Lung segmentation using a fully convolutional neural network with weekly supervision

Most supervised methods for lung CT image segmentation require all training examples to be labeled with segmentation masks. This requirement makes it expensive to annotate varies categories of lung diseases. The goal of this paper is to propose a new weekly supervised training scheme, together with a lung patch feature extraction method, that enables training segmentation models on a large set of self-generated texture mosaics images, but only a small fraction of which have mask annotations. Such feature extraction is implemented by the empirical wavelet transform (EWT) followed by a fully convolutional neural network which consists of the final segmentation step. A generative adversarial networks (GAN) based partial supervised learning is also utilized to further refine the correction of the segmentation result. Our method deals with lung segmentation issue under normal or severe pathological conditions. The proposed method is tested on two public datasets and our experiment results without heavy work of mask annotations give similar results compared with the approach with fully labeled mask.