Diagnosing coronary artery disease, the most common type of heart disease in the U.S., can be improved by AI, according to a new, multicenter international study published in The Journal of Nuclear Medicine.
Currently, coronary artery disease, or CAD, is often diagnosed by upright and supine single photon emissions computed tomography (SPECT) myocardial perfusion imaging (MPI), which reveal a heart’s ability to pump and examine blood flow through the heart during exercise and at rest. The images are routinely taken in two positions––semi-upright and supine, which are both used to calculate the combined total perfusion deficit (TPD) to analyze MPI data.
For the study, researchers from Cedars-Sinai Medical Center in Los Angeles, compared the standard TPD analysis of 1,160 patients without known CAD to a deep learning analysis of data from the two-position stress MPI. To collect the images, new-generation solid-state SPECT scanners were used in four different centers as patients underwent stress MPIs and had on-site clinical reads and invasive coronary angiography correlation within six months of MPI.
Of the more than 1,100 patients, 62%, or 712, and 37%, or 1,272, of 3,480 arteries had obstructive disease––defined as at least 70% narrowing of the three major coronary arteries and at least 50% for the left main coronary artery.
Between the two methods of diagnosis, deep learning was more sensitive than TPD for predicting obstructive disease and improved current quantitative methods, researchers found.
"These findings were demonstrated for the first time in a rigorous, repeated external validation," Piotr J. Slomka, PhD, at Cedars-Sinai Medical Center, said in a statement, affirming that "the latest developments in artificial intelligence can be efficiently leveraged to enhance the accuracy of existing nuclear medicine techniques."
The sensitivity of the diagnosis improved per-patient from 61.8% with TPD to 65.6% with deep learning. Per vessel, sensitivity improved from 54.6% with TPD to 59.1% with deep learning. For on-site clinical read, deep learning had a sensitivity of 84.8% compared to 82.6% for TPD.
According to Slomka, the method can be “easily and immediately deployed clinically,” he wrote.