Chest X-ray segmentation using Sauvola thresholding and Gaussian derivatives responses

Abstract

This paper presents a simple, flexible and an effective lung segmentation technique called ST-GD (Sauvola thresholding-Gaussian derivatives) method. In this technique Sauvola thresholding method and four Gaussian derivatives responses are used. This technique for extraction of lung field area is consist of six main steps. (1) For the purpose of enhancement the image is preprocessed. This is achieved by using adaptive contrast enhancement and normalization. (2) The average image is calculated from a Gaussian derivatives of four different magnitudes in such a way that it highlights the outer boundary of the lung region. (3) Preprocessed image is then thresholded by using Sauvola image thresholding which mostly highlights the inner area of the lung region. (4) To emphasize the lung region completely the Sauvola thresholded image and gradient average image is combined. (5) Once the image is combined, to remove the noisy area such as trachea, clavicle region and outer body, XOR is taken between similar X-rays average image and combined image. (6) Finally, morphology is used to remove the noise that has been occurred during the formation of lung shape. This developed system tested on JSRT, Montgomery and a self collected dataset. The self-collected database has been collected from Northwest General Hospital and Research Center, Peshawar, Pakistan. The proposed system produced an accuracy of 94.57% on JSRT dataset, 90.75% accuracy on Montgomery dataset and 65.25% on Northwest dataset using Jaccard coefficient. Furthermore, it is also investigated that the proposed study has outperformed as compared to the state-of-the-art methods.