Sanjay Aneja

M.D.

Yale School of Medicine

Assistant Professor of Therapeutic Radiology

NLST (Learn more about this study)

NLST-911

Apr 25, 2022

Applying deep learning image segmentation to improve coronary artery calcium scoring for cardiovascular outcome prediction

While cardiovascular disease remains a leading cause of morbidity and mortality in the United States, major adverse cardiac events (MACE) such as stroke or myocardial infarction are often undetected until their clinical presentation. Previous studies have risk stratified patients by quantifying the burden of atherosclerotic plaques in coronary vessels through coronary artery calcium (CAC) scores.

In the 1990s, Agatston et. al. demonstrated that coronary artery calcium burden could be calculated using CT scans. The Agatston score (AS) was calculated as the product of coronary artery calcium volume and weighted calcium peak CT attenuation factor (Hounsfield units). A variety of studies have since demonstrated that an increasing Agatston score is highly predictive of MACE and outperforms other risk factors in asymptomatic populations. More recently, Zeleznik et. al. applied deep learning methods to automate calcium score calculations using the Agatston method. In this proof-of-concept study, they developed a deep learning based coronary calcium measurement, which relied on heart segmentation of the CT scans followed by volumetric implementation of AS.

This automated CAC score has also been studied in cancer patients. Cancer patients are a unique population that have an increased mortality risk of cardiovascular disease, however only half of high-risk patients are medically optimized with GDMT. Additionally, cancer patients often require a staging or therapy planning CT scan, which provides a unique opportunity to intervene based on cardiovascular risk. Recently, Atkins et. al. demonstrated the deep learning-based CAC score was significantly predictive of MACE in lung cancer patients. Another study in breast cancer patients demonstrated automated CAC scoring was strongly associated with coronary artery disease.

While the current AS is predictive of MACE , the AS does not provide information on the location or distribution of the plaque. Multiple studies have looked at CAC distribution amongst coronary vessels by manual segmentation. These studies demonstrated the number of coronary vessels with calcium burden and the presence of CAC in the proximal dominant coronary artery independently predict MACE and significantly improved discriminatory capacity of AS to predict MACE. Another study extracted previously defined radiomic features from segmented CAC and generated a radiomic score using 20 of these key features. They demonstrated that when the radiomic score was used with AS, it significantly improved MACE prediction. Another limitation of the current AS is that it measures calcium scores as a product of volume and density. However, previous studies have demonstrated that while volume of calcium burden does correlate with increased risk of major cardiovascular events, density might in fact be inversely correlated and protective of adverse events.

Given recent literature demonstrating the utility in applying deep learning methods to automate calcium scoring and the known limitations of the Agatston scoring method, we aim to use deep learning to develop a more robust imaging based predictive biomarker of MACE. We hope this new biomarker would increase predictive power, enabling clinicians to appropriately risk-stratify and provide preventive care for patients.

Cancer patients provide a unique population for which there is increased risk of MACE and available CT data. We aim to use the NLST cohort as our training and validation cohort. The NLST lung screening trial has low dose non contrast CT scans for 14,959 patients with 12.5 years of follow-up.

Aim 1: Previous studies have demonstrated that calculation of the Agatston score can be automated using deep learning techniques. We will validate the automated Agatston score amongst NLST patients.
• Aim 1A: Using non gated CT scans from NLST, we will train a deep learning model using U-Net to segment coronary vessels from provided CT images. We will then calculate an Agatston score using the coronary vessel image segments and validate it as a predictive biomarker of MACE amongst NLST patients.
• Aim 1B: Previous literature demonstrates calcium scoring varies across demographic data (sex, race, age) and imaging parameters. We will stress test our automated calcium score against demographic and imaging parameters to understand how our model performs in different populations.

Aim 2: Because of the limitations of the Agatston score, we wish to better understand atherosclerotic plaque features in CT to identify unique image features that are predictive of MACE.
• Aim 2A: We hope to use both a supervised and unsupervised deep learning architecture with the segmented CAC vessels to identify key image features that are predictive of MACE. We will compare these features against other radiomic features for predictive power.
• Aim 2B: Previous studies largely rely on image features contained within the coronary vasculature, and we wish to identify if there are any other features in CT scans including heart size, locations of atherosclerotic plaques, or lung lesions that may be predictive of MACE. We will use the whole CT scan as input into a deep learning architecture to identify any other image features that are predictive of MACE.