Using AI and machine learning to more accurately detect breast cancer is not new, but as it continues, the amount of data and length of experiments is highlighting its value.
A recently study from the University of Warwick in the UK explored the potential of artificial intelligence (AI) tools in evaluating the risk of breast cancer using digital mammograms, especially after a negative screening.
Researchers hypothesized that AI algorithms known for their effectiveness in detecting breast cancer could also be highly efficient in assessing the risk of developing the disease.
To test this, a detailed case-control study was designed, utilizing mammograms from the OPTIMAM registry, which included data from women diagnosed with breast cancer between 2010 and 2019 through the English breast screening program.
The study involved comparing mammograms taken at the time of diagnosis with those from a previous screening, three years earlier, when no cancer was detected.
This comparison was made for 3,386 women, with cases being those diagnosed with cancer and controls matched by age, screening location, and mammography machine type.
The study’s main tool for risk assessment was the Mirai algorithm, designed specifically for this purpose, alongside three other deep-learning algorithms originally developed for cancer detection.
The performance of these algorithms was measured using a statistic called the matched area under the curve (AUC), with results showing that the Mirai algorithm performed best for both risk assessment and cancer detection.
Notably, the study found a correlation between an algorithm’s performance in detecting cancer and its ability to assess risk, particularly with larger cancers showing higher accuracy.
The findings suggest that enhancing AI algorithms’ ability to detect smaller cancers could significantly improve their performance in risk assessment.
This has implications for the future of breast cancer screening, hinting at the possibility of repurposing state-of-the-art cancer detection algorithms for risk assessment. Such advancements could enable more personalized screening strategies, directing more intensive screening efforts towards those at higher risk of developing breast cancer.
The study’s strengths include its design, which directly compares algorithm performance on the same women’s mammograms at different times, and its external validation of algorithms without additional training.
However, there are limitations, such as the black-box nature of AI algorithms, making it unclear exactly what features the algorithms are identifying as risk factors. Additionally, the study’s retrospective design and focus on a specific dataset may limit its generalizability.
Seeing AI/ML’s value opens avenues for future research and development in AI-driven breast cancer screening, with the potential to refine screening protocols and offer more targeted and effective early detection strategies.
