Add Predicting Heart Disease Mortality Risk to LDCT Capabilities

Low-dose CT (LDCT) isn’t just for lung cancer screening. Data gathered from the scans can also help predict a patient’s five-year risk of death from cardiovascular disease.

In a study published April 15 in Radiology: Cardiothoracic Imaging, investigators from The Netherlands detailed how implementing a deep learning algorithm that is fueled with information captured during a low-dose CT exam for lung cancer screening can offer valuable details about calcification in the heart and the aorta.

Using LDCT in this way is a two-for-one opportunity to screen patients for both potentially deadly conditions. The scan is used to detect lung cancer in high-risk individuals, including heavy smokers, but for those individuals, cardiovascular disease actually surpasses lung cancer as the leading cause of death. By picking up these areas of calcification that are linked to plaque build-up, the scan can give providers a strong predictor of heart attacks, stroke, and mortality from cardiovascular disease.

Given the recent recommendation update from the U.S Preventive Services Task Force that opened up LDCT lung cancer screening to high-risk individuals ages 50-to-80 who have at least a 20-pack year smoking history, it is possible that providers will be able to pick up on cardiovascular disease mortality risk in an even wider population of patients, the team said.

In the past, data pulled from CT images has been used in conjunction with other risk factors, such as cholesterol levels, blood pressure, and self-reported history of illness. But, for this study, the team, led by Bob D. de Vos, Ph.D., from Amsterdam University Medical Center and the Image Sciences Institute at the University Medical Center Utrecht, drew upon deep learning to create an automated method that can quickly predict five-year cardiovascular disease mortality without putting a heavier work burden on providers.

“We have shown that five-year cardiovascular disease mortality can be predicted for lung screening participants in less than half a second, using only site-specific calcium scores automatically derived from lung screening low-dose CT,” the team said. “Hence, the proposed image-based analysis could aid in identification of lung screening participants at risk for cardiovascular disease mortality, without relying on self-reported participant data.”

Figure 2: Projections of all aligned chest CT scans show feasibility of slab-based quantification of calcium, resulting in an average image. For alignment, only translation, rotation, and scaling were allowed, resulting in a blurry image, because not all anatomy is exactly the same across participants. From left to right, the center axial, sagittal, and coronal sections are shown. Note that field of view is similar to cardiac CT, which is a consequence of image alignment by the used automatic calcium scoring method. Image alignment allowed the determination of calcification distributions into slabs as a proxy measure for proximal and distal calcifications, of which the borders are indicated by the horizontal lines.

Credit: RSNA

For the study, they pulled data from LDCT scans conducted between August 2002 and April 2004 on 4,451 participants with an average age of 61 who were part of the National Lung Screening Trial. Using those details, the team trained their deep learning method to quantify six types of vascular calcification: thoracic aorta calcification (TAC), aortic and mitral valve calcification, coronary artery calcification (CAC) of the left main, the left anterior descending, and the right coronary artery. After training, they tested their method on data gleaned from 1,113 participants.

The team compared the methods with semi-automatic baseline prediction using self-reported participant characteristics, such as age, smoking history, and history of illness. According to their analysis, the prediction model using calcium scores performed better than the baseline model.

Specifically, the prediction model that used calcium scores reached a C statistic of 0.74 while the baseline model that relied on only participant characteristics achieved 0.69. The model had the best results, they said, when all variables were combined – a 0.76 C statistic.

“The analysis shows we found predictors that are typically not described in a literature, possibly because we performed analysis in lung cancer screening participants we are already at high risk of cardiovascular disease from a history of heaving smoking and the presence of extensive arterial calcification,” de Vos explained.

The method works in two stages, he said, first using deep learning to pinpoint the amount and location of arterial calcification in the coronary arteries and aorta. The second stage relies on a more conventional statistical approach to mortality prediction and identifies features that are most predictive of five-year mortality.

Incorporating the deep learning method into lung cancer screening would be simple, de Vos said, because it requires no special equipment and does not tack on any exam time. And, the benefits to providers and patients could be immense.

“The method uses only image information, it is fully automatic, and it is fast. The method obtains calcium scores in a complete chest CT in less than half a second. This means that the method should be easy to implement to routine patient work ups and screening,” he said. “Lung screening studies show that heavy smokers die from cardiovascular disease as much as from lung cancer. But, we also see that some people with very high calcium scores survive, while others with low scores do suffer from major cardiac events.”

Their results point to a need for additional research that can more specifically point to calcifications that are dangerous. To date, they have developed several automatic scoring methods that can use varied types of data, but they are continuing on with development. Now, de Vos said, they are concentrating on creating a calcium-scoring method that can use low-quality data, such as data affected by cardiac motion, low image resolution, and high noise level, to detect arterial calcification.

“We developed a method, for example, that can detect coronary calcification even when the lesions are below the clinically used threshold,” he said. “This way, we hope to increase the reproducibility of calcium scoring and enable more accurate prediction.”

But, for now, he said, their work has shown that it is possible to achieve real-time mortality prediction with lung screening chest CT by using only data from automatic deep learning-quantified arterial calcium.

For more coverage based on industry expert insights and research, subscribe to the Diagnostic Imaging e-Newsletter here.