In comparison to traditional models, artificial intelligence (AI) may be more than three times as effective at identifying women at high risk for breast cancer, according to newly published research.
For the retrospective study, recently published in Radiology, researchers compared breast cancer risk assessments for a mammography-based deep learning model, the Tyrer-Cuzick (TC) risk model and the National Cancer Institute Breast Cancer Risk Assessment Tool (BCRAT). The cohort included 2,168 patients who had a total of 4,247 breast magnetic resonance imaging (MRI) exams. According to the study authors, the deep learning model was trained on over 210,000 mammograms from over 56,000 patients with subsequent validation and testing on over 25,600 screening mammograms.
For patients identified as high-risk by the deep learning five-year model, subsequent breast MRI identified 12 breast cancers out of 583 exams (or a 20.6 cancer detection rate (CDR) per 1,000 examinations). In contrast, for patients identified as high-risk by the TC lifetime risk model, there were seven breast cancer cases diagnosed out of 1,166 MRI exams (a 6.0 CDR per 1,000 exams). Of the high-risk patients identified with the BCRAT lifetime risk model, MRI detected five breast cancer cases in 737 exams (a 6.8 CDR per 1,000 exams), according to the study.
“Decisions regarding whether to perform supplemental screening, and whether insurance companies will pay for it, are driven by an individual’s breast cancer risk score, which has historically been calculated with traditional methods,” wrote Leslie R. Lamb, M.D., an assistant professor of radiology in the Department of Radiology at Massachusettes General Hospital and Harvard Medical School, and colleagues.
“The DL risk stratification model used in this study — trained, validated, and tested on screening mammographic images alone — has been previously shown to reliably identify patients at high risk of developing breast cancer, outperforming traditional risk models in screening mammography cohorts across multiple institutions globally.”
The researchers also noted that high breast cancer risk scores with the deep learning model were associated with significantly higher positive predictive values (PPVs) than the TC and BCRAT models.
Specifically, the study authors said the deep learning model had a 14.6 percent PPV for abnormal findings at screening in comparison to 5 percent for the TC model and 5.5 percent for the BCRAT model. For recommended biopsies, the MRI exams associated with the deep learning model had a 32.4 percent PPV in comparison to 12.7 percent for the TC model and 11.1 percent for the BCRAT model.
Three Key Takeaways
1. AI improves breast cancer risk assessment. The study suggests that a mammography-based deep learning model is more than three times as effective as traditional risk assessment tools at identifying women at high risk for breast cancer. This AI model outperformed the Tyrer-Cuzick risk model and the National Cancer Institute Breast Cancer Risk Assessment Tool (BCRAT) in identifying high-risk patients.
2. Increased cancer detection rate. Patients identified as high-risk by the deep learning model had a significantly higher cancer detection rate on breast MRI compared to those identified as high-risk by traditional models. The deep learning model achieved a 20.6 cancer detection rate (CDR) per 1,000 examinations, whereas the TC model had a CDR of 6.0 per 1,000 exams, and the BCRAT model had a CDR of 6.8 per 1,000 exams.
3. Higher positive predictive values (PPVs). The deep learning model not only detected more cancers but also had significantly higher positive predictive values (PPVs) for abnormal findings at screening and recommended biopsies. The deep learning model had a 14.6 percent PPV for abnormal findings at screening, compared to 5 percent for the TC model and 5.5 percent for the BCRAT model. For recommended biopsies, the deep learning model had a 32.4 percent PPV, whereas the TC model had 12.7 percent and the BCRAT model had 11.1 percent.
“We found that patients classified as having increased or high risk by the (deep learning) model, as compared with the Tyrer-Cuzick and Breast Cancer Risk Assessment Tool models, had higher cancer detection rate and higher positive predictive value (PPV) for abnormal findings at screening, PPV for biopsies recommended, and PPV for biopsies performed at MRI, but no difference in abnormal interpretation rate was observed,” noted Lamb and colleagues.
(Editor’s note: For related content, see “The Reading Room Podcast: Emerging Concepts in Breast Cancer Screening and Health Equity Implications, Part 3,” “Combining AI Lesion Detection, Mammographic Texture Model Improves Breast Cancer Risk Assessment” and “Mammography-Based Deep Learning Model May Help Detect Precancerous Changes in High-Risk Women.”)
Beyond the inherent limitations of a retrospective, single-center study, the authors acknowledged that 91.8 percent of the patient cohort was White and emphasized the need to validate their findings in larger, more racially diverse populations. Noting that women at high risk for breast cancer commonly forego MRI, the researchers conceded an inherent selection bias with a cohort focused on those who have had MRI exams. The study authors also pointed out that a significant number of women, who did not have a bilateral screening mammogram at the facility in the two years prior to the study, were excluded from the study.
