Deep Neural Networks for Semantic Segmentation of Lung Nodules

Abstract

Segmentation of lung nodule from computed tomography (CT) images is of great importance for lung nodule analysis. In order to improve the segmentation accuracy of lung nodules, a lot of research work has been done. The two main challenges in segmentation of lung nodules are lung nodules have the similar characteristics as surrounding tissues and lung nodules are extremely unbalanced with the background. In this study, we investigated the performance of U-Net to segment the lung nodules in CT images. Two imaging pre-process methods, image resizing and intensity normalization, are applied to alleviate the problem of small nodules in lung and improve the quality of training images. In addition, to increase the diversity and variation of training data, we propose a fusion over-sampling method to augment the data. The dice coefficient loss function was also introduced to solve the problem of imbalance distribution of positive and negative samples. Experiments on the LUNA16 dataset demonstrate that this method can improve the performance of lung nodules segmentation network, which is better than the state-of -the-art methods in the aspect of lung nodules segmentation.