Generative AI Shows Promise in TB Detection, But Human Oversight Remains Essential

Generative artificial intelligence (AI) continues to make waves in radiology, offering potential benefits such as streamlined reporting, improved communication of findings, and expanded access in low-resource settings. But a new study urges caution, noting that current models still fall short of independent clinical use.

Researchers recently tested a generative AI model’s ability to detect tuberculosis (TB)-associated abnormalities on chest radiographs—images that are crucial for screening in high-risk and underserved regions. In areas with limited medical expertise, such technology could be transformative.

“Chest radiographs play a crucial role in tuberculosis screening in high-prevalence regions, although widespread radiographic screening requires expertise that may be unavailable in settings with limited medical resources,” wrote lead author Eun Kyoung Hong, MD, PhD, of Mass General Brigham in Boston, and colleagues.

The study used two public TB screening datasets to train the AI model, which then generated free-text reports identifying TB-related abnormalities and their locations. Radiologists independently reviewed the images and decided whether they would accept the AI-generated reports. A separate pair of radiologists established the ground truth.

Out of 800 radiographs, just over half showed TB-related abnormalities. The AI achieved a sensitivity of 95.2%, specificity of 86.7%, and overall accuracy of 90.8%. Although the model performed admirably, the radiologists’ results were slightly better, especially in pinpointing the exact location of abnormalities.

When evaluating the AI’s reports, one radiologist said they would accept 91.7% of the reports for normal images and 52.4% for abnormal ones. The second radiologist reported 83.2% and 37.0% acceptability, respectively.

Despite the strong numbers, the researchers emphasized the importance of caution. “While AI-generated reports may augment radiologists' diagnostic assessments, the current model requires human oversight given inferior standalone performance,” the authors concluded.

As interest in generative AI grows, particularly for its potential to support low-resource healthcare settings, this study highlights both the promise and the current limitations of these tools. With further refinement and close monitoring, generative AI could one day become a valuable aid in global TB screening efforts—but it's not ready to operate alone just yet.