Lung CT Screening With 3D Convolutional Neural Network Architectures

Abstract

The standard tests for diagnosis of pulmonary cancer are imaging, sputum cytology and lung biopsy, with chest computed tomography (CT) playing a major role in the early detection of nodules, which increases the patients survival. The challenge is to analyze these images automatically, for example, the nodules density often resembles other pulmonary structures evidenced in CTs. This paper proposes an automated algorithm to classify pulmonary nodules into benign or malignant. Our contribution is to design and test 3D Convolutional Neural Networks using a public CT image collection, optimize the results of the proposed approach considering varying input sizes and numbers of convolutional layers, as well as compare with several previous approaches on CT analysis. Promising results show accuracy of 0.9040, kappa of 0.7624, sensitivity of 0.8630, specificity of 0.9191 and AUC of 0.8911 during malignant nodule detection.