A Novel Deep Learning Network and Its Application for Pulmonary Nodule Segmentation

Abstract

Pulmonary nodules are the early manifestation of lung cancer, which appear as circular shadow of no more than 3 cm on the computed tomography (CT) image. Accurate segmentation of the contours of pulmonary nodules can help doctors improve the efficiency of diagnosis. Deep learning has achieved great success in computer vision. In this study, we propose a novel network for pulmonary nodule segmentation from CT images based on U-NET. The proposed network has two merits: one is that it introduces dense connection to transfer and utilize features. Additionally, the problem of gradient disappearance can be avoided. The second is that it introduces a new loss function which is tolerance on the pixels near the borders of the nodule. Experimental results show that the proposed network at least achieves 1% improvement compared with other state-of-art networks in terms of different criteria.