Reading the Mammogram: AI Imaging and the Caribbean Detection Gap

Note from the President

October is Breast Cancer Awareness Month. In the Caribbean, breast cancer is the most diagnosed cancer in women and the most common cause of cancer death among us. The reasons our outcomes lag are not mysterious. We are diagnosed later than women in higher income countries because we screen less, and we screen less because mammography capacity in our region is uneven, expensive, and concentrated in capital cities. The introduction of artificial intelligence into diagnostic imaging is one of the few developments in the last decade that could meaningfully change this picture.

I asked Dr. Yolande Pierre-Louis, who has spent her career evaluating diagnostic tools for Caribbean populations, to tell us what we should believe about AI mammography and what we should not. Her piece is data-heavy. That is on purpose. In this domain, hype kills.

If you have not had a mammogram in the last two years and you are forty or older, please book one before this newsletter goes out next month. Nothing else in this issue matters more than that.

Adrian Dunkley Founder and President, Caribbean AI Association

---

Feature

Reading the Mammogram: AI Imaging and the Caribbean Detection Gap

By Dr. Yolande Pierre-Louis

The promise of AI in mammography is, in plain terms, a second reader that never gets tired, never has a bad morning, and can be deployed anywhere there is a digital mammography unit and a working internet connection. The published evidence in 2024 supports the claim that AI second reading raises detection rates and reduces recall rates in screening populations. Three large studies, the MASAI trial published in Lancet Oncology in 2023, the PRAIM evaluation in Germany earlier this year, and the Capio retrospective from Sweden, all reported sensitivity gains in the range of seven to twenty percent against the unaided radiologist, with comparable specificity and reduced workload.

That is the headline. Now the small print, because the small print is where Caribbean policy lives.

First, none of those studies were performed on Caribbean women. The training and validation sets for the major commercial AI mammography systems, including those from Lunit, Therapixel, and Volpara, are drawn predominantly from European and East Asian populations. We have known for years that breast density, lesion morphology, and tumour subtype distributions differ in women of African and South Asian ancestry, and our region carries a meaningful share of both. There is no published evidence at the time of this writing that any of the leading systems perform at parity on Caribbean women. There is also no published evidence that they do not. The honest answer is that we do not know, and we should not deploy what we have not tested.

Second, the cost structure is unforgiving. A commercial AI second reader licence runs between four and seven US dollars per case in the markets where it is sold today. For a screening programme that processes thirty thousand cases per year, that is between one hundred and twenty thousand and two hundred and ten thousand US dollars in annual software cost alone, before the hardware upgrade required to deliver images in the format the system expects. That figure is achievable for Trinidad and Tobago and Barbados. It is challenging for Jamaica without external support. It is not credible for Dominica or Saint Vincent without a regional procurement arrangement that pools volume across small markets. CARPHA is the natural convener of such an arrangement, and CAIRA has begun a conversation with them.

Third, the operational model is not plug and play. AI second reading changes the radiologist workflow. It changes the recall conversation with patients. It changes liability. It changes how the screening programme reports outcomes to the ministry. Each of these is a project of its own. The Swedish trials succeeded because the screening programmes were already mature, the radiologist communities were trained, and the data infrastructure was in place. None of those conditions are guaranteed across our region.

Fourth, the alternative paths matter. AI is not the only lever. The single highest yield intervention for Caribbean breast cancer outcomes remains getting more women screened in the first place. A study of Jamaica's public mammography uptake from 2022 found that fewer than one in three eligible women had had a mammogram in the prior two years, against a target of two in three. Closing that gap by half would save more lives in the next five years than any imaginable AI deployment, because the limiting factor is not interpretation quality, it is patients walking through the door. Mobile mammography vans, evening clinic hours, and employer subsidized screening days are decidedly low-tech interventions that any health ministry can act on now.

So what should we do.

For Caribbean ministries of health, I would propose a four step path. Step one, fund and publish national breast cancer screening statistics annually. We cannot improve what we do not measure, and in several CARICOM states the most recent published figures are more than five years old. Step two, mandate digital mammography in all public sector units within thirty six months. AI requires digital input. Without that, the rest is moot. Step three, commission an independent regional evaluation of AI mammography on Caribbean cases, conducted through CARPHA in partnership with at least two of our teaching hospitals. Step four, only then, begin staged procurement on a regional pooling basis.

For private sector providers and insurers, I would urge restraint on the marketing. I have already seen advertising in two of our capital cities promoting AI-enhanced mammography as a premium service. The implication that the public service is therefore reading scans with less accuracy is, on present evidence, unsupported. If your facility has deployed an AI tool, the responsible disclosure is the name of the system, its regulatory status, the population it was validated on, and how its outputs are integrated into the radiologist's decision. Anything less is a marketing claim, not a clinical one.

For CAIRA members in radiology, oncology, and primary care, I would ask you to do three things this month. Audit your own knowledge of the AI systems your institutions are evaluating. Speak to your patients about screening, in the actual language and registers they live in, not the language of a regulatory leaflet. And join the CAIRA Working Group on AI in Health, which is consolidating Caribbean clinical perspectives into a position paper we will publish before the regional health ministers meet in December.

The arc of this technology bends, eventually, toward better detection for more women in more places. That is worth working toward. It is also worth being honest that the work is not yet done, the evidence on our population is not yet in, and the women we owe answers to are the ones in front of us today. Get them screened. Then get the technology ready for the ones who come after.

---

Dr. Yolande Pierre-Louis is a consultant radiologist at the University Hospital of the West Indies and chairs the CAIRA Working Group on AI in Health.

Originally published in The Monthly Intelligence Report, October 2024.