Skip to main content

Comparing the Performance of Man and Machine for TB Screening in Chest Radiographs

Friday, November 08, 2013 — Poster Session IV

2:00 p.m. – 4:00 p.m.

FAES Academic Center (Upper-Level Terrace)




  • A. Karargyris
  • L. Folio
  • J. Siegelman
  • F. Callaghan
  • S. Candemir
  • Z. Xue
  • P.-X. Lu
  • Y.-X. Wang
  • S. Antani
  • G. Thoma
  • S. Jaeger


To reduce the global burden of tuberculosis (TB), the development of new and more effective tools to detect TB is important. Due to the large number of world-wide TB infections, an automated TB detection approach is desirable. The higher throughput of automated systems, compared to manual screening, allows efficient testing of large populations for TB, with lower costs per tested individual. However, in order to become generally accepted, an automated screening approach needs to provide a performance that approaches the performance of human experts. Here, we describe our automatic image processing system for detecting TB in chest radiographs, including the techniques that we use to detect manifestations of TB in the lung. We evaluate our system on two x-ray sets from the local TB clinic of Montgomery County and from Shenzhen Hospital China. For these sets, we achieve an accuracy of about 80%, and an area under the ROC curve of about 88%. For the Montgomery County set, we compared the performance of our system with the performance of two expert radiologists. When trying not to miss any positive cases, the performance of the radiologists drops significantly, and is not significantly better than general system accuracy.

back to top