Racial, Ethnic Minorities Are Underrepresented in AI Mammogram Interpretation

A study published in the European Journal of Cancer raises concerns about the fairness and equity of AI-driven mammogram interpretation due to the underrepresentation of racial and ethnic diversity in datasets. While AI holds promise for improving mammogram analysis, especially in underserved areas, the study warns that limited diversity in both datasets and the researchers involved in developing these models could undermine the generalizability and fairness of the technology.

Researchers conducted a scientometric review of studies from 2017, 2018, 2022, and 2023 that used screening or diagnostic mammograms to train or validate AI algorithms. Out of 5,774 identified studies, only 264 met the criteria for inclusion. Despite a 311% increase in studies—from 28 in 2017–2018 to 115 in 2022–2023—only 0–25% reported on the race or ethnicity of participants, most of whom were identified as Caucasian. Patient cohorts primarily came from high-income countries, with none from low-income regions, and the majority of authors were also based in wealthier nations. Gender imbalances were noted among both first and last authors.

The study concludes that this lack of racial, ethnic, and geographic diversity—both in datasets and among researchers—could negatively impact the accuracy and fairness of AI-based mammogram interpretation, potentially worsening existing health disparities. Algorithms trained predominantly on Caucasian populations may produce inaccurate results for underrepresented groups, increasing the risk of misdiagnosis and poorer outcomes.

To ensure equitable advances in breast cancer care, the authors stress the importance of collecting diverse datasets and fostering international collaboration, particularly with researchers from low- and middle-income countries. They argue that without such efforts, AI tools risk systematically disadvantaging certain racial, ethnic, or socio-demographic groups.

This concern emerges amid broader industry efforts to improve cancer detection and treatment through AI. In February, Google partnered with the Institute of Women’s Cancers at France’s Institut Curie to explore AI’s role in forecasting cancer progression and relapse risk, focusing on aggressive forms like triple-negative breast cancer. In 2024, AI biotech firm Owkin teamed up with AstraZeneca to develop an AI tool for pre-screening gBRCA mutations in breast cancer patients using digitized pathology slides, aiming to increase access to genetic testing.

That same year, AI cancer diagnostics company Lunit partnered with Volpara Health to build an integrated AI ecosystem for early cancer detection and risk prediction. Lunit later acquired Volpara, incorporating tools like the Scorecard breast density assessment into its platform. Before this acquisition, Lunit had already expanded its reach by supplying its AI-powered mammography analysis software, Lunit INSIGHT MMG, to Capio S:t Göran Hospital in Sweden, helping analyze breast images for about 78,000 patients annually.