Quantification of pulmonary tuberculosis characteristics from digital chest x-rays using radiomics

Abstract

Pulmonary tuberculosis (PTB) is internationally one of the leading causes of death from a single infectious agent, and South Africa remains in the top 8 countries globally with the highest number of new infections. A chest x-ray (CXR) is still the most common radiological imaging procedure for PTB screening, diagnosis and monitoring, but it cannot be used as a standalone diagnostic tool due to the subjectivity associated with reporting. This can be addressed by quantifying digital CXR with tools such as radiomic feature extraction. In this thesis a unique sliding window segmentation method was developed to eliminate the difficult and time-consuming task of accurate PTB disease segmentation from planar images. It was applied as a secondary segmentation, superimposed on a primary automatic lung segmentation, that divided the entire lung region into uniform windows that overlapped while sliding over the CXR in both image dimensions. When radiomic features were extracted from each sliding window, it allowed the distribution of the features across the lung region to be evaluated. Three different outcomes were achieved when radiomic feature extraction was applied to chest x-rays using the sliding window segmentation. Firstly a model was developed that can automatically differentiate normal CXR from CXR with PTB cavities, which could improve the accuracy of CXR reporting currently regaining prominence as a high-volume screening tool. Secondly, signature parameter maps that showed a strong correlation to the lung pathology were constructed. This might be valuable as a quantitative supplementary indicator in the management of PTB disease and further increase the acceptance of CXR as a tool for assessing the TB response in medical research and clinical practice. Finally, a radiomics score was constructed that was able to quantify the change in the disease characteristics as seen from digital CXR of patients diagnosed with PTB. This radiomic score analysis of serial x-rays taken while patients receive TB therapy has the potential to be a quantitative monitoring tool of response to therapy. Radiomics was therefore successfully applied in this study to quantify the characteristics of PTB from chest x-rays.