Dutch researchers have found that a deep-learning algorithm refines cardiac CT results by automatically eliminating exams negative for coronary artery calcium (CAC) -- streamlining radiology workflow, according to a study published on 3 June in the European Journal of Radiology.
"[Our study showed that] scans without calcium were safely excluded without false-negative predictions," a group led by Dr. Leonardus van den Oever of the University Medical Center Groningen in the Netherlands reported. "This algorithm can reduce the radiologist's workload by filtering out normal scans."
Prior research has demonstrated that a patient's CAC score is an effective predictor of future heart disease and that early detection of disease via CAC scoring could reduce the number of deaths. But having to read exams that turn out not to have CAC can slow the workflow in a busy radiology department. That's where deep learning can help, according to van den Oever and colleagues.
"Screening programs [for cardiovascular disease] would add a large number of additional scans to be seen by radiologists, due to the large number of eligible participants," the team wrote. "Automatically excluding participants without coronary artery calcium from the workflow would result in an enormous reduction of the total screening workload."
Van den Oever's group explored the feasibility of using a deep-learning algorithm to do just that. The team collected data from the ROBINSCA (Risk Or Benefit IN Screening for CArdiovascular diseases) study, a randomized CT screening trial, and the ImaLife (Imaging in Lifelines) study, which investigates imaging biomarkers for lung cancer, chronic obstructive pulmonary disease, and coronary artery disease.
The investigators used 60 CT scans taken from the ROBINSCA study to train the deep-learning algorithm. For internal validation, they used 100 scans, also taken from that study (50 of participants without CAC and 50 of participants with CAC); for external validation of the algorithm, the team used data from the ImaLife study (also 100 scans, divided into two groups, one with CAC and one without).
The deep-learning algorithm correctly classified almost two-thirds of negative CAC exams in the internal validation set and more than three-quarters in the external validation set. The researchers found no false-negative results.
| Deep-learning algorithm performance identifying CAC, internal and external validation sets | ||
| Internal validation dataset | External validation dataset | |
| Study participants with CAC | 100% | 96% |
| Study participants without CAC | 62% | 86% |
The study results are good news for busy radiology practices, according to the researchers.
"Assuming a prevalence for CAC of 60% in a screening population of elevated risk, deploying our model would allow for a direct CAC negative classification of 34 out of 100 scans," they concluded. "That implies a 34% reduction in the number of scans due for manual evaluation, and represents a considerable reduction in radiologists' workload in such a screening setting."
