Why is finding tuberculosis still hard?
Tuberculosis is an infection caused by a bacterium that mainly affects the lungs. It spreads through the air, so finding it quickly matters. In countries where it is common, screening centers see many people, and there are not always enough radiologists to read the chest X-rays.
The reference test in this study is called Xpert: it looks for the bacterium's DNA in a sample of phlegm. Testing everyone this way takes a lot of resources. So the idea is to triage: a chest X-ray first, then an Xpert test only for people whose X-ray looks abnormal. Several commercial AIs can read these X-rays, but independent comparisons between them were lacking.
What did the researchers do?
Between 2014 and 2016, people aged 15 and over attending three screening centers in Dhaka, Bangladesh, were enrolled one after another. Each person answered questions about symptoms, then had a chest X-ray and an Xpert test.
Every X-ray was read by radiologists and by five AI programs: CAD4TB, InferRead DR, Lunit INSIGHT CXR, JF CXR-1 and qXR. Importantly, none of these images had ever been used to train any of the AIs. The researchers compared the programs with each other, with the radiologists, and with the World Health Organization's targets for a triage test: finding at least 90% of cases, with at least 70% specificity, meaning the ability to avoid false alarms.
What were the results?
X-rays from 23,954 people were analyzed. All five AIs did significantly better than the radiologists. Their overall score, the area under the curve (a mark that would be 100% for a perfect test), ranged from 84.89% to 90.81%. Two programs, qXR and CAD4TB, met the WHO targets while finding 90% of cases.
Used for triage, all five AIs halved the number of Xpert tests needed, while still finding more than 90% of cases. However, all of them performed worse in people over 60 and in people who had had tuberculosis before.
Key figures
- 23,954people whose chest X-ray was analyzed
- 5 out of 5AI programs did better than the radiologists
- 50%fewer Xpert tests, while still finding more than 90% of cases
After a chest X-ray, with Mima
This study mainly concerns countries where tuberculosis is common, where AI helps decide who needs a lab test. The diagnosis itself still relies on that test and on a doctor's judgment.
If you have a lasting cough or lung tests to follow up, Mima helps you note your symptoms day by day, keep your results and prepare your questions, such as: "Was my X-ray read with the help of AI?"
What could this mean for patients?
In regions where tuberculosis is common, AI can read X-rays where radiologists are scarce. Lab tests can then be kept for the people who need them most, so the same resources can reach more people. The authors conclude that these AIs can be highly accurate and useful triage tools in such regions.
The study also offers a method to help health teams choose a program and set its alert threshold, depending on how many tests they can carry out.
What are the limits of this study?
The data come from one country and from screening centers in one large city. Results could differ elsewhere, with other populations or other X-ray machines. The software also changes quickly, so the versions tested here may not be the ones in use today.
The study looks at how X-rays were read, not at what happened to patients afterwards. It does not directly show how many people were treated sooner thanks to AI. The weaker results in people over 60 and in those who had had tuberculosis before also call for caution.
Frequently asked questions
Is a chest X-ray enough to diagnose tuberculosis?
No. In this study the X-ray is used for triage. The diagnosis relies on a lab test, such as Xpert, that looks for the bacterium.
Do these AIs replace radiologists?
The study compares them with radiologists for triage. It shows they can help where radiologists are scarce, not that they should replace them.
Why does the AI work less well for some people?
All five programs performed worse in people over 60 and in people who had had tuberculosis before. Health teams need to take this into account.
Where did the study take place?
In three screening centers in Dhaka, Bangladesh, where tuberculosis is common, among people aged 15 and over.
The original study
Qin ZZ, Ahmed S, Sarker MS, et al. “Tuberculosis detection from chest x-rays for triaging in a high tuberculosis-burden setting: an evaluation of five artificial intelligence algorithms” The Lancet Digital Health, September 2021. DOI : 10.1016/S2589-7500(21)00116-3
Read the full study in The Lancet Digital HealthThis article summarizes a scientific study and is not medical advice.