A deep learning-shape driven level set synergism for pulmonary nodule segmentation

Abstract

Accurate pulmonary nodule segmentation, an essential pre-requisite in every computer-aided diagnosis (CAD) system, significantly helps in the risk assessment of lung cancer. In this paper, we propose a synergistic combination of deep learning and shape driven level sets for automated and accurate lung nodule segmentation. A coarse-to-fine solution is adopted, where, a deep fully convolutional network is employed to obtain coarse segmentation. To achieve fine segmentation, shape driven evolution of level sets is designed. The seed points for initializing the level sets are obtained from the coarse segmentation of deep network in an automated manner. Perimeter and circularity of the evolving contours are employed for guiding the evolution of level sets. Experiments on the publicly available LIDC/IDRI dataset clearly reveal that our method outperforms several state-of-the-art competitors as well as its constituent parts, i.e., deep network and level set, when applied in isolation.