Deep Learning for Segmentation of Solid and Non-Solid Pulmonary Nodules in CT

In this study, we aimed to segment three types lung nodules (solid, part-solid, non-solid) using the deep learning. A total of 25,993 images and 9,159 images were used for experiments on the solid and non-solid regions, respectively. Deep learning was performed using the U-Net architecture based on the tensorflow. In this study, we have validated the deep learning models for the segmentation of the solid and non-solid regions with test sets. As a result, the solid region showed an average sensitivity of 91% and an average DSC of 0.9 and the non-solid region showed an average sensitivity of 82% and an average DSC of 0.8. In CT, deep learning based models showed relatively high accuracy for all solid, part solid, and non-solid types of lung nodule.