Transferring Deep Convolutional Network Representations from SPECT to Improve PET Cardiac Outcome Prediction

Purpose Cardiac PET, less common than SPECT, is rapidly growing and offers the additional benefit of first-pass absolute myocardial blood flow measurements. However, multicenter cardiac PET databases are not well established. We used multicenter SPECT data to improve PET cardiac risk stratification via a deep learning knowledge transfer mechanism. Methods SPECT and PET patients are followed-up for major adverse cardiac events (MACE): all-cause death, unstable angina, myocardial infarction, or revascularization. SPECT cohort: 20,401 cases (5 sites; 57% male). PET cohort: single-site, split temporally for internal 5-fold cross-validation (n=2,231; 56% male) and simulated prospective testing (n=2,056; 58% male). A convolutional neural network is pretrained on SPECT polar map images to estimate MACE risk. Knowledge transfer is achieved by leveraging trained SPECT model parameters for PET models. Unique PET imaging parameters (e.g. myocardial blood flow reserve) are integrated using ensemble fully connected networks. Methodologies of increasing complexity: PET (vanilla), PET (transfer), PET (transfer+freeze) and clinical baselines: summed stress score (SSS), summed rest score (SRS), summed difference score (SDS), were compared using area under the receiver operating characteristic curve. Results MACE prediction performance in the internal set with transfer learning: PET (transfer+freeze) (0.744) and PET (transfer) (0.743) were significantly higher than PET (vanilla) (0.727), SSS (0.645), SRS (0.620), and SDS (0.607) (all p<0.001). In hold-out testing, PET (transfer+freeze) (0.741) and PET (transfer) (0.732) improved over PET (vanilla) (0.722) (p<0.001, p<0.05), SSS (0.658), SRS (0.612), and SDS (0.613) (all p<0.001). Conclusion We demonstrate a novel methodology for improving cardiac PET MACE prediction via transfer learning from SPECT.