An AI model using mammography images to estimate the five-year risk of developing breast cancer has demonstrated more accurate and reliable categorization of women into risk groups compared to traditional breast density assessment. These are the conclusions reached by researchers in a new study that will be presented at the annual conference of the Radiological Society of North America (RSNA).
Clairity Breast is the first FDA-cleared AI breast cancer risk prediction system based solely on mammography images. It was trained on a dataset of 421,499 mammograms obtained at 27 medical institutions in Europe, South America and the United States. The training included both images of women who were subsequently diagnosed with cancer and those who did not develop it within five years. This allowed AI to identify features of tissue structure associated with the future occurrence of the disease. To generate five-year risk forecasts, the model was calibrated on an independent dataset using a deep convolutional neural network.
The model was then applied to a large cohort of 236,422 bilateral 2D screening mammograms from five US medical centers and 8,810 mammograms from a clinic in Europe, obtained between 2011 and 2017. Data on breast density category and information on cancer diagnoses over a five-year period were obtained from medical records and cancer registries. AI predictions were categorized into risk levels according to National Comprehensive Cancer Network thresholds: low (less than 1.7%), intermediate (1.7–3.0%), and high (more than 3.0%). The analysis was carried out using statistical models taking into account the duration of observation and cases of censoring.
Even after taking into account breast density, it was found that women classified as high-risk by AI had cancer more than four times more often than women at average risk (5.9% vs. 1.3%). While breast density itself showed only small differences in incidence – 3.2% in women with dense tissue and 2.7% in women with loose tissue.
These results show that AI risk models significantly outperform traditional approaches based on breast density alone, providing more accurate predictions of cancer risk over the next five years.