CAD systems in medical application: Detection of tuberculosis from CXR-images using convolutional neural based networks (case study: Nigerian public health)

Tuberculosis (TB) is a very dangerous contagious bacterial infection that usually attacks the lungs and has spread widely around the globe in the last 4 to 5 decades. Early detection of TB is pivotal to the decrease in morbidity and mortality. The diagnosis of TB is made through conventional methods such as blood tests, skin tests, CT scans, and sputum tests, which are tedious and take weeks or even more. Therefore, to lower the detection time and raise the accuracy of diagnosis, due to the time-consuming task that typically requires expert radiologists to read the test, leading to fatigue-based diagnostic error and lack of trained experts in areas of the world where radiologists are not available or limited. In recent years, the application of Artificial intelligence (AI) and the breakthrough of deep learning (DL) approaches like convolutional neural networks (CNN) architectures in object detection and classification have attracted many researchers to apply this sophisticated technology in different fields of science, primarily in the medical field to achieve expert-level performance in the interpretation of different diseases, which is powered by the availability of labeled digital dataset. Hence chest radiograph (CXR) has become critical in detecting thoracic diseases such as TB. This study used a public and private dataset of CXR images to build models that can effectively detect TB using pretrained deep neural networks. In our experiment, the best performing model was MobileNetV1 out of the models used, in which we were able to achieve an accuracy of 94%, specificity of 96%, sensitivity of 92%, with an f1-score, precision of 94% respectively on a private dataset we collected from a densely affected region of Africa, Nigeria to be precise. Meanwhile, the same model on a public dataset scored an accuracy of 100%, sensitivity of 99%, specificity, precision and F1-score of 100% respectively. Hence, the performance of the deep neural networks (DNN) model in our experiment shows the efficacy of DNN in detecting tuberculosis from CXR images.