A study published in The Lancet Digital Health takes a close look at how well AI-powered computer-aided detection (CAD) tools can help identify tuberculosis (TB) in some of the world’s most at-risk populations. With South Africa, a country that faces high rates of both TB and HIV, as the setting, researchers compared 12 different CAD products to see how well they could spot TB in chest X-rays. While AI offers promise in quickly identifying cases, the study shows that these tools aren’t without limitations—especially when it comes to vulnerable groups.
Tuberculosis is still a leading cause of death globally, particularly in areas where HIV rates are high. People living with HIV are more likely to develop TB, and catching the disease early can save lives. But in many high-risk areas, there simply aren’t enough medical staff or resources to test everyone. This is where CAD systems, which use AI to scan X-rays and flag potential cases, could make a real difference by speeding up the screening process.
In the study, 774 participants from a South African TB survey had their X-rays analyzed by 12 different CAD systems. The participants were also tested using lab methods, so researchers had a clear way to check the AI’s accuracy. Of the group, 258 were confirmed to have TB, while 516 tested negative.
The main question: How well could each CAD tool identify those with TB? The researchers measured this using something called the area under the receiver operating characteristic curve, which basically tells you how good a test is at telling the difference between positive and negative cases. A score closer to 1 means the system is more accurate.
The results showed a wide range of performance:
- Lunit and Nexus were the top performers, with AUC scores near 0.9, meaning they did a great job of detecting TB.
- qXR, JF CXR-2, InferRead, Xvision, and ChestEye followed closely with scores between 0.8 and 0.9, still performing well.
- XrayAME, RADIFY, and TiSepX-TB didn’t do as well, with scores under 0.8, indicating lower accuracy.
Some tools, like Lunit and Nexus, were able to maintain high sensitivity—catching more than 90% of TB cases—across a wide range of thresholds. This is important because it means fewer people would need additional testing, which is a big deal in areas where healthcare resources are limited.
One of the more concerning findings was that the CAD tools didn’t perform as well in certain high-risk groups. Older people, those with previous TB, and individuals living with HIV saw lower detection accuracy across most tools. This suggests that the technology struggles with cases where the person’s lungs might already be damaged or where TB symptoms might be less obvious. It’s a reminder that while AI can be helpful, it’s not a magic solution for everyone.
Another key point from the study was the importance of setting the right threshold for detection. Different CAD systems—and even different versions of the same system—varied in when they flagged a chest X-ray as suspicious. Setting the threshold too low could result in unnecessary testing, while setting it too high might mean missing cases. This means that healthcare providers need to be careful about how these tools are used and where they set the bar for TB detection.
This study highlights both the potential and the challenges of using AI to combat TB. On one hand, these tools can help overwhelmed healthcare systems in high-risk areas by quickly identifying who needs further testing. On the other hand, there’s a clear need for ongoing evaluation to ensure these tools work across different populations and settings.
As more CAD systems come onto the market, the researchers call for a global strategy to validate and standardize these tools. Without that, there’s a risk that some versions won’t perform well in certain areas, leaving gaps in TB detection. In short, AI can help, but it’s not a one-size-fits-all solution, and careful testing is needed to ensure it works where it’s needed most.