which is also based on models of deep convolutional neural networks, consists of three main steps: (1) initial segmentation, (2) final segmentation and (3) segmentation refinement. The methods were evaluated in a set of 131 CT images from the LiTS database. The liver segmentation method obtained a sensitivity of 95.45%, specificity of 99.86%, Dice coefficient of 95.64%, VOE of 8.28% and RVD of -0.41%. The method for liver lesions segmentation, when the liver is marked by the specialist, obtained a sensitivity of 84.52%, specificity of 99.96%, Dice coefficient of 83.84%, VOE of 27.19% and RVD of -0.72%. When using the liver marked by the proposed method, the method for liver lesions segmentation obtained a sensitivity of 82.20%, specificity of 99.95%, Dice coefficient of 80.60%, VOE of 30.75% and RVD of -0.25%. From these results, it is demonstrated that the problem of liver and lesions segmentation in CT images can be efficiently solved using deep convolutional neural networks to reduce the scope of the problem and segmentation of the liver and lesions.