Detection of Drug Resistant Tuberculosis in Chest Radiographs via Mediastinum Analysis

Diagnosis of Tuberculosis (TB) using Chest Radiographs (CXRs) is non-trivial, and identifying drug-sensitive and resistant TB is even more challenging. Analysis of mediastinum in CXRs could provide significant information for detecting TB conditions. The aim of this study is to develop a computerized framework to detect Extensively Drug Resistant (XDR), Multi Drug Resistant (MDR), and Drug Sensitive (DS) TB in CXRs using analysis of mediastinum. The images are acquired from a publicly available database and are subjected to edge-aware contrast enhancement to strengthen local edge contrast. The image projection profiling technique is employed to identify the mediastinal and thoracic borders. Measures that quantify the structural variations are extracted from the profile. Further, ratiometric indices are explored to assess the anatomical differences. Machine learning techniques, including support vector machine, linear discriminant analysis, and multilayer perceptron, are employed for binary classification of XDR, MDR and DS TB. The results indicate that the projection profiling technique can detect mediastinum and lung borders. Features extracted from the profile are found to characterize alterations in TB conditions. The mediastinal width ratio provided a mean difference of 43.9%, 31.6% and 15.4% between XDR vs DS, XDR vs MDR, and MDR vs DS, respectively. In addition, the mediastinal thoracic ratio achieved a mean difference of 35.3%, 0.1% and 43.8%, respectively. Moreover, the features between XDR and DS are statistically significant. Furthermore, accuracy of 79.5%, 77.5%, and 72.0%, and area under the curve of 84.6%, 81.0%, and 78.6% are obtained to differentiate XDR vs DS, MDR vs DS, and XDR vs MDR, respectively. This shows better differentiation between XDR and DS, with the least performance for XDR vs MDR, which aligns with the literature. Thus, it is evident that the proposed methodology could be utilized to identify drug-resistant TB conditions.