Computer-aided diagnosis of X-rays in a screening for pulmonary tuberculosis of a prison population in Tanzania

Background: Recent studies have shown that computer-aided diagnosis (CAD) of chest X-rays (CXR) allows the detection of pulmonary tuberculosis (TB) with a performance similar to that of clinical officers. This study investigated the performance of the new CAD4TB system in a Tanzanian prison. Design/Methods: Between August 2013 and June 2014, all inmates of the Ukonga Prison, Dar es Salaam, are screened for TB. X-rays are classified as "normal", "abnormal, not suggestive of TB", "abnormal, consistent with TB", and "abnormal, highly suggestive of active TB" by an assistant medical officer with two years of training in radiology (reader A) and processed by the software CAD4TB v3.07 (Diagnostic Image Analysis Group, Nijmegen University, The Netherlands). A subset of 517 consecutive images was re-evaluated by reader A, as well as by an independent clinical officer (reader B), and by a TB expert. The CXR findings of the TB expert, based on a case definition which included X-ray results either "consistent with TB" or "highly suggestive of active TB", determined the radiological reference that was used for CAD4TB and the two other readers. Results: On the 466 negative and 51 positive images, readers A and B performed with a sensitivity of 52.9% and 43.1%, and a specificity of 92.4% and 77.5%, respectively. CAD4TB performed significantly better than reader B, but not as good as reader A (AUC 0.751, e.g. sensitivity 52.9% and specificity 83.5% for a defined threshold). The intra-reader agreement of reader A was good (Cohen's i:31.80), but the agreement between A and B was low (13.36). Although all readers, including CAD4TB, analyzed every CXR in under two minutes, human readers delayed 10% of images by more than 20 hours during the screening. Conclusion: This first evaluation of CAD4TB in a real-world screening situation showed, in line with previous studies, that CAD4TB performs better than a clinical officer (B), but not as good as a more experienced X-ray reader (A). Moreover, the overall time to detection is kept predictably short, which is an important criterion in multistage screening algorithms.