Artificial intelligence (AI) algorithms have surpassed the standard clinical risk model in predicting the five-year risk of breast cancer, according to a large-scale study published in the journal Radiology by the Radiological Society of North America (RSNA). The study analysed thousands of mammograms and found that AI algorithms outperformed traditional models in forecasting breast cancer risk.
Typically, a woman's risk of breast cancer is determined using clinical models like the Breast Cancer Surveillance Consortium (BCSC) risk model, which considers factors such as age, family history, childbirth history, and breast density. However, these models often lack comprehensive and readily available data.
Lead researcher Dr. Vignesh A. Arasu, a research scientist and practicing radiologist at Kaiser Permanente Northern California, highlighted the limitations of clinical risk models and the potential of AI. Recent advancements in AI deep learning have enabled the extraction of numerous additional features from mammograms, providing a wealth of valuable information.
The retrospective study analyzed negative screening 2D mammograms conducted in 2016 at Kaiser Permanente Northern California. A random sub-cohort of 13,628 women and an additional 4,584 patients diagnosed with cancer within five years were selected for analysis. The study followed these women until 2021.
The research team divided the five-year study period into three segments: interval cancer risk (diagnosed between 0 and 1 year), future cancer risk (diagnosed between 1 and 5 years), and all cancer risk (diagnosed between 0 and 5 years).
Using five AI algorithms, including two academic and three commercially available ones, the researchers generated risk scores for breast cancer over the five-year period using the 2016 mammograms. They compared these scores with each other and with the BCSC clinical risk score.
The study revealed that all five AI algorithms outperformed the BCSC risk model in predicting breast cancer risk within the five-year timeframe. This suggests that AI can identify missed cancers and breast tissue features that aid in predicting future cancer development, providing valuable insights into the "black box" of AI.
Certain AI algorithms excelled in identifying patients at high risk of interval cancer, a particularly aggressive form that often requires additional screening or follow-up imaging. For instance, when evaluating the top 10% highest-risk individuals, AI predicted up to 28% of cancers, compared to the 21% predicted by the BCSC model.
Interestingly, even AI algorithms trained for shorter time horizons (as low as 3 months) successfully predicted the future risk of cancer up to five years, even when no cancer was initially detected by screening mammography. Combining AI with the BCSC risk model further enhanced the accuracy of cancer prediction.
Dr. Arasu emphasized the practical advantages of mammography-based AI risk models, as they utilise a single data source—the mammogram itself. Some institutions are already implementing AI to aid radiologists in detecting cancer on mammograms. Integrating an individual's future risk score, quickly generated by AI, into the radiology report shared with patients and physicians could provide personalized care and precision medicine on a national scale.
The study's findings demonstrate the potential of AI in revolutionising breast cancer risk prediction, offering accurate, efficient, and scalable means of assessing individual risk. By leveraging AI technology, healthcare providers can enhance early detection and deliver personalized care tailored to each woman's unique risk profile.
