The effect of kernel size of CNNs for lung nodule classification

Early detection in lung nodule will be helpful for lung cancer diagnosis. Computer-aided detection (CAD) system to automatic detection of pulmonary nodules is one of the most effective methods to decrease the burden on radiologists where they have to analyze a huge number of thoracic Computed Tomography (CT) scans to find out suspicious nodules. Lung nodule classification is crucial to implement a trustable lung nodule detection system. With the rapid development of deep learning in the field of object recognition, good performance on lung nodules classification has been achieved with Convolutional Neural Network (CNN). In this study, we propose three CNN architectures which are adapted to represent small, normal and large networks. We implement different CNN architectures with various kernel sizes to compare the performances of different combinations of CNN architectures and convolution kernels. The method is evaluated on the public Lung Image Database Consortium (LIDC) dataset of 1018 patients scans. The experiment shows the relation of convolution layers and kernel size has affection on the sensitivity of result in our model. The proposed method achieved a sensitivity of 88.22%∼94.18%.