AI tuberculosis screening uses software, known as computer-aided detection (CAD), to read chest X-rays and score how likely they are to show TB. Since 2021 WHO has recommended it for screening and triage in people aged 15 and over, in place of a human reader. It works best as a filter that decides who gets a confirmatory test.
Picture a screening van parked outside a market in a high-burden district. A portable X-ray unit runs off a generator. There is no radiologist within a hundred kilometres, and there never will be. A few seconds after each image is taken, software returns a number between 0 and 100. Above the threshold, the person gives a sputum sample for a molecular test. Below it, they go home.
This is not a pilot fantasy. It is how a growing share of TB case-finding now works, and it is the closest thing global health has to an unambiguous AI success. That makes it worth studying closely, because the reasons it worked are more specific than the hype suggests. The thesis here is simple: CAD for TB succeeded because it was given a narrow job inside a system that already had a confirmatory test, and its hardest problems were never about accuracy. Those are the lessons that transfer.
Why does TB need AI screening at all?
Tuberculosis remains the world’s leading infectious killer. WHO’s Global Tuberculosis Report 2025 estimates that 10.7 million people fell ill with TB in 2024 and 1.23 million died. Only 8.3 million people were newly diagnosed and reported, about 78% of the estimated total. The remaining gap, well over two million people a year, is the core problem. They are sick, often infectious, and not in any system.
Finding them means screening people who have not sought care, and many people with TB have few or no symptoms at the time. A chest X-ray is far more sensitive than asking about a cough. But X-rays need to be read, and radiologists are exactly the specialists that high-burden countries lack. For years this made mass X-ray screening impractical in precisely the places it was most needed.
CAD changed that equation. The software reads the image in seconds, never tires, and runs on a laptop in a van.
How accurate is AI at reading chest X-rays for TB?
The best-known independent evaluation, led by Zhi Zhen Qin and colleagues at the Stop TB Partnership and published in The Lancet Digital Health in 2021, tested five commercial products on chest X-rays from people attending three TB screening centres in Dhaka, Bangladesh. Every person also had a molecular test, which gave a reliable reference standard.
All five products outperformed the experienced local radiologists who read the same images. Areas under the curve ranged from roughly 0.85 to 0.91. The practical finding mattered more than the headline: according to the study, all five could cut the number of confirmatory Xpert tests needed by about half while keeping sensitivity above 90%.
That is the economic heart of it. Molecular tests are the expensive, cartridge-limited step. If software can safely rule out half the people before they need one, a programme can screen twice as many people for the same testing budget.
The same study also found something every programme manager needs to know. All five products performed worse in older people and in people who had previously had TB. Old TB leaves scars on the lung, and scars look abnormal. In a population with lots of past TB, the software flags many people who do not have active disease.
In March 2021, WHO’s consolidated guidelines on TB screening recommended that CAD may replace human readers for interpreting digital chest X-rays for screening and triage in people aged 15 and over. That recommendation was limited to adults and older adolescents because the evidence in children was thin, and it applies to products shown to perform at least as well as those WHO reviewed.
Why did TB become AI’s clearest win?
Four features came together, and most other AI use cases lack at least two of them.
The task was narrow. The software does not diagnose TB. It triages: this image looks abnormal enough to justify a test. Nobody is asked to trust the software’s final answer, because the final answer comes from a microbiological test.
There was a gold standard downstream. Because every flagged person gets a molecular test, errors in one direction are caught automatically. False positives cost a cartridge. They do not cost a wrong treatment.
The counterfactual was nothing. In a rural screening camp, the alternative to AI reading was not a radiologist. It was no reading, or a symptom questionnaire that misses many cases. The bar for being useful was low, and the software cleared it easily.
Independent evaluation existed early. The Stop TB Partnership and partners set up comparative evaluations of multiple commercial products on shared datasets, so programmes could compare vendors on the same terms instead of trusting marketing claims.
Compare that with, say, AI that reads mammograms in a well-staffed European screening programme. There the counterfactual is a skilled radiologist, the downstream workup is expensive, and errors in both directions carry real harms. The same technology faces a much harder test. We made a similar argument in our earlier piece: AI diagnostic evidence is strongest where the question is narrowest.
What goes wrong when programmes deploy TB CAD?
The hard problems turned out to be operational, not algorithmic.
Thresholds. The software gives a score, not a verdict. Someone has to choose the cut-off above which people get tested. Set it low, and sensitivity is high but you burn through cartridges. Set it high, and you save tests but miss cases. The right threshold depends on local prevalence, the age of the population, the history of past TB and the testing budget. WHO has published guidance and a calibration toolkit precisely because a threshold borrowed from another country can quietly give the wrong trade-off.
Software updates. This is the least discussed and most transferable lesson. A 2022 study in PLOS Digital Health, again led by Qin, compared two versions each of two leading products. The newer versions were more accurate. But the same threshold score meant different things in each version. With one product, a threshold of 60 on the older version gave 92% sensitivity and saved 43% of tests. On the newer version, the same threshold gave 88% sensitivity and saved 55% of tests. A programme that installed the update and kept its settings would have silently traded away cases for cartridges. The authors concluded that new versions will likely require thresholds to be recalibrated.
Hardware and people. CAD is only as useful as the X-ray unit it runs on, the generator that powers it, the radiographer who positions patients, and the courier who takes sputum to the lab. Ultra-portable X-ray machines have helped, but a programme that buys software without funding the rest of the chain has bought a very accurate idle laptop.
