Segmentação da região pulmonar em imagens de radiografia torácica utilizando redes neurais convolutivas

Abstract

According to data from the National Cancer Institute, lung cancer is one of the most frequent tumors in the Brazilian population. The process for its diagnosis sometimes involves the need to segment the pulmonary region in an image exam, a phase that requires hours from a medical professional. Therefore, the use of tools that apply automated techniques to accomplish this task could help them. This dissertation develops an automatic methodology, based on convolutive neural networks, to segment the lung region in chest X-ray images. Three architectures are developed (CNN1, CNN2 and CNN3), where the CNN1 and CNN2 architectures are of direct network, while the CNN3 architecture is a topology of directed acyclic graphs (DAG). In conjunction with the architectures, three different regularization methods (Dropout, L2 and Dropout + L2) and three different optimization methods (SGDM, RMSPROP and ADAM) are investigated. The database used for this study is the JSRT - Japanese Society of Radiological Technology, which contains 247 images of chest radiography. As a way of measuring the performance of the studied networks, six performance metrics were used, they are: Global Accuracy, Accuracy, Jaccard Coefficient, Weighted Jaccard Coefficient, Score F1 and Dice Index. At the end of all simulations, the best results were achieved using the CNN3 network, which makes use of the DAG topology, together with the Dropout + L2 regularization method and the ADAM optimization method. The metrics obtained were: Global Accuracy equal to 0.99139 ± 0.00098; Accuracy equal to 0.98927 ± 0.00161; Jaccard coefficient of 0.97967 ± 0.00232; Weighted Jaccard coefficient equal to 0.98294 ± 0.00191; F1 Score of 0.97475 ± 0.00357 and, finally, a Dice Index of 0.98921 ± 0.00163.