Modelo computacional para classificação de nódulos pulmonares utilizando redes neurais convolucionais

Abstract

Cancer is a disease that occurs when the organism’s cells suddenly begin to grow disorderly and over time invading tissues and organs, spreading (metastasizing) to other regions of the body. Among the more than 100 cancers, lung cancer is just the one that causes more deaths in the world, wherein every 5 deaths, 1 is caused by this cancer. However, if the diagnosis is made at the beginning of the disease, the 1-year survival rates are approximately of 15-19%. The main manifestation of lung cancer is the initiation of the pulmonary nodule, which hás larger diameters between 3mm and 30mm. The diagnosis of the pulmonary nodule is performed mainly through Computed Tomography (CT) images, however, performing this diagnosis still represents a complex and challenging task for the specialists, since the nodules can be located in complex structures of the lung and the number of exams or images that come to be examined is increasing. Thus, it is essential to integrate a tool of Computer Diagnostic Assistance (CADx) in the interpretation of medical images, which aims to act as a second opinion to specialists through a suggestion provided by a computational model. There are a number of papers proposed in the literature to better assist specialists in pulmonary nodule diagnosis, whether using radiomics attributes associated with a machine learning technique or using deep learning techniques, which have called the attention both in the scientific community as well as in industry. More specifically, Convolutional Neural Networks (CNNs) has become a trend for images. However, there is still a search for a computational model that improves the performance of CADx systems. In this context, the objective of this work was to research and develop a computational model to classify lung nodules in benign or malignant using a CNN together with a hyperparametric optimization technique. The model obtained the following results: sensitivity of 95%, specificity of 100%, an accuracy of 85% and under the ROC curve (AUC) of 0.93 in a set of medical images composed of solid pulmonary nodules of CT with a diameter between 3mm and 30mm. The results showed the importance of analyzing the parenchyma region in a very close proportion of the nodules to achieve a higher performance in the classification of the pulmonary nodules, and also that the analysis of the isolated parenchyma leads to better results than analyzing only the nodule or the nodule with the parenchyma.