Computer-Assisted Diagnosis of Tuberculosis: A First Order Statistical Approach to Chest Radiograph

Abstract

Textural properties of normal and tuberculosis posterior-anterior chest radiographs were looked into in this investigation. The proposed computerized scheme segmented the lung field of interest using a user-guided snake algorithm and extracted the corresponding pixel data. For both normal and tuberculosis radiographs, the grayscale intensity distribution within the region of interest was analyzed to study their respective characteristics, and fed to classifiers for automated classification. Statistically the tuberculosis infected radiographs manifested a higher variance, third moment, entropy and a lower mean value in their intensity distributions, compared to their normal peers. The greater disparities between a particular radiograph and the confidence interval determined by our normal groups on some of the features were observed to be related to the level of haziness at the upper lobe. Lastly, the C4.5 (a decision tree based classifier)-adaboost achieved an accuracy of 94.9% in normal-tuberculosis classification. An integrated index, called tuberculosis index (TI), is proposed based on texture features to discriminate normal and tuberculosis chest radiographs using just one index or number. We hope this TI can be used as an adjunct tool by the radiographers in their daily screening.