Automated Lung Diseases Classification from Chest Radiographs using Machine Learning Techniques

Abstract

Nowadays, the most typical examination type in a radiology department is a chest newlineradiography imaging. The use of automatic illness classification can help radiologists newlineprovide better patient care and lessen their workload. Over the past ten years,convolutional newlineneural networks (CNNs) have achieved superhuman performance in many image newlineclassification, segmentation, and quantification tasks, leading to a paradigm shift in newlinemedical image processing. CNNs are being used on chest radiography images, but there newlineare still many research obstacles to overcome before they can be used in a therapeutic newlinesetting due to the high spatial resolution, the dearth of big datasets with trustworthy newlineground truth, and the wide range of disorders. Interestingly, the innovative contributions provided throughout this thesis are motivated by these problems. newlineAccording to the esteemed World Health Organization (WHO), the primary causes newlineof death in the world are Pneumonia (PNA), COVID-19 (C19), Tuberculosis (TB), newlineand Pneumothorax (PNTRX). These ailments are accompanied by symptoms such as newlinecoughing, sneezing, fever, and shortness of breath, which are commonly observed. In newlineorder to identify these symptoms, one must undergo various tests including molecular newlinetests (RT-PCR), antigen tests, Monteux tuberculin skin test (TST), and complete blood count (CBC) tests. However, these tests are known to be time-consuming and possess newlinean error rate of 20% along with a sensitivity of 80%. Hence, radiographic examinations newlinelike computed tomography (CCT) and X-rays (CXR) are employed to identify lung diseases newline(LDs) under the guidance of a medical professional. However, the concern lies in newlinethe possibility of these LDs overlapping in their diagnosis when observed through chest newlineradiograph. The automation of this process is required in order to accurately classify newlinethese diseases using healthy (HY) images. Currently, there exists no established method newlinefor identifying and categorizing these LDs. Consequently, we were motivated to utilize newlineeight pre-trained CNNs to classif