Machine learning (ML) can be used to assess coronary artery calcification (CAC) on CT lung screening scans, according to research presented at RSNA 2020.
Coronary artery calcification – calcium buildup within the walls of the arteries of the heart – suggests the presence of atherosclerotic coronary artery disease (CAD) and the potential for serious cardiac events such as heart attack or stroke. The severity of CAD is an independent indicator of a patient’s risk of such an event.
Fuhrman
The patient’s risk is established through the CAC score derived through imaging and which typically measures CAC volume and density.
“Usually these are assessed on cardiac-gated CT scans,” said presenter Jordan Fuhrman, a PhD candidate, Department of Radiology, University of Chicago. “However, recent efforts have attempted to quantify coronary artery calcium on low-dose CT scans obtained from lung cancer screening. A complete evaluation of these scans can be time-consuming and inconsistent.”
Therefore, Fuhrman and colleagues sought to use ML to develop an automatic, complete assessment of CAC. Using thoracic CT scans obtained from the International-Early Lung Cancer Action Program, researchers divided 814 scans into four categories of CAC severity – none, mild, moderate and severe.
Fuhrman explained that these images could also include other sources of calcium not within the coronary arteries. Efforts to identify CAC from an input image have utilized convolutional neural networks (CNNS), and, according to Fuhrman, some studies have attempted to reduce these false positive detections through the use of multiple CNNs.
“However, the use of these additional networks is, in my opinion, not necessary, and we can achieve equivalent performance with reduced computation using only one convolution neural network,” Fuhrman said.
U-Net Architecture Identifies CAC
The U-Net architecture is commonly implemented for biomedical image segmentation. In the study, Fuhrman and his colleagues used a revised U-Net architecture that simultaneously identifies CAC and assigns ordinal scores of between 0 and 3 to each of the four coronary artery branches – the left main artery, left anterior descending artery, right coronary artery, and the circumflex – which corresponds to the volume of CAC in each. The total sum is the CAC volume score (0-12).
Using the revised U-Net architecture, the researchers were able to determine whether a low-dose lung screening CT scan contained CAC with an area under the curve (AUC) of 0.96, establishing a baseline metric that allows them to specify whether a case has CAC. Fuhrman and colleagues further investigated this approach by investigating each of the four coronary artery branches. They determined that the left anterior descending artery performed the best (AUC=0.95), and the left main artery performed the worst (AUC=0.84).
“This was expected, since the left main artery is closest to the aorta and most likely to experience false positive detections through aortic valve calcifications,” Fuhrman said.
In examining confusion matrices for each of the four main coronary artery branches they observed that in most cases they misclassified by a single ordinal severity score, which generally does not have huge clinical implications, Fuhrman said. However, in looking at the left main artery he noted that there were several cases in which predicted severity and ground truth severity differed by an ordinal score of 3, “which is the most severe mistake we could make.”
“We’re still looking into why this is the case,” he said.
As for the study’s limitations, Fuhrman pointed out that that inter-reader variability and motion artifacts were not assessed and could impact results. In addition, 3D information was not included in the revised U-Net architecture, which may improve results at the cost of increased computational requirements, he said.
Going forward, Fuhrman and his colleagues plan to conduct a Free Response Operating Characteristic analysis with the goal of more completely assessing the revised U-Net capabilities. The ultimate goal, he added, is for this technique to be used in the clinic as a computer-aided diagnosis system to reduce reading time and improve reader consistency.
Previous studies have indicated that coronary artery calcium lesion location information is important in assessing a patient's risk of experiencing a major adverse cardiac event, with some branches suggesting higher risk than others. Thus, the goal of this research is to provide a method for the detection of coronary artery calcium lesions in lung cancer screening CT scans and simultaneously determine in which coronary artery branch each lesion lies. The four possible branch assignments were the right coronary artery (RCA), left main artery (LMA), left anterior descending artery (LAD), and circumflex (CFX). This figure identifies an example input image and output segmentations with lesions correctly identified in the LAD and CFX.
For More Information:
View the RSNA 2020 session A Machine Learning Pipeline for Segmentation and Classification of Coronary Artery Calcium Lesions in Lung Cancer Screening CT's — SSCA07 at RSNA2020.RSNA.org.
