Related Experiment Video
Updated: May 5, 2026
07:15
Machine Learning Algorithms for Early Detection of Bone Metastases in an Experimental Rat Model
Published on: August 16, 2020
5.8K
A Deep Learning Breast Cancer Risk Model for Precise Supplemental Screening
Leslie R Lamb1,2, Sarah F Mercaldo1,2, Andrew Carney1
1Department of Radiology, Massachusetts General Hospital, Boston.
JAMA Network Open
|May 4, 2026
Summary
A deep learning (DL) breast cancer risk model significantly outperformed radiologist-assessed breast density in predicting future cancer and false-negative screening results. This suggests a shift towards DL models for personalized breast cancer screening.
Area of Science:
- Radiology
- Artificial Intelligence
- Oncology
Background:
- Current federal legislation requires informing patients of breast density, a factor associated with increased breast cancer risk and potential masking of tumors.
- Breast density assessment is subjective and varies between readers, limiting its effectiveness in guiding supplemental imaging decisions.
- A binary dense/nondense classification applies to 40-50% of women, highlighting the need for more precise risk stratification methods.
Purpose of the Study:
- To compare the performance of a deep learning (DL) breast cancer risk model against radiologist-assessed breast density.
- To evaluate the models' efficacy in estimating future breast cancer risk and identifying false-negative (FN) screening results.
Main Methods:
- A retrospective cohort study analyzed screening mammograms from 67,019 women aged 30+ over a 10-year period (2009-2018).
- A DL risk model was applied to mammograms, and breast density was categorized using BI-RADS criteria.
- Primary outcomes included 5-year breast cancer diagnoses and FN screening results; performance was assessed using AUROC.
Main Results:
- The DL model showed significantly higher accuracy in predicting future cancer (AUROC 0.71) compared to breast density (AUROC 0.53).
- False-negative rates increased across DL risk groups, while women with dense breasts had higher FN rates than those with nondense breasts.
- Adding breast density information to the DL model did not enhance its predictive performance.
Conclusions:
- A DL risk model demonstrates superior performance over subjective breast density assessment for predicting breast cancer risk and FN screening results.
- These findings advocate for transitioning from density-based policies to more accurate, image-derived risk models for guiding supplemental screening.
- Implementing DL models can lead to more precise and personalized breast cancer screening strategies.