A more equitable AI model for breast cancer risk assessment

The tool can perform accurately across diverse racial backgrounds

Image credit: Image Creator from Microsoft Designer/DALL.E and Canva

At the recent annual meeting of the Radiological Society of North America (RSNA), researchers presented a groundbreaking advancement in breast cancer risk assessment.

They introduced a deep learning artificial intelligence model designed solely using mammogram images. This innovative model exhibited remarkable accuracy in predicting both ductal carcinoma in situ (DCIS) and invasive breast cancer, marking a significant leap forward in breast cancer risk evaluation.

Traditionally, assessing breast cancer risk relied on models that gathered information from patient questionnaires, mainly centered around medical and reproductive history. However, these traditional models often lacked precision at the individual level and demonstrated biases, particularly across different racial groups. These biases were primarily attributed to the data used to develop these models, which were often derived from predominantly European Caucasian populations.

The research team, led by Leslie R. Lamb, M.D., M.Sc., a breast radiologist at Massachusetts General Hospital (MGH), addressed these limitations by developing an AI model that solely relied on mammographic images. Their aim was to create a risk assessment tool that could perform accurately across diverse racial backgrounds, thus improving early detection and survival rates for all populations.

The AI model's performance was evaluated using the Area Under the Curve (AUC) measurement, a standard gauge of predictive accuracy ranging from 0 to 1. Impressively, this AI model consistently scored 0.71 across all racial groups in predicting both DCIS and invasive breast cancer. In contrast, traditional risk assessment models typically yielded lower scores, especially for non-white populations.

The study included a large dataset of 129,340 routine bilateral screening mammograms performed in 71,479 women between 2009 and 2018, with a five-year follow-up. The racial composition of the study group comprised white, Black, Asian, self-reported other races, and unknown categories.

The AI model's superior performance in predicting breast cancer risk, without showing biases across different races, could mark a significant step forward in breast cancer risk assessment.