Using deep learning to analyze upright and supine SPECT myocardial perfusion images (MPI) could improve the diagnosis of obstructive coronary artery disease, according to a multicenter international study published in the May issue of the Journal of Nuclear Medicine.
Currently, the method of analyzing MPI data quantitatively is to calculate the combined total perfusion deficit from the semi-upright and supine SPECT MPI positions. Clinicians then visually evaluate the results of the two perspectives. By contrast, deep-learning convolutional neural networks (CNNs) are trained on the visual data, learn from that information, and create findings based on the analyses.
For this study, deep-learning CNN analysis of two-position stress MPI scans was compared with the standard total perfusion deficit method on 1,160 patients with no known evidence of coronary artery disease. Patients underwent stress MPI with technetium-99m sestamibi on solid-state SPECT cameras at four different centers. All patients also had onsite clinical reads and invasive coronary angiography correlations within six months of MPI (JNM, May 2019, Vol. 60:5, pp. 664-670).
The researchers trained four different deep-learning models using data from three centers before testing the approach at the fourth facility. Predictions for all four centers were merged to create an overall estimation of the multicenter performance.
The findings show that 718 patients (62%) and 1,272 arteries (37%) exhibited obstructive disease. On a per-patient basis, sensitivity improved to 66% with deep learning, compared with 62% with the total perfusion deficit method. The per-vessel analysis showed sensitivity of 59% with deep learning, compared with 55% through the total perfusion deficit approach.
