A deep unsupervised saliency model for lung segmentation in chest X-ray images

Abstract

Medical image segmentation is a critical and laborious step in the diagnosis workflow by separating regions of organs or anomalies to be analyzed further by a physician, or classified automatically by an intelligent tool. Deep machine learning techniques have achieved impressive results in image segmentation tasks whenever immense amounts of labeled data are provided. However, having big amounts of manually supervised image data is labor intensive, and costly impractical. Unsupervised learning, contrariwise, can provide results using unlabeled images by reaching decisions, clustering, and deriving insights from the data only. We propose a deep unsupervised learning approach to perform lung segmentation in chest X-ray images, which is guided by heuristic saliency maps as a noisy labeling framework, which through a composite loss of reconstruction and regularization reach saliency estimation as final segmentation. Tests were performed using public chest X-ray images from the Japanese Society of Radiological Technology (JSRT), and from Montgomery County (MC) datasets. The proposed model based on saliency scored a dice coefficient of 0.87, against a state-of-the-art unsupervised method, which scored 0.68, on JSRT, and on the MC dataset 0.88, and 0.64, respectively. Qualitatively comparing, the saliency model worked better in reducing background noise. Considering that unsupervised models can be used as robust preparation steps, and as baselines for further different tasks, either guided by constant feedback from physicians, or as input to semi-supervised learning, the results have shown this to be a promise line of research to medical image segmentation problems.