Artificial intelligence-reported chest X-ray findings of culture-confirmed pulmonary tuberculosis in people with and without diabetes

Abstract

Objectives: We applied computer-aided detection (CAD) software for chest X-ray (CXR) analysis to determine if diabetes affects the radiographic presentation of tuberculosis. Methods: From March 2017-July 2018, we consecutively enrolled adults being evaluated for pulmonary tuberculosis in Karachi, Pakistan. Participants had same-day CXR, two sputum mycobacterial cultures, and random blood glucose measurement. We identified diabetes through self-report or glucose >11.1mMol/L. We included participants with culture-confirmed tuberculosis for this analysis. We used linear regression to estimate associations between CAD-reported tuberculosis abnormality score (range 0.00 to 1.00) and diabetes, adjusting for age, body mass index, sputum smear-status, and prior tuberculosis. We also compared radiographic abnormalities between participants with and without diabetes. Results: 63/272 (23%) of included participants had diabetes. After adjustment, diabetes was associated with higher CAD tuberculosis abnormality scores (p < 0.001). Diabetes was not associated with frequency of CAD-reported radiographic abnormalities apart from cavitary disease; participants with diabetes were more likely to have cavitary disease (74.6% vs 61.2% p = 0.07), particularly non-upper zone cavitary disease (17% vs 7.8%, p = 0.09). Conclusions: CAD analysis of CXR suggests diabetes is associated with more extensive radiographic abnormalities and with greater likelihood of cavities outside upper lung zones.

Keywords

• Diabetes
• Tuberculosis
• Chest X-ray
• Deep learning