📄 radiology and imaging

Evaluating Deep-Learning Based Quantification of Breast Arterial Calcification on Mammography for Cardiovascular Risk Assessment

This study demonstrates that a deep learning model can accurately quantify breast arterial calcification from screening mammograms to identify women at elevated cardiovascular risk, although adding this AI-derived biomarker to existing risk scores like PREVENT offers minimal improvement in overall predictive discrimination.

Original authors: Singh, P., Platt, S., Bussey, O., Heacock, L., Verdone, A., Chen, W., Reynolds, H. R., Yu, C., Shen, Y., Bredella, M. A.

Published 2026-06-18

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Singh, P., Platt, S., Bussey, O., Heacock, L., Verdone, A., Chen, W., Reynolds, H. R., Yu, C., Shen, Y., Bredella, M. A.

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer

The Big Idea: Finding a "Hidden Clue" in Routine Photos

Imagine you go to the doctor for a routine checkup, like getting your eyes tested. While the doctor is looking at your eyes, they accidentally notice a small scratch on your glasses that hints you might be driving too fast. You didn't go there to talk about your driving, but that little clue could help you avoid a future crash.

This study is about finding a similar "hidden clue" in mammograms (routine breast cancer screening photos). The researchers wanted to see if they could use Artificial Intelligence (AI) to spot Breast Arterial Calcification (BAC).

What is BAC? Think of your arteries as water pipes. Over time, they can get clogged with mineral deposits (calcium), just like old pipes get clogged with rust or limescale. When these deposits show up in the blood vessels of the breast, it's called BAC.
The Problem: Doctors usually don't measure these deposits on mammograms because it's hard to see them clearly, and the photos are taken for a different reason (checking for cancer).
The Solution: The team built a "smart robot" (a deep learning model) that can look at these routine photos and automatically count exactly how much "rust" (calcium) is in the pipes.

How They Built the "Smart Robot"

The researchers taught their AI using a massive library of mammograms.

The Training: They showed the AI thousands of photos where human experts had already drawn outlines around the calcium deposits. It was like showing a student thousands of math problems with the answers written in red ink so they could learn the pattern.
The Test: They tested the AI on new photos it had never seen before. The robot was incredibly good at finding the calcium (97% accuracy in spotting it) and very good at measuring how much was there.

What They Discovered

Once the robot was ready, they used it to scan 202,000 women who had routine mammograms. They then waited to see what happened to these women over the next 5 to 10 years.

Here is what they found:

The "Rust" Increases with Age: Just like an old car has more rust than a new one, older women had more calcium deposits.
The "Rust" Predicts Trouble: Women who had a lot of these calcium deposits were much more likely to have a major heart event (like a heart attack or stroke) in the future compared to women with no deposits.
- The Analogy: If you look at a group of people and see that the ones with the most "rust" on their pipes are the ones whose pipes eventually burst, you know that "rust" is a warning sign.
The Numbers:
- Women with no calcium had a low risk of heart trouble (about 1.5% chance in 5 years).
- Women with high amounts of calcium had a much higher risk (about 6.9% chance in 5 years).

Did the AI Replace the Doctor's Current Tools?

The researchers compared their new AI method against the standard tool doctors use today, called the PREVENT score. The PREVENT score is like a calculator that takes your age, blood pressure, cholesterol, and smoking habits to guess your heart risk.

The Result: The PREVENT score was still the better predictor on its own.
The Combination: When they added the AI's calcium measurement to the PREVENT score, it didn't make the prediction much better. It was like adding a tiny bit of extra spice to a dish that was already perfectly seasoned; the flavor didn't change much.

The Bottom Line

The study concludes that:

It Works: We can use AI to automatically find and measure calcium in breast arteries from routine mammograms.
It Matters: The amount of calcium found is a real warning sign for future heart trouble.
The Role: This AI tool shouldn't replace the standard heart risk calculators (like PREVENT). Instead, think of it as a scalable "opportunity." Since millions of women already get mammograms, this AI could act as a safety net. If a woman's standard risk score looks low, but the AI sees a lot of "rust" in her breast arteries, it could be a signal to pay closer attention to her heart health.

Important Note: The paper explicitly states this is a research study and the results have not yet been certified by peer review. It is not yet a rule for doctors to change how they treat patients, but it proves the technology is ready to be tested further.

Technical Summary: Evaluating Deep Learning–Based Quantification of Breast Arterial Calcification on Mammography for Cardiovascular Risk Assessment

Problem Statement
Cardiovascular disease (CVD) remains the leading cause of death among women, yet traditional risk prediction tools often misclassify women, particularly younger and middle-aged cohorts, as low-risk. This underestimation leads to delayed diagnosis and undertreatment. While arterial calcifications are established risk factors for major adverse cardiovascular events (MACE), their detection typically requires specialized testing ordered only after clinical suspicion arises. Screening mammography, performed annually on approximately 40 million women in the U.S., offers a vast, untapped resource for "opportunistic imaging"—the systematic quantification of biomarkers from exams performed for other clinical reasons. However, breast arterial calcifications (BAC) are rarely quantified or systematically reported on routine mammograms. The challenge lies in developing a scalable, automated method to detect and quantify BAC to assess its utility as a predictor of future cardiovascular risk.

