D-SPECT combined with deep learning predicts obstructive coronary artery disease.
Authors
Affiliations (4)
Affiliations (4)
- Department of Nuclear Medicine, First Hospital of Shanxi Medical University, Taiyuan, Shanxi, China.
- Shanxi Key Laboratory of Molecular Imaging, Shanxi Medical University, Taiyuan, Shanxi, China.
- Collaborative Innovation Center for Molecular Imaging of Precision Medicine, Shanxi Medical University, Taiyuan, Shanxi, China.
- Nuclear Medicine Department, Taiyuan People's Hospital, Taiyuan, Shanxi, China.
Abstract
Deep learning models trained on dynamic single-photon emission computed tomography myocardial perfusion imaging (D-SPECT MPI) data may enhance the predictive capability of D-SPECT MPI images for obstructive CAD (OCAD). This study aimed to evaluate the predictive performance of deep learning algorithm based on D-SPECT MPI for OCAD. A total of 92 patients with suspected coronary artery disease were ultimately included in the study. They underwent D-SPECT MPI and coronary angiography within 6 months. The deep learning OCAD prediction model (DL-OCAD) was trained using perfusion and motion images, and its predictive capability for obstructive stenosis was evaluated through a stratified 5-fold cross-validation process. The comparison of the predictive capabilities for OCAD among DL-OCAD, coronary flow reserve (CFR), stress myocardial blood flow (sMBF), stress total perfusion deficit (sTPD), and summed difference score (SSS). The total of 56 patients (61%) had OCAD, and obstructive lesions were present in 160 of 276 arteries (58%). The overall diagnostic performance of DL-OCAD (AUC, 0.85; 95% CI, 0.75-0.95) was higher than that of sTPD (AUC, 0.64; 95% CI, 0.50-0.78, <i>P</i> = 0.044) and SSS (AUC, 0.63; 95% CI, 0.49-0.78, <i>P</i> = 0.028). The sensitivity of DL-OCAD was significantly higher than that of sTPD (81.3% vs. 58.2%, <i>p</i> = 0.002) and SSS (81.3% vs. 57.3%, <i>P</i> = 0.001). The calibration curves of DL-OCAD, CFR, sMBF, sTPD and SSS models were generally in good agreement with the ideal diagonal. Within the clinically meaningful threshold probability range of 0.05 to 0.5, the DL-OCAD, CFR, sMBF, sTPD, and SSS prediction model can obtain a positive net clinical benefit compared with the strategy of not further examining all patients, and the DL-OCAD curve was always better than the CFR, sMBF, sTPD, and SSS curves. Compared with existing clinical methods, deep learning has potential clinical value in improving the predictive ability of D-SPECT MPI for OCAD. However, its clinical value still needs to be validated in external studies involving multiple centers and large samples.