Effect of deep learning attenuation correction on accuracy of stress myocardial perfusion imaging

Attenuation correction (AC) improves the accuracy of myocardial perfusion imaging for detecting coronary artery disease (CAD). There has been limited application of deep learning attenuation correction (DLAC) in cadmium zinc telluride (CZT) cameras.

We employed DLAC for CZT imaging in a consecutive series of 100 patients with suspected CAD who had myocardial perfusion imaging with a DSPECT camera and coronary angiography. CAD was defined as left main stenosis ≥50% or another stenosis ≥70%. Automated analysis was compared with visual interpretation for the summed stress score (SSS) and was employed before (non-attenuation correction, NAC) and after DLAC. Accuracy was compared across upright, supine, and combined imaging positions using receiver operating characteristic analysis of the SSS, summed difference score (SDS), and a pixel-by-pixel approach.

Automated analysis of SSS was not dissimilar to detailed visual analysis in the first 41 patients and was superior to visual analysis for the entire series. The SSS with NAC was highly accurate in detecting CAD (area under curve (AUC): 0.915) and was similar for SDS. DLAC was less accurate than NAC (P = 0.01 for SSS and 0.0003 for SDS). For SDS, upright imaging was superior to combined images (P = 0.002). The pixel-by-pixel analysis for reversibility correctly classified 93% of patients with NAC and 74% of patients with DLAC.

For detecting CAD with a DSPECT camera, automated analysis for SSS had greater accuracy than visual analysis. While DLAC was effective, non-AC methods proved to be more accurate. Upright imaging was superior to combined imaging for detecting ischemia by SDS, highlighting opportunities to optimize CZT imaging protocols.