Who Should Make Sure AI Reaches Everyone?

Every few months, another imaging AI tool earns regulatory clearance, promising faster reads, fewer misses, and relief for overworked radiologists. Yet for many SIIM members practicing in lower-resource settings, the question is not which algorithm to buy. It’s whether the infrastructure exists to run one at all. If AI is going to transform imaging, we have to ask an uncomfortable question: transform it for whom? 

The gap is bigger than the algorithm

The Lancet Oncology Commission on medical imaging found that access to basic imaging is profoundly unequal: low-income countries have a small fraction of the scanners and radiologists per capita that high-income countries have. In some countries, a handful of radiologists serve millions of people, and many reside in capital cities, far from the rural populations who need them most. In these places, AI could matter most, by triaging chest X-rays, flagging urgent findings, and extending scarce expertise to clinics that have never had a radiologist on site. There is real evidence that this works. The WHO now recommends computer-aided detection for tuberculosis screening on chest radiographs, which makes it one of the clearest examples of AI changing practice where specialists are scarce. Mobile screening vans with digital X-ray units and on-board AI can now reach communities that once waited weeks for a read. 

But an algorithm is the last mile of a long road. It needs digitized images, a functioning PACS, reliable power and connectivity, interoperable standards like DICOM and HL7 FHIR, cybersecurity, and someone trained to act on its output. Many facilities still print film or store images on standalone workstations. Without that foundation, even the best model sits unused. This is where imaging informatics, rather than AI research alone, becomes the deciding factor. 

Models trained elsewhere don’t always travel well

Most imaging AI is trained on data from a narrow set of high-income institutions. When these models meet different patient populations, disease prevalence, and equipment (older scanners, portable units, unfamiliar acquisition protocols), performance can degrade quietly. A model tuned for a low-TB-prevalence population may behave very differently where TB, HIV coinfection, and post-infectious lung changes are common. Research has also shown that AI can systematically underdiagnose patients from underserved groups, even within high-income health systems. 

Deploying a model that wasn’t validated locally isn’t charity. It’s a risk. Local validation, ongoing performance monitoring, and clear processes for reporting failures are just as essential in Lagos or Lahore as they are in Boston. 

The economics are not neutral

Market forces naturally flow toward health systems that can pay. Per-study licensing fees, cloud subscription costs, and hardware requirements built for well-funded hospitals can put tools out of reach. Meanwhile, the data from lower-resource settings, which is often exactly what’s needed to make models robust, is sometimes collected without lasting benefit to the institutions that provided it. If we don’t intentionally design for equity, the default outcome will be a widening gap. 

So, who is responsible? The honest answer is everyone, but in specific ways. 

• Vendors should build for constrained environments: offline-capable tools, lightweight hardware requirements, support for older modalities, and tiered pricing that doesn’t assume high-income reimbursement structures. 

• AI developers are beginning to face this question too. Some frontier labs have started setting aside equity for public benefit. Anthropic, for example, announced in October 2026 that it would commit a share of its stock to a fund for nonprofits and affected communities, though governance details are still to come. Whether commitments like these ever reach health systems in lower-resource countries may depend on communities like ours making the case. 

• Governments and global health bodies must invest in digital infrastructure and regulatory capacity, as WHO’s guidance on AI ethics urges, so that local authorities can evaluate tools rather than simply import them. Funding agencies can tie support to sustainability, so that projects don’t disappear when a grant ends. 

• Academic centers should treat data partnerships as two-way streets, co-developing models with institutions in low- and middle-income countries, sharing authorship and ownership, and building local research capacity rather than simply extracting data. 

• Hospitals and local leaders in lower-resource settings deserve a seat at the table when tools are designed, and they should push for transparency about how models were trained and tested. 

• Professional societies like SIIM are uniquely positioned to help. We bring together the informaticists, radiologists, engineers, and vendors who make systems work together. We can share expertise in open standards, publish practical guidance on deploying AI with limited infrastructure, mentor imaging informatics professionals worldwide, support local validation, and broaden participation through virtual programs and accessible membership. Above all, we can ensure members from lower-resource countries help shape the conversation rather than remain on the sidelines. 

A call to the community

Equity in imaging AI won’t happen by default. It will happen because people who understand both the technology and the plumbing decide to make it happen. If you work in a lower-resource setting, your experience of what breaks, what’s missing, and what actually works is exactly the knowledge this field needs. Share it. If you work in a well-resourced institution, look for partnerships that build capacity rather than dependency. 

The question isn’t whether AI will reshape imaging. It’s whether we’ll let it widen the gap or help close it. 

Read More: 

1. Hricak, H., Abdel-Wahab, M., Atun, R., Lette, M. M., Paez, D., Brink, J. A., … & Scott, A. M. (2021). Medical imaging and nuclear medicine: a Lancet Oncology Commission. The Lancet Oncology, 22(4), e136-e172. 

1. Mollura, D. J., Culp, M. P., Pollack, E., Battino, G., Scheel, J. R., Mango, V. L., … & Dako, F. (2020). Artificial intelligence in low-and middle-income countries: innovating global health radiology. Radiology, 297(3), 513-520. 

1. Guidance, W. H. O. (2021). Ethics and governance of artificial intelligence for health. World Health Organization, 1. 

1. Seyyed-Kalantari, L., Zhang, H., McDermott, M. B., Chen, I. Y., & Ghassemi, M. (2021). Underdiagnosis bias of artificial intelligence algorithms applied to chest radiographs in under-served patient populations. Nature medicine, 27(12), 2176-2182.  

1. World Health Organization. (2021). WHO consolidated guidelines on tuberculosis. Module 2: screening-systematic screening for tuberculosis disease. World Health Organization. 

1. Anthropic. Investing in the public benefit. 18 October 2026. https://www.anthropic.com/news/public-benefit-fund