Computer-Aided Detection for Tuberculosis and Silicosis in Chest Radiographs of Southern African Gold Miners

Abstract

Background: The South African gold mining sector has been for over one hundred years a considerable source of tuberculosis and silicosis disease burden across southern Africa. Reading chest radiographs (CXRs) is an expert and time intensive process necessary for the screening and diagnosis of lung disease and the provision of evidence for compensation claims. Our study explores the use of computer-aided detection (CAD) of both tuberculosis (TB) and silicosis in a population with a high incidence of both diseases. Methods: A set of 330 CXRs with human expert-determined classifications of silicosis, tuberculosis, silcotuberculosis, and normal were provided to each of four health technology companies. Ability of each of their respective CAD systems to provide accurate quantitative predictions for each CXR was assessed using receiver operating characteristic (ROC) curve analysis of the under the curve (AUC) metric. Findings Three of the four systems differentiated accurately between TB and normal images (AUCs 0·989 [95% CI 0·975-1·0], 0·963 [0·930-0·996], and 0·980 [0·961-0·998]), while two differentiated accurately between silicosis and normal images (AUCs 0·986 [0·960-1·0], 0·939 [0·901-0·978]). Inclusion of silicotuberculosis images reduced each system's ability to detect either disease from the set of CXRs. In differentiating between any abnormal CXR from normal CXRs, the most accurate system achieved a sensitivity and specificity both of 98·2%. Interpretation: The use of CAD as a tool in mass screening for TB and silicosis in a population with a high burden of both diseases shows considerable promise, but current ability of CAD to differentiate between the two diseases is limited. Retraining of the systems with more silicosis CXRs and adding information about past work exposure into the screening tool may improve accuracy. Funding Statement: Canadian Institutes of Health Research, Canada Research Council, and the Department of International Development, United Kingdom. Declaration of Interests: The authors declare no competing interests. Ethics Approval Statement: Ethics approval was provided by the University of British Columbia Behavioural Research Ethics Board (H18-01793) and the University of Cape Town Faculty of Health Sciences Human Research Ethics Committee (HREC REF: 563/2019).