Automatic lung segmentation in chest radiographs using shadow filter and multilevel thresholding

Abstract

Lung segmentation in chest radiographs is a requisite pre-processing step in the Computer-aided Diagnosis (CAD) system for the detection of chest diseases. This paper proposes an unsupervised lung segmentation method in chest radiographs based on shadow filter and multilevel thresholding. The method consists of three main processes: pre-processing, initial lung field estimation and noise elimination. For First step, resize the original image and adjust contrast. Then, shadow filter is applied to enhance each lung outlines. After that, the initial lung fields are estimated by using local thresholding, delete outer body regions, fill holes, filter regions from their properties, and using multilevel thresholding to remove unwanted regions. Finally, morphological operations and refine edge techniques are used to eliminate the noise in the result. All 247 chest radiographs from a public JSRT dataset were used to evaluate the performance. The performance measures of the proposed method (overlap, accuracy, sensitivity, specificity, precision, and F-score) are above 91% and the average computation time for 512 by 512 pixels resolutions is 8.46 seconds, improved from our previous work. The accuracy and overlap are 97.28% and 91.36%, respectively. The results show that our proposed unsupervised method is performed accurately.