Lung Fields Segmentation Based on Shape Compactness in Chest X-Ray Images

Lung Fields Segmentation Based on Shape Compactness in Chest X-Ray Images

Computer-Aided Diagnosis (CAD) benefits from its early diagnosis and accurate treatment. As the preprocessing step of CAD-based chest radiograph analysis, lung segmentation affects the precision of lesion recognition and classification. With the development of artificial intelligent technologies, a lot of powerful algorithms based on machine learning, such as convolutional neural networks, are used to extract lung areas from X-ray images. However, these state-of-the-art segmentation algorithms have become inapplicable with limited training data, varied boundaries and poor contrasts. In order to overcome these problems, this paper proposes a novel lung segmentation method which integrates Graph-cut and neural network. Different from traditional methods, the proposed method is designed with an energy function which involves a shape compactness, and the conditional probabilities are calculated according to the outputs of U-Net. Furthermore, the objective function is transformed into an iterative form and decomposed into a series of easier sub-problems based on ADMM algorithm, which is used to reduce the complexity of high-order optimization. Compared with the previous methods on JSRT dataset, the segmentation results of the proposed method show a higher Dice-Coefficient. By using the proposed method, we can achieve 97.1% accuracy compared to 94.87% using the baseline U-Net model, and the segmentation accuracy of each image in JSRT dataset is improved.