Skip to main content
Diagnostic imaging

Diagnostic imaging

Open-source algorithm predicts heart attack risk from chest CT scan

16 Feb 2021 Samuel Vennin 
Chest CT scans

A heart attack occurs when the coronary arteries responsible for supplying the heart with blood and oxygen become blocked. This does not happen overnight and is usually the result of fatty or calcium plaques being slowly deposited. At first, these plaques hinder the efficient supply of blood to the heart muscle (myocardial perfusion). Eventually, the rupture of one plaque can cause a blood clot to form, blocking the coronary arteries and preventing myocardial perfusion. For this reason, coronary artery calcification is an important and independent predictor of adverse cardiovascular events such as heart attacks.

But despite this knowledge, and the fact that it can be assessed from any chest CT scan, quantification of coronary artery calcium (CAC) is not automatically integrated in the patient pathway as it requires radiological expertise, time and specialized equipment. To remedy this, a multidisciplinary team from Brigham and Women’s Hospital‘s Artificial Intelligence in Medicine Program led by Hugo Aerts, and the Massachusetts General Hospital’s Cardiovascular Imaging Research Center led by Udo Hoffmann, developed and tested a deep-learning algorithm that can automatically quantify CAC from any chest CT scan. Their findings are reported in Nature Communications and the algorithm is available as free open source software at the AIM website.

“In theory, the deep-learning system does a lot of what a human would do to quantify calcium,” said first author Roman Zeleznik. “Our paper shows that it may be possible to do this in an automated fashion.”

Developed from 1636 scans, tested on 20,084 patients

To develop the algorithm, the researchers used scans from the Framingham Heart Study – a seminal yet ongoing cardiovascular study that since 1948 has been investigating the health of the residents of Framingham, MA. They used 1636 CT scans from the study (acquired from the third generation onward) to identify and quantify CAC, using manual segmentations performed by expert CT readers as ground truth to train the deep-learning system.

The deep-learning system uses three consecutive convolutional neural networks to predict the heart centre, segment the heart, and segment and identify coronary calcium in less than 2 s. It then computes the CAC scores and stratifies them into clinically relevant categories:  very low (CAC=0); low (CAC=1–100); moderate (CAC=101–300); and high (CAC>300).

The strength of the study lies in the breadth and scope of the datasets that the deep-learning system was subsequently tested on. The team used four cohorts focusing on different pathologies: 663 Framingham Heart Study participants who underwent cardiac CT (none of whom were in the training group); 14,959 heavy smokers having lung cancer screening CT (NLST trial); 4021 patients with stable chest pain having cardiac CT (PROMISE trial); and 441 patients with acute chest pain having cardiac CT (ROMICAT-II trial).

First, the team investigated whether the system was accurate in 5521 scans where expert’s measurements were available. The algorithm’s CAC scores were highly correlated with the manual scoring and most differences occurred between adjacent risk categories.

Predicting future cardiovascular events

The team also investigated the predictive potential of their system. As patients enrolled in these studies usually had a follow-up visit, the researchers looked for correlations between CAC scores and cardiovascular events. For example, in the NLST study, median follow-up time was 6.7 years after the first CT scan was acquired. They saw a clear association between CAC scores and death from cardiovascular disease: taking the very low CAC score group as reference, the low, moderate and high score groups had 57%, 179% and 287% more such deaths, respectively. Similar trends were observed in the other studies.

The diversity of the datasets used strengthens the generalizability of these results to clinical settings. “The coronary artery calcium score can help patients and physicians make informed, personalized decisions about whether to take a statin [a medication which reduces cholesterol and the risk of heart attack]. From a clinical perspective, our long-term goal is to implement this deep-learning system in electronic health records, to automatically identify the patients at high risk,” concludes co-senior author Michael Lu.

Copyright © 2024 by IOP Publishing Ltd and individual contributors