A large retrospective study of women with no prior history of breast cancer suggests that artificial intelligence (AI) may be able to identify women at high risk for the development of breast cancer and interval breast cancer four to six years prior to detection.
For the retrospective study, recently published in JAMA Network Open, researchers examined the use of an AI mammography system (Insight MMG, version 1.1.7.2, Lunit) in three rounds of biennial breast cancer screening to score the likelihood of breast cancer from 0 to 100 with 100 representing the highest likelihood of breast cancer.
In the first round of biennial screening, researchers found that the mean AI score in breast developing screening-detected cancer (SDC) was 19.2 in comparison to 9.5 in breasts not developing SDC. The difference in mean AI scores become more pronounced in the second (30.8 vs. 8.2) and third biennial screenings (82.7 vs. 5), according to the study authors.
The researchers also noted significant differences with AI scoring in biennial screening of women diagnosed with interval cancers. In the first screening round, breasts with developing interval cancer had a mean AI score of 17.8 in comparison to 10.1 in the contralateral breast not developing an interval cancer. In the second and third biennial screening rounds, the study authors found that breasts with developing interval cancer had mean AI scores of 20.1 and 33.1, respectively, nearly double and quadruple the AI scoring for contralateral breasts (10.1 and 8.4 respectively).
“Although current commercial AI tools, such as the one used in our study, were not developed or optimized for future cancer risk estimations, we found that the AI system’s discriminatory accuracy for estimating future screening-detected or interval cancer risk 4 to 6 years prior to diagnosis met or exceeded the performance of established risk calculators currently in wide use,” wrote study co-author Solveig Hofvind, Ph.D., head of the Norwegian Breast Cancer Screening Program and professor of radiography at the Oslo and Akershus University College of Applied Sciences in Oslo, Norway, and colleagues.
In differentiating between women with SDC and no breast cancer, the study authors noted the AI platform had a 64 percent area under the curve (AUC) at the first biennial screening, 73 percent at the second round of screening and 97 percent at the third screening round. The AI AUC for interval cancer detection increased from 66 percent for the initial screening round to 78 percent at the third biennial screening, according to the researchers.
The study authors noted that current breast cancer risk calculator models, such as the Tyrer-Cuzick model, the Breast Cancer Risk Assessment Tool (BCRAT) and the Breast Cancer Surveillance Consortium model, have AUCs ranging between 62 to 71 percent, 56 to 68 percent, and 64 to 69 percent, respectively.
“Information about common risk factors for breast cancer is usually not available to radiologists during the interpretation of the screening mammography. An AI system that indicates the woman’s individual risk for breast cancer based solely on mammograms could provide a streamlined, more efficient approach to risk-based screening decisions if image-based AI is found to be as accurate as, or more accurate than, existing risk calculators,” added Hofvind and colleagues.
Three Key Takeaways
1. AI's predictive ability. The AI system (Insight MMG) was able to identify women at high risk of developing both screening-detected cancers (SDCs) and interval breast cancers up to 4–6 years before detection, showing improved discriminatory accuracy over time.
2. Comparative performance. The AI tool's predictive accuracy, as measured by the area under the curve (AUC), outperformed traditional risk calculators like the Tyrer-Cuzick model, BCRAT, and others, especially by the third round of biennial screening.
3. Faster development of interval cancers. Interval cancers showed a lower and slower increase in AI scores compared to SDCs, suggesting they may develop more rapidly and are harder to detect on screening mammograms, reinforcing the need for more advanced detection strategies.
The study authors also observed that SDCs demonstrated higher AI scores at initial biennial screening and had a more accelerated increase in scoring in successive screening rounds than interval cancers.
“This finding suggests that interval cancers develop faster and may be less likely to show suspicious features on screening mammograms compared with screening-detected cancers, indicating that many interval cancers are truly mammographically occult at the time of screening and may not be detectable by the interpreting radiologists.”
(Editor’s note: For related content, see “Mammography Study: Can Stand-Alone AI Enhance Detection of Interval Breast Cancer?,” “Can Multimodal AI Enhance Prediction of Axillary Lymph Node Metastasis Beyond MRI or Ultrasound-Based Models?” and “Mammography Study Shows Supplemental Ultrasound Has Higher Sensitivity than Adjunctive AI in Dense Breasts.”)
In regard to study limitations, the authors noted the retrospective trial design and the lack of racial diversity in the cohort. The researchers also cautioned against broader extrapolation of the study results as the research was based on the assessment of one AI platform.
