Feature Selection for Contour-based Tuberculosis Detection from Chest X-Ray Images

Although the mortality rate of Tuberculosis (TB) has been reduced dramatically over the past decade, detection of TB remains a challenge and an active research problem in recent years. Compared to more advanced imaging systems such as SPECT, CT, and PET scan, the routine Chest X-Ray (CXR) imaging is a prevalent, inexpensive, fast, and with a much less radiation doses for detecting TB patients. Since pulmonary diseases can make some geometric variations on lung shape, analysis of the lung shape plays an effective role in determining the type of the lung disease. The goal of this study is to find the optimal feature vectors in order to TB detection in CXR images. In this study, using contour-based shape descriptors and Two Dimensional Principal Component Analysis (2DPCA), we proposed a novel approach to feature selection of the lung shape in order to TB detection. In our knowledge, texture based features are used in all of prior studies on TB detection instead of the contour based features. Using K-Nearest Neighbor classifier on one of the publicly dataset (namely the Montgomery dataset), we achieved the maximum accuracy (ACC) and AUC of 92.86% and 91.67% respectively.