Methodology
This retrospective study, conducted at NYU Langone Health, utilized a cohort of 202,006 women (mean age 54.8 ± 11.7 years) who underwent screening full-field digital mammography (FFDM) between 2010 and 2022, excluding those with a prior history of MACE.

Dataset Construction: An expert-annotated dataset of 2,676 FFDM images was created for model development. Images were selected via an enrichment strategy balancing BAC-positive and BAC-negative cases based on radiology reports, with final labels determined by manual review and freehand annotation by a radiology resident, verified by an attending breast radiologist. The dataset was split into training (1,907 images), validation (360 images), and test (409 images) sets.
Deep Learning Architecture: A multi-task convolutional neural network was developed using a U-Net architecture with a ResNet-18 encoder. The model was trained to jointly output:
1. A pixel-level BAC segmentation map.
2. An image-level probability of BAC presence.
  The training objective combined Focal-Tversky loss for segmentation and binary cross-entropy for classification, with a weighted loss function ( $\lambda = 0.1$ ) favoring segmentation.
Quantification Pipeline: Post-segmentation, binary masks were generated using a 0.5 probability threshold. BAC burden was quantified as physical area (mm²) by multiplying the foreground pixel count by the DICOM-encoded pixel spacing values. Exam-level burden was calculated by averaging the area across all four views (craniocaudal and mediolateral oblique) for a patient.
Risk Modeling: The study evaluated the predictive value of AI-derived BAC burden for incident MACE over 5- and 10-year horizons. Four Cox Proportional Hazards models were developed: (i) BAC burden alone, (ii) the ACC/AHA PREVENT score alone, (iii) a combined BAC + PREVENT model, and (iv) a combined model of standard clinical risk factors + BAC. Performance was evaluated on an internal test set and a geographically held-out test cohort (Brooklyn) to assess generalizability.

Key Contributions

Automated Quantification: The study presents a deep learning pipeline capable of detecting and quantifying BAC burden in millimeters squared from routine screening mammograms with high accuracy.
Large-Scale Validation: The model was validated on a massive cohort of over 200,000 women, with a specific geographically held-out test set (Brooklyn) demonstrating robustness across different demographic populations.
Risk Stratification: The research establishes a graded relationship between AI-derived BAC burden and MACE incidence, moving beyond binary presence/absence to continuous burden stratification.
Comparative Analysis: The paper rigorously compares the predictive power of AI-derived BAC against the established PREVENT score and standard clinical risk factors, providing a realistic assessment of BAC's incremental value.

Results

Model Performance: On the geographically held-out test set, the BAC detection model achieved an AUROC of 0.97. For segmentation, the Dice score was 0.6678. The correlation between AI-derived and manually annotated BAC burden was strong (Pearson $r = 0.961$ ), though the model showed a slight systematic overestimation (mean bias +10.3 mm²).
Prevalence and Distribution: AI-detected BAC was present in 23.1% of the cohort. BAC burden increased with age and was right-skewed, with a median burden of 19.51 mm² among BAC-positive patients.
Association with MACE: BAC burden was significantly associated with future MACE. Five-year MACE incidence rose from 1.5% in women with no BAC to 6.9% in those with high BAC burden. Similarly, 10-year incidence increased from 2.6% to 10.4% across these groups.
Predictive Performance:
- BAC Alone: Showed modest discrimination, with 5-year and 10-year AUROCs of 0.661 and 0.650, respectively.
- PREVENT Score Alone: Performed significantly better, with 5-year and 10-year AUROCs of 0.781 and 0.771.
- Combined Models: Adding BAC burden to the PREVENT score resulted in minimal improvement in discrimination (e.g., 5-year AUROC increased by only 0.0001).

Significance and Claims
The authors conclude that deep learning-based quantification of BAC from routine mammography is feasible, accurate, and associated with future cardiovascular risk. However, they maintain a modest stance regarding its clinical utility as a standalone predictor. The paper claims that while BAC burden alone does not significantly improve discrimination when added to the PREVENT score, it serves as a valuable scalable opportunistic imaging biomarker.

The significance lies in the potential to make cardiovascular risk "visible and actionable" for women who might otherwise be classified as low-risk by traditional tools. By leveraging the existing infrastructure of mammography, this approach could identify women at elevated risk without additional cost, radiation, or patient burden, potentially prompting earlier risk assessment, lipid evaluation, and preventive care. The authors position BAC not as a replacement for clinical risk models, but as a complementary, standardized imaging marker that can trigger clinical decision support and preventive interventions.