AI Offers Groundbreaking Hope for Improved Breast Cancer Diagnostics in Low-Resource Settings

Researchers have just revealed a groundbreaking use of artificial intelligence to address a critical gap in healthcare diagnostics in low- to middle-income countries.  

The findings, published in an abstract presentation by the American Society of Clinical Oncology, revealed that when a deep learning model is trained on widely available pathology slides, it can accurately predict a breast cancer patient’s tumor estrogen receptor status without the need for more expensive immunohistochemistry tests.   

This breakthrough may soon help low- to middle-income countries access breast cancer diagnostics that were once out of reach, leading to better outcomes for patients, and may have significant implications for global health. 

An urgent need for innovation in global cancer care 

While breast cancer survival rates are improving in high-income countries, many low- and middle-income countries aren’t experiencing the same. The absence of affordable and accurate diagnostic tools has forced doctors to guess at the best treatments, often with poor outcomes.   

“The situation in Haiti is especially dire, with a mortality-to-incidence ratio for breast cancer patients over 60%, among the highest in the world,” says Dagoberto Pulido-Arias, a research scientist at Athinoula A. Martinos Center for Biomedical Imaging Massachusetts, and co-author of the study. “This disparity highlights a public health emergency.”   

Haiti’s mortality ratio is so high because the medical community lacks easy access to immunohistochemistry (IHC), a test that is routinely used to determine tumor estrogen receptor (ER) status.  

“In settings like Haiti, ER assessment with IHC is not available, either because government-funded hospitals cannot afford to pay for it and/or lab professionals don’t have the technical skills to perform the test,” Mr. Pulido-Arias says. “As a result, it is often out of reach financially for patients to afford the test from a private lab, and it is often poorly quality controlled in public-funded hospital laboratories. Our tool bypasses the need for this separate, costly, and logistically complex test.”  

Developing a deep learning model that can help  

In response to this urgent diagnostic gap, the authors developed a deep learning model they named ESPWA, which means “hope” in Haitian Creole.   

The researchers trained ESPWA to use hematoxylin and eosin (H&E) stained whole slide images to accurately predict a tumor’s ER status without the need for IHC. The model can do this by detecting nuances in the slides that human eyes often miss.   

“A typical H&E slide contains hundreds of thousands of cells, and our deep learning model is trained to identify subtle spatial patterns in the cell populations that are indicative of the tumor's ER status,” Mr. Pulido-Arias says. “While many of these patterns are recognized by well-trained human pathologists, the model can learn them by analyzing thousands of examples, and read them faster, more accurately, and potentially see features the human eye misses.”   

Training the AI on population-specific datasets  

Initially, ESPWA was trained on a large U.S. dataset from The Cancer Genome Atlas (TCGA), achieving high accuracy when tested on slides from U.S. patients.   

But when the same AI model was given slides from Haitian patients to analyze, its accuracy dropped.   

“Its performance [called ‘area under receiver operating characteristic’ or AUROC] dropped significantly, from 0.846 to 0.671,” Mr. Pulido-Arias says. “This demonstrates that models trained on one population, in this case a predominantly Caucasian one, are unlikely to generalize well to other populations, such as Black patients in Haiti.”  

This phenomenon is known as domain shift.  

Once the deep learning model was fine-tuned with breast cancer slides from Zanmi Lasant, a hospital and non-government healthcare provider in Haiti, its AUROC improved to 0.85 when analyzing Haitian patients’ slides.   

“We believe this is primarily because there is a different demographic of breast tumor biology in Haiti, as is also seen across Africa, with more aggressive tumors overall, and more aggressive ER-positive tumors than typically seen in Caucasian populations,” Mr. Pulido-Arias says.  

This outcome underscores the importance of training AI models on diverse, population-specific datasets.  

Precision medicine without the price tag  

Knowing if a patient’s breast cancer is ER-positive or negative determines whether the patient is likely to benefit from endocrine therapy. In Haiti, where immunohistochemistry can be nearly impossible to access, most patients are given endocrine therapy as the default, regardless of whether their ER status is known or not.   

But this practice can actually increase the mortality rate for some patients.   

“This ‘one-size-fits-all’ strategy can result in increased morbidity, as endocrine therapy in patients with ER-negative disease can worsen their disease burden and cause considerable toxicity,” Mr. Pulido-Arias says.   

Thanks to their new deep learning model, this approach can be changed.   

“ESPWA can help clinicians avoid this, enabling precision-based delivery of endocrine therapies and optimizing therapeutic decision-making for each patient,” he says.  

By offering a low-cost, scalable method for accurately assessing ER status, ESPWA can allow Haitian clinicians to tailor their treatments for the specific kind of cancer their patients have. This will improve outcomes for patients who might otherwise fall through the cracks.   

Catching what medical professionals might miss 

Another impressive feature of ESPWA is how it can even outperform human doctors in predicting estrogen receptor status.   

“ESPWA demonstrated promising accuracy across different tumor types and grades, and in a head-to-head comparison, it was able to outperform an expert pathologist in accurately determining ER status on our dataset of 3,448 images,” Mr. Pulido-Arias says.   

This impressive performance demonstrates how vital ESPWA and similar AI algorithms will be in real-world settings.   

A major step forward for global health equity  

The next major step is a clinical trial with patients at Zanmi Lasante once it reopens.   

“The essential next step is to validate these findings in a real-world setting,” Mr. Pulido-Arias says.   

Because the process only requires tools already available in most pathology laboratories — like a digital slide scanner and standard H&E slides — it can be easily applied to more countries once the clinical trial is completed.   

“The key advantage is that our tool works with the standard H&E slides that are already prepared for initial diagnosis and eliminates the need for the specialized reagents and equipment for IHC,” Mr. Pulido-Arias says. “This makes the technology far more accessible for any laboratory that has, or can acquire, a slide scanner.”  

Mr. Pulido-Arias and his co-authors believe this deep learning algorithm represents a case study in how artificial intelligence can be used to promote global health equity and reshape the future of patients in low-income countries.   

“In resource-limited settings like Haiti, where timely and complete pathologic assessment is nearly impossible, AI can help standardize and accelerate the analysis of complex data,” Mr. Pulido-Arias says. “It can provide critical information to guide treatment, moving away from empirical strategies that can harm patients.”  

And as deep learning tools continue to evolve, their potential to uncover new patterns invisible to the human eye could fundamentally reshape how pathology functions on a global scale.   

“The H&E slide contains a massive amount of data that we are just beginning to understand,” Mr. Pulido-Arias says. “The future lies in using these computational tools to extract a much deeper layer of information from this routine diagnostic test, which has the potential to transform how we make therapeutic decisions for cancer patients everywhere.” 

You can read the preprint of the study here. 

Mr. Pulido-Arias collaborated with the rest of the study’s authors to answer Critical Values’ questions: Rebecca Henderson, Marie Djenane Josert, Christophe Millien, Joarly Lormil, Michael Mathelier, Maisha Corrieulus, Gabriel Flambe, Azin Mashayekhi, Jane Brock, Scott Kilcoyne, Ali Brown, Dan Milner, Ken Landgraf, Albert Kim, and Christopher Bridge.