Why computer-aided detection still misses the patients who need it most

Computer-aided detection for chest imaging is no longer experimental. It's recommended, deployed, and in some settings already standard. And yet the people with the most to gain from it — those in high-burden, under-resourced settings, far from a radiologist — are the ones it serves least well. This is not a failure of ambition. It's a structural consequence of two gaps that compound each other: who the models were trained to read, and who can actually reach them. This piece is about both.

A tool that works, unevenly

Start with what's genuinely true: CAD for chest radiography works. In 2021 the World Health Organization went so far as to recommend computer-aided detection software as an aid in tuberculosis screening — a rare institutional endorsement of AI in frontline diagnosis, made in part because the need is so acute. Of an estimated 10 million TB cases in 2019, only about 7.1 million were diagnosed and reported; roughly 29% went undiagnosed or unreported entirely. That gap — the missing millions — is exactly the space a scalable reader is supposed to fill, in precisely the places where radiologists are scarcest.

So the tool exists, the endorsement exists, and the need is overwhelming. The problem is what happens when you look at whom the tool reads well.

The first gap: who the model was trained to see

Modern chest-imaging AI is trained overwhelmingly on a handful of large public datasets — MIMIC-CXR, CheXpert, ChestX-ray14 — all drawn from a small number of well-resourced hospital systems in the United States. A model learns the population it's shown. When that population is narrow, the model's competence is narrow with it, in ways that don't announce themselves.

The clearest evidence came in 2021, when Seyyed-Kalantari and colleagues, writing in Nature Medicine, examined chest-X-ray classifiers trained on those very datasets and measured how often the models wrongly declared a sick patient healthy. The bias was systematic. Female patients, Black patients, Hispanic patients, younger patients, and patients on Medicaid were all underdiagnosed at higher rates — and the effect compounded at intersections, so that, for instance, Hispanic female patients fared worse than white female patients. Underdiagnosis is the most dangerous kind of error a screening tool can make: it sends a patient home reassured, when they needed care.

The point is not that anyone built these models to fail certain groups. It's that a model trained on one population silently encodes that population as the definition of "normal," and every group underrepresented in the data inherits a quieter, less accurate reader.

The second gap: the model doesn't transfer

You might hope that a model good enough in one place is good enough everywhere, once deployed. The evidence says otherwise — and the most instructive case is TB CAD in exactly the high-burden settings the WHO recommendation was written for.

In 2024, a study in The Lancet Digital Health externally validated twelve commercial CAD products against a South African national TB prevalence survey. Two findings matter here. First, the manufacturers' own default thresholds — the settings the software ships with — did not hold. One widely used product, run at its default threshold, achieved only 64% sensitivity on this population; reaching the target sensitivity required a completely different threshold than the one its documentation implied. Second, performance sagged in exactly the subgroups a high-burden program cannot afford to miss: every product performed worse in people living with HIV, and all performed worst in older patients.

The lesson is uncomfortable. A CAD tool is not a fixed instrument you can ship and trust; its accuracy is contingent on the population it meets, and the settings that make it safe in one place can make it unsafe in another. The authors' own conclusion was blunt — programs should be wary of extrapolating thresholds from the literature, and on-site calibration is essential. A tool that has to be re-tuned, per site, by people with the expertise to do so, is a tool whose burden falls hardest on the places with the least capacity to carry it.

Why the two gaps compound

Taken separately, each gap is a known problem with known mitigations. Taken together, they reinforce each other in a way that specifically disadvantages the patients CAD was meant to help most.

The populations least represented in the training data are, very often, the same populations with the least access to the infrastructure that modern CAD assumes — the connectivity, the compute, the technical staff to validate and recalibrate. So the underrepresentation makes the model less accurate for these patients, and the access gap makes it harder to catch and correct that inaccuracy locally. A better-resourced hospital can commission its own validation study and tune its thresholds; a rural clinic screening for TB cannot. The gap in the data and the gap in access are not two separate problems. They are one problem, seen from two sides.

What closing it actually requires

None of this is an argument against computer-aided detection. It's an argument about what has to be true for it to serve the people it was endorsed to serve. Three things follow.

Representative data isn't a nicety; it's the fix for the first gap. A model's competence is bounded by whom it has seen. Closing the underdiagnosis gap means training and validating on the populations who will actually be read — not treating a handful of US hospital datasets as a universal proxy for the human chest.

Portability has to be designed in, not assumed. If accuracy is contingent on population and threshold, then a responsible tool is one that can be validated and calibrated where it's deployed — not one that demands a data-center round-trip and a specialist to do it.

And where it runs decides who it reaches. A tool that needs connectivity and central compute is a tool that reaches the well-connected first, which is to say the already-served first. Access is not a deployment detail bolted on at the end; it's upstream of whether the second gap ever closes.

This is the frame we build GeoRadX inside. It's why Rad is built to run on the device, in the clinic, on local data and without a connection — so that the population being read and the population the model was shaped for can be the same population, and so that being far from a radiologist stops meaning being last in line for a second read. We don't claim to have closed either gap. We're saying, plainly, that any tool that ignores them will keep missing the people who need it most — and that the two gaps have to be treated as the single problem they are.