Computerized Reading of Chest Radiographs in The Gambia National Tuberculosis Prevalence Survey: Retrospective Comparison with Human Experts
P. Maduskar, I. Adetifa, J. van den Hombergh, E. Leroy-Terquem, A. Fasan-Odunsi, C.I. Sánchez, U. d'Alessandro and B. van Ginneken
in: Union World Conference on Lung Health, 2015
Rationale: Tuberculosis (TB) prevalence surveys require recruitment of trained personnel for reading chest radiographs (CXRs), which is scarce in high TB burden countries. Computerized CXR reading could replace human readers in such a scenario and we retrospectively investigate this possibility on data from the 2011-2013 TB prevalence survey in The Gambia. Methods: Computerized readings were compared with field and central readings on 4,552 CXRs. The survey participants were screened based on symptoms and CXR findings. Field readers judged CXRs on radiological findings as: Normal, Abnormal, suggestive of TB? or Other abnormalities; the latter two were considered abnormal. In case of symptoms and/or abnormal CXR, sputum samples were collected and bacteriological tests (fluorescence microscopy and BACTEC MGIT culture) were performed for confirmatory TB diagnosis. Following the analysis, 73 subjects with proven TB were considered Abnormal and the remainder as Normal. The CXRs were centrally audited by an expert reader and one of the following categories was assigned to each CXR: Normal, Active TB, Abnormal, healed TB, or Other pathology. The software (CAD4TB, Radboud University Medical Center, Nijmegen, The Netherlands) computed a TB score between 0-100 (0-Normal, 100-Abnormal) for each CXR. The area under receiver operating characteristic curve (AUC) was calculated for CAD4TB and various cut-off points were chosen on the TB score to compare specificities at the sensitivities of the field and central readings. Results: The field reading achieved a sensitivity of 86.3% at 72.8% specificity. The central reading had a sensitivity of 67.1% at 93.3% specificity, and when the healed TB category was considered abnormal, sensitivity increased to 87.7% with a decrease in specificity to 65.2%. CAD4TB attained an AUC of 0.90 and at all three sensitivity levels of human readers, specificity was nearly identical and not significantly different (p-value>0.05): at field sensitivity, CAD4TB had a specificity of 72.2%, and for central sensitivity levels, specificities of 93.7% and 66.0% (healed TB as abnormal) were obtained. Conclusions: When selecting subjects in a prevalence survey on the basis of chest radiography to undergo confirmatory testing for TB, the performance of computerized reading is not significantly different from field and central readings by human experts. The software has potential to improve the efficiency of TB prevalence surveys.