A recent study has revealed that artificial intelligence can identify subtle indicators of aggressive breast cancers in routine mammograms, potentially years before these cancers are diagnosed. Conducted by researchers and published in the journal npj Digital Medicine, the study analyzed a substantial dataset of 112,621 mammograms collected from the UK’s National Health Service (NHS) between 2014 and 2017.
The study aimed to evaluate the effectiveness of four advanced Deep Learning (DL) algorithms in predicting “interval cancers,” which account for around 30% of breast cancers diagnosed after a negative screening mammogram and before the next scheduled examination. These cancers often develop aggressively and pose significant challenges within current screening protocols.
Evaluation of Deep Learning Models
The retrospective validation study involved a head-to-head comparison of four leading DL models: Mirai, developed by the Massachusetts Institute of Technology (MIT); iCAD ProFound AI Risk; Transpara Risk; and Google Health’s Risk Model. The analysis found that the Mirai model exhibited the highest predictive performance for identifying interval cancers, achieving an Area Under the Curve (AUC) score of 0.77 for interval cancer detection. This model flagged approximately 27.5% of interval cancers from the cohort by identifying the top 4% of “normal” mammogram images as high-risk.
While performance varied across the different mammography machines, the findings indicate that these AI tools could enhance risk-stratified screening strategies. However, the authors emphasized the need for further prospective clinical evaluations before these tools can be integrated into routine practice.
Addressing the Challenge of Interval Cancers
Interval cancers present a significant diagnostic challenge in breast cancer screening. Traditionally, screening recommendations have involved mammograms every few years. Although effective in identifying many cancers, these screenings often miss those that arise between appointments. Interval cancers tend to be more aggressive, resulting in poorer prognoses and outcomes.
Historically, clinicians have relied on genetic assessments and family history evaluations to estimate individual risk. Yet, these methods are not routinely implemented and often lack comprehensiveness. Advances in DL algorithms, trained on extensive datasets of mammogram images, suggest that AI models can discern subtle imaging patterns that may elude human radiologists.
The study’s findings underscore the potential of AI to significantly improve early detection of interval cancers. By identifying women at higher risk based on prior “normal” screenings, healthcare providers could implement more personalized monitoring strategies.
The dataset used in the study included high-resolution mammograms from two distinct NHS screening sites, allowing researchers to track participants for five years. Approximately 1,225 cases of breast cancer were identified during the follow-up period, including interval cancers.
The evaluation also examined how well these algorithms performed across different mammography hardware from manufacturers like Philips and GE. While most models demonstrated consistent performance, the Transpara model showed a modest advantage when analyzing images from GE machines compared to those from Philips.
Despite the promising results, the study acknowledged several limitations. The exclusion of certain mammograms and incomplete ethnicity data may restrict the generalizability of the findings. Additionally, retrospective validation may not fully capture the clinical utility of these models, as some cancers could emerge through other imaging pathways.
The research highlights that while AI models like Mirai can identify significant indicators of hidden cancers, further investigation through prospective clinical trials is essential before their integration into routine breast cancer screening practices. The potential for these AI tools to reshape breast cancer detection strategies is evident, paving the way for more effective and personalized patient care.
