Detection of Chest X-ray abnormalities and tuberculosis using computer-aided detection vs interpretation by radiologists and a clinical officer

🔗2014

🔗Journal/Publication: Proceeding from Union World Conference on Lung Health

🔗Read it in full version: https://www.diagnijmegen.nl/publications/khan14/

Abstract

Background: There is renewed interest in chest radiography (CXR) as a screening tool for tuberculosis in high burden countries. Limited availability of skilled personnel in low-resource settings is a major barrier to its wider-scale adoption. Computer-aided detection (CAD) has been shown to be effective in supporting TB diagnostic algorithms in African settings; however, limited real-world applications of such systems have been evaluated. This study compares automated readings of a CAD system for CXR abnormalities with the interpretations of radiologists and a clinical officer.

Design/Methods: Suspects for lung diseases were identified through active screening in the private sector at Family-Practitioner clinics. These suspects were referred to the lung health diagnosis and screening center for CXR from three low-income towns of Karachi,Pakistan. CXRs were analyzed using CAD4TB v3.07(Diagnostic Image Analysis Group,Nijmegen,The Netherlands), a system designed for TB diagnosis that computes an abnormality score (0-100) on the basis of shape, symmetry and texture of the lung fields. CXRs were reviewed by two radiologists and a clinical officer who were masked to the CAD4TB score and patient disease status. Interobserver Agreement (Cohen’s Kappa) between CAD4TB and each radiologist and clinical officer for CXR abnormality was analyzed at various cut-points of CAD4TB.

Results: CXRs from 186 suspects were included in the study for analysis. A CAD4TB score of 80 maximized agreement between CAD4TB and each reviewer (Radiologist 1: Kappa 0.62, Radiologist 2: 0.54, Clinical Officer: 0.60). Mean sensitivity and specificity for CXR abnormality of CAD4TB compared to the reviewers was 78% and 79% respectively. A high proportion of non-TB CXR abnormalities as reported by the reviewers were also classified as abnormal by CAD4TB at a cut-point of 80 (Radiologist 1:88.2%, Radiologist 2: 71.9%, Clinical Officer 60%). Agreement between CAD4TB and each reviewer subsequently increased after including only TB CXR abnormalities (Radiologist 1: Kappa 0.63, RadiolA,A!ogist 2: 0.67, Clinical Officer: 0.71).

Conclusion: CAD4TB has the potential as a screening tool for TB in high burden countries. Detection of non-TB abnormalities by CAD4TB may lead to challenges in interpretation of results in programmatic settings. Use of additional data relevant to local disease epidemiology such as patient demographics as well as symptom screening may improve accuracy of CAD4TB when utilized in TB control programs.