What comes after the test. Screening that finds more people with TB needs treatment slots, drug supply and follow-up to match. Finding a case is the start of an obligation, not the end of one, and it lands on programmes that, as we noted in our piece on how the biggest killers in emerging markets get the smallest share of aid, are rarely funded to match the burden.
A worked example: choosing a threshold
This is a hypothetical example with round numbers to show the trade-off, not data from a real programme.
A district programme plans to screen 20,000 adults this year and has budget for 3,000 molecular tests. Surveys suggest about 1% of the screened population has active TB, so roughly 200 cases are there to be found.
At a low threshold, the software flags 30% of people: 6,000 tests needed, double the budget, but it would catch almost everyone with TB. At a higher threshold, it flags 15%: 3,000 tests, within budget, with sensitivity falling to, say, 88%. That means about 176 cases found and 24 missed.
The programme picks the higher threshold because the budget forces it. Six months later the vendor pushes an update. Nobody recalibrates. The new version at the same threshold flags only 11% of people, so fewer tests are used and the monthly report looks efficient. But sensitivity has fallen further, and nobody notices until a prevalence survey years later shows more missed cases than expected.
Nothing in this story involves a bad algorithm. Every failure is a governance failure: a budget that set the clinical threshold, and an update process with no owner.
The strongest counterargument: is this win overstated?
A fair critic would say the TB CAD story is flattered by its comparator. Beating local radiologists in a screening centre is not the same as improving outcomes. The evidence base is mostly diagnostic accuracy, not trials showing fewer deaths or less transmission. Cost-effectiveness depends heavily on prevalence: in low-prevalence populations, screening thousands to find a handful of cases can be a poor use of scarce money, whether a human or software reads the film.
The critic would also point out that much of the evaluation has been done or funded by a small circle of global health organisations, and that performance in children, people with HIV and populations with lots of old TB remains less certain.
This is a fair challenge, and the honest answer is partial. The accuracy evidence is strong and independent enough to trust for triage. The outcome evidence is weaker, as it is for most screening. But the logic of the win does not depend on software being magical. It depends on software being good enough to make X-ray screening possible where it was otherwise impossible. That is a lower bar, and it has been cleared.
What this means for health ministries, donors and AI founders
For TB programmes and ministries, the priority is governance over procurement. Assign someone to own thresholds, document the version in use, recalibrate after every update and audit a sample of results against molecular tests. Treat CAD as a clinical instrument with a maintenance schedule, not an app.
For donors, fund the whole chain: X-ray units, power, radiographers, sample transport, cartridges and treatment. Funding the software alone moves the bottleneck to the next link. It is also worth funding the unglamorous independent evaluation infrastructure that made TB CAD trustworthy in the first place. Our piece on why TB declines are exactly when it comes back explains why case-finding has to be sustained, not campaign-based.
For AI founders working in other areas of global health, copy the structure, not the product. Find a narrow triage task, with a confirmatory step downstream, where the realistic alternative is no expert at all, and invite independent head-to-head evaluation early. And plan for version control from day one: if your update changes what your scores mean, your customers need to know before they install it.
Frequently asked questions
Does WHO recommend AI for TB screening?
Yes. Since March 2021, WHO’s consolidated guidelines on TB screening have recommended that computer-aided detection software may be used instead of human readers to interpret digital chest X-rays for TB screening and triage in people aged 15 and over. The recommendation covers products that perform at least as well as those WHO reviewed, and programmes are advised to calibrate thresholds for their own setting.
Can AI diagnose tuberculosis from a chest X-ray alone?
No. CAD software estimates whether an X-ray looks consistent with TB, which makes it a triage tool. A diagnosis still requires a confirmatory test, usually a rapid molecular test on sputum. The value of AI is in deciding who needs that test, allowing programmes to screen many more people without a radiologist and to spend expensive tests where they are most likely to find disease.
Is AI TB screening used for children?
WHO’s 2021 recommendation is limited to people aged 15 and over, because evidence on CAD performance in children was insufficient at the time. Children’s chest X-rays look different, childhood TB is harder to confirm microbiologically, and the products were mostly trained and evaluated on adult images. Research is ongoing, but programmes should not assume adult performance carries over to younger children.
What is a CAD threshold score?
CAD products give each X-ray an abnormality score, often on a scale of 0 to 100. The programme chooses a threshold above which people are referred for testing. A lower threshold catches more cases but requires more tests; a higher one saves tests but misses more. Because the best threshold depends on local conditions and can change between software versions, it should be set and periodically rechecked locally.
Which companies make AI TB screening software?
Products evaluated in the independent 2021 Lancet Digital Health study included CAD4TB, qXR, Lunit INSIGHT CXR, InferRead DR and JF CXR-1. Other products have entered the market since. Programmes should rely on independent, head-to-head evaluations on data similar to their own population, and check which version was evaluated, rather than choosing on vendor-reported accuracy alone.
The takeaway
TB screening shows what AI in global health looks like when it works: a narrow job, a real test behind it, and an alternative that was nothing at all. The next wins will come from programmes that copy that structure, and the next failures from programmes that install an update on a Friday and never check what their threshold now means.
Has your programme recalibrated its CAD threshold after a software update, and what did you find?
