Deep Learning Model Outperforms Traditional Models for Predicting Breast Cancer in Women at High Risk
Key findings
- This study compared the performance of a deep learning risk assessment model and that of two traditional models in 2,168 women at high risk of breast cancer who underwent mammography and supplemental MRI screening
- Women classified by the deep learning model as being at increased or high risk had higher cancer detection rates than women classified with the Tyrer–Cuzick model or Breast Cancer Risk Assessment Tool as being at increased or high risk
- Positive predictive values for abnormal findings at screening, biopsies recommended, and biopsies performed at MRI were higher in women classified as increased or high risk by the deep learning model compared to traditional models
- There was no difference between groups in the abnormal interpretation rate
- In previous research, the deep learning model improved risk discrimination across women of diverse ages, races, and ethnicities, and collectively, these results support its use in routine clinical practice
For individuals whose lifetime risk of developing breast cancer is at least 20%, national guidelines recommend annual surveillance with breast MRI in addition to mammography. Risk assessment is typically performed with the Tyrer–Cuzick (TC) model, also known as the International Breast Cancer Intervention Study (IBIS) model, or the National Cancer Institute Breast Cancer Risk Assessment Tool (BCRAT), also known as the Gail model.
Subscribe to the latest updates from Radiology Advances in Motion
Recent publications in Science Translational Medicine, the Journal of Clinical Oncology, and the Journal of the National Cancer Institute, have shown a deep learning (DL) risk stratification model, Mirai, substantially improves breast cancer risk discrimination compared with traditional models in patients of diverse ages, races, and ethnicities. The image-only model uses the four standard screening mammographic views to predict an individual's risk of developing breast cancer over the next five years.
Now, researchers at Massachusetts General Hospital have shown the DL model also performs better than traditional models in identifying at-risk women who are most likely to benefit from supplemental breast MRI. Leslie R. Lamb, MD, MSc, and Constance D. Lehman, MD, PhD, diagnostic radiologists and investigators in the Department of Radiology, and colleagues report in Radiology.
Methods
The study included 2,168 women 40 years of age or older who were considered at high risk of breast cancer because of a genetic mutation, family or personal history of breast cancer, history of a high-risk lesion, or history of chest radiation. They collectively underwent 4,247 screening MRI examinations in Mass General facilities between September 2017 and September 2020.
For each screening MRI examination, the researchers obtained the corresponding traditional risk assessment score from the most recent screening mammographic examination.
Patients were considered to have increased (intermediate) five-year risk if the TC or BCRAT risk score was 1.67% or higher, and high (lifetime) risk if the TC or BCRAT risk score was 20% or higher. Using the DL model, an absolute score of 2.3 or higher was defined as increased risk, and 6.6 or higher was defined as high risk.
Cancer Detection Rates
For MRI examinations with any risk score available, the cancer detection rate (CDR) per 1,000 examinations was higher for women classified by the DL model as increased or high risk than for women classified as such by the TC model or BCRAT:
Increased risk
- DL model—CDR, 17.1
- TC model—CDR, 6.8 (P<0.01 vs. DL model)
- BCRAT—CDR, 5.3 (P<0.01 vs. DL model)
High risk
- DL model—CDR, 20.6
- TC model—CDR, 6.0 (P<0.01 vs. DL model)
- BCRAT—CDR, 6.8 (P=0.04 vs. DL model)
Positive Predictive Values (PPV)
Likewise, PPV1 (abnormal findings at screening), PPV2 (biopsies recommended), and PPV3 (biopsies performed) were higher in women classified by the DL model as increased or high risk than women classified as such by TC or BCRAT:
Increased risk
- DL model—PPV1, 13.5%; PPV2, 27.9%; PPV3, 30.7%
- 5-year TC model—5.7%; 13.8%; 16.2% (P-value range, <0.01–0.02)
- 5-year BCRAT—5.1%, 11.9%; 14.5% (P-value range, <0.01–0.03)
High risk
- DL model—PPV1, 14.6%; PPV2, 32.4%; PPV3, 36.4%
- Lifetime TC model—5.0%; 12.7%; 13.5% (P-value range, 0.02–0.03)
- Lifetime BCRAT—5.5%; 11.1%; 12.5% (P-value range, 0.02–0.05)
Abnormal Interpretation Rate
There was no evidence of a significant difference in abnormal interpretation rate (percentage of examinations interpreted as positive) between patients found by the DL model to be at increased risk or high risk and those found to be at increased or high risk with the TC model or BCRAT.
These results, in conjunction with the earlier research on the DL model, support its routine use in clinical practice.
view original journal article Subscription may be required
Learn more about the Breast Imaging Division
Learn more about research in the Department of Radiology