Coronary artery calcification assessment in National Lung Screening Trial CT images (DeepCAC2)

This paper introduces DeepCAC2, a publicly available dataset and reproducible deep learning pipeline that automatically generates coronary artery calcification scores and risk categories from 127,776 low-dose chest CT scans in the National Lung Screening Trial to facilitate opportunistic cardiovascular risk assessment.

Leonard Nürnberg, Simon Bernatz, Borek Foldyna, Michael T. Lu, Andrey Fedorov, Hugo JWL Aerts

Published 2026-03-27
📖 4 min read☕ Coffee break read

Imagine your heart is like a car engine. Over time, just like an old car, it can develop "rust" or "scale" inside its pipes. In medical terms, this is called Coronary Artery Calcification (CAC). If you have a lot of this "rust," it's a big warning sign that your heart might struggle or fail in the future.

Usually, to check for this rust, doctors need to take a special, expensive photo of the heart using a specific type of X-ray machine (a CT scan) designed just for the heart. But here's the problem: most people getting chest X-rays aren't there for their hearts; they are there to check for lung issues (like cancer). These standard lung scans are "low-dose" and not set up to see the heart's rust clearly. So, even though the heart is right there in the picture, doctors often ignore it because it's hard to measure without a special protocol.

Enter "DeepCAC2": The AI Detective

This paper introduces a new, free tool called DeepCAC2. Think of it as a super-smart AI detective that can look at those standard lung X-rays and say, "Hey, I can actually see the rust on the heart pipes in this picture!"

Here is how they did it, broken down into simple steps:

1. The Training (Teaching the Detective)

You can't just tell an AI to "look for rust" and expect it to work. You have to teach it.

  • The Teachers: The researchers used 778 high-quality heart scans where human experts had already carefully marked every single speck of rust.
  • The Student: They built a digital brain (a 3D U-Net model) and fed it these images. The AI learned to spot the difference between healthy tissue and the white, rocky spots of calcium.
  • The Trick: To make the AI really good, they taught it to look at the heart in tiny 3D chunks (like looking at a wall brick by brick) and made sure it saw plenty of both "rusty" and "clean" examples so it wouldn't get confused.

2. The Big Job (The National Lung Screening Trial)

Once the AI was trained, they gave it a massive homework assignment. They handed it 127,776 chest scans from the National Lung Screening Trial (NLST).

  • The Scale: This is a huge library of scans from over 26,000 people who were being screened for lung cancer.
  • The Result: The AI went through every single scan, found the heart, measured the "rust," and gave each person a score. It did this in about 6 days using powerful computers.

3. The Scorecard (What did they find?)

The AI didn't just say "rust" or "no rust." It gave a score, similar to a credit score, but for your heart health:

  • Score 0: No rust found (Very Low Risk).
  • Score 1-100: A little bit of rust (Low Risk).
  • Score 101-300: Moderate rust (Moderate Risk).
  • Score 300+: Heavy rust (High Risk).

4. Did it work? (The Proof)

The researchers didn't just trust the AI; they tested it.

  • The Test: They compared the AI's scores against the scores of human experts on a smaller set of 390 scans.
  • The Result: The AI and the humans agreed almost perfectly. It was like having two detectives who wrote down the exact same report.
  • The Real-World Proof: They looked at the people in the study over time. They found that people with high "rust" scores (high CAC) were much more likely to pass away from heart-related issues than those with low scores. This proves the AI's scores actually predict real health outcomes.

5. Why is this a big deal? (The "Free Gift")

Usually, if you want to study heart disease, you need a special, expensive heart scan. But this paper says: "You don't need a new scan!"

  • Opportunistic Screening: Since the AI can find heart rust in lung scans, we can now check the heart health of millions of people who are already getting lung scans, without them needing another appointment or radiation.
  • Open Source: The researchers are giving away the "detective" (the AI code), the "case files" (the data), and a "dashboard" (a website where you can look at the results) for free. Anyone can download it, check the work, and use it for their own research.

The Bottom Line

This paper is like discovering that you can check your car's engine oil just by looking at the photo of the car's tires. The DeepCAC2 tool turns a routine lung scan into a heart health check-up, automatically, for free, and with high accuracy. It opens the door to catching heart disease earlier in people who might otherwise be missed.