Automatic lung field segmentation using novel feature extraction and unsupervised learning

Obtaining accurate and automated lung field segmentation is a challenging step in the development of Computer-Aided Diagnosis (CAD) system. In this paper, fully automatic lung field segmentation is proposed. Initially, novel features are extracted by considering spatial interaction of the neighbouring pixels. Then constrained non-negative matrix factorization (CNMF) factorized the data matrix obtained from the visual appearance model into basis and coefficient matrices. Initial lung segmentation is achieved by applying unsupervised learning on the coefficient matrix. 2-D region growing operation removes trachea and bronchi appearing in the initial lung segmentation. The experimental results on different database shows that the proposed method produces significant DSC 0.973 as compared to the existing lung segmentation algorithms.