A novel algorithm developed using machine learning and a J-ACCESS database can estimate defect scores from myocardial perfusion single-photon emission tomography images

Background Stress myocardial perfusion single-photon emission computed tomography (SPECT) imaging (MPI) has been used to diagnose and predict the prognoses of patients with coronary artery disease (CAD). An ongoing multicenter collaboration established a Japanese database (J-ACCESS) in 2001 that includes a risk model and expert interpretations. The present study aimed to develop a novel algorithm using machine learning (ML) and resources from the J-ACCESS database to aid SPECT image interpretation. Methods We analyzed data from 1288 patients in J-ACCESS 3 and 4 databases. Three-dimensional (3D) stereoscopic images of left ventricular myocardial perfusion were reconstructed with linear transformation from the original short-axis data. Segments were extracted from U-Net, then features were extracted from each segment during the ML process. We estimated segmental scores based on weighted features obtained from fully connected layers. Correlations between segment scores interpreted by nuclear cardiology experts and estimated by ML were evaluated using a 17-segment model, summed stress (SSS), summed rest (SRS), and summed difference (SDS) scores, and ratios (%) of summed different scores (%SDS). Results The complete concordance rate of scores assessed by the experts and estimated by ML was 79.6%. The underestimated and overestimated rates were 10.3% and 10.0%, respectively. Associations between defect scores assessed by experts and ML were close, with correlation coefficients ( r ) of 0.923, 0.917, 0.842 and 0.853 for SSS, SRS, SDS, %SDS, respectively ( p