AI tool CAD4TB helps detect TB on chest X-rays in Indonesian hospitals

Tuberculosis remains a major health challenge worldwide, especially in countries with many cases and limited access to specialist medical imaging. In Indonesia, early detection is critical but can be hampered when experienced radiologists are not available. To explore whether an automated system could help, researcher Erlina Burhan and colleagues evaluated CAD4TB, a computer-aided detection system, for reading chest X-ray images of people suspected of having pulmonary tuberculosis in a hospital setting. The team used a retrospective, cross-sectional design and drew chest radiographs from more than 1,100 adult patients recorded in the national tuberculosis database. Every image was processed with CAD4TB and independently reviewed by two experienced radiologists. Bacteriological test results were used as the diagnostic reference standard against which CAD4TB and the radiologists were judged. The study focused on how well the automated tool performed compared with human readers under real hospital conditions, aiming to assess whether CAD4TB could be a practical aid in places where radiological expertise is limited.

The study tested CAD4TB at an index cutoff score of 60 and compared its output to the readings of two experienced radiologists, using bacteriological test results as the reference standard. At that cutoff, CAD4TB achieved a sensitivity of 81.04% and a specificity of 63.80%. By comparison, radiologist interpretations reached a sensitivity of 88.20% and a specificity of 58.18%. The researchers also looked at clinical subgroups. In patients without pleural effusion, CAD4TB sensitivity rose to 83.65% while radiologist sensitivity was 90.35%. Across other subgroups, including people with a prior tuberculosis history and those with HIV-negative status, CAD4TB showed consistent specificity advantages. These results came from processing over 1,100 chest radiographs drawn from the national tuberculosis database in a hospital setting in Indonesia, demonstrating how CAD4TB and human readers performed against bacteriological confirmation in a real-world sample.

The findings suggest practical implications for hospitals and tuberculosis programs in high-burden areas. CAD4TB did not outperform experienced radiologists overall, who remained slightly more sensitive, but the automated system showed strengths such as comparable detection in many cases and higher specificity in several subgroups. Its consistent performance in patients without pleural effusion and in groups defined by prior tuberculosis history and HIV-negative status indicates it can handle a range of patient complexities. Because CAD4TB can analyze X-rays quickly and is easy to use, the authors argue it could assist radiologists and clinical teams, speed up triage, and provide a reliable second opinion in settings where expert readers are scarce. The study supports integrating CAD4TB into diagnostic workflows as a tool to help identify likely tuberculosis cases for confirmatory testing and timely treatment, while recognizing that expert radiological interpretation still adds value.