Utilizing a deep learning algorithm could help radiologists determine valuable coronary artery calcium scores (CACS) in a fraction of the time.
Such scores have proven to be more predictive of cardiovascular risk than any other biomarker, but quantifying CACS via imaging remains a time-consuming and labor-intensive task, wrote authors of a new study published Nov. 11 in Clinical Radiology. Those researchers found deep learning could change that, offering high-quality measurements in less time.
“There have been no studies to date evaluating the use of AI for the quantification of CACS from CT calcium score imaging,” W. Wang, with Beijing Anzhen Hospital’s Department of Radiology, and colleagues wrote. “This study demonstrates that the deep learning algorithm provides reliable calcium score and risk stratification with immense convenience by automatically quantifying CACS in CT calcium score imaging.”
Wang and co-authors retrospectively analyzed CT data from 530 patients who underwent CACS scans, while a radiologist manually quantified CACS using Agatston, mass (both commonly used measures) and volume scoring (found to be highly reproducible).
Meanwhile, data from 300 of those patients was used to train the deep learning algorithm. It was validated on a subset of 90 participants and tested on a new set of data from 140 patients.
After comparing measurements derived from the algorithm to manual scoring, the researchers reported no differences. Agatston score categories and cardiovascular risk stratification were very similar, they explained.
The authors noted that while total calcium score was calculated for each patient, the scores of various coronary artery branches were highly variable—an important limitation, they explained. Despite this, they believe, deep learning could offer a “low-cost” and “labor-effective” strategy to one day automatically assess a patient’s risk for cardiovascular disease.
“The present study constitutes the first attempt to evaluate the use of AI in CACS quantification from CT calcium score imaging,” the group concluded. “Although larger-scaled studies are required to further refine the approach, the results indicate that the technique brings automated CACS quantification one step closer to clinical translation.”