Dual Encoding Fusion For Atypical Lung Nodule Segmentation

Cavitary lung nodules are the main hindrance to the clinical application of lung-nodule-segmentation methods. In this study, to avoid the disturbance of high granularity in computed tomography images, we design a novel Dual Encoding Fusion Network (DEF-Net) for atypical lung nodule segmentation. DEF-Net is composed of three parts: multi-level representations, dual encoding, and fusion decoding. The global and local characteristics of lung nodules can be better captured by fusing the two encodings of the original inputs. Furthermore, we design two dilated residual blocks and incorporate them into the dual encoding process, which can recognize the morphology of unusual lung nodules at multiple scales. We evaluate the proposed DEF-Net on the Lung Image Database Consortium and Image Database Resource Initiative dataset; experimental results demonstrate that our DEF-Net can achieve excellent segmentation for atypical lung nodules and consistently outperform state-of-the-art approaches.