TY - JOUR T1 - ­­­­Improved risk assessmentof myocardial SPECT using deep learning : report from REFINE SPECT registry JF - Journal of Nuclear Medicine JO - J Nucl Med SP - 50 LP - 50 VL - 62 IS - supplement 1 AU - Ananya Singh AU - Robert Miller AU - Yuka Otaki AU - Paul Kavanagh AU - Serge Van Kriekinge AU - Wei Chih-Chun AU - Tejas Parekh AU - Tali Sharir AU - Andrew Einstein AU - Mathews Fish AU - Terrence Ruddy AU - Philipp Kaufmann AU - Albert Sinusas AU - Edward Miller AU - Timothy Bateman AU - Sharmila Dorbala AU - Marcelo Di Carli AU - Joanna Liang AU - Damini Dey AU - Daniel Berman AU - Piotr Slomka Y1 - 2021/05/01 UR - http://jnm.snmjournals.org/content/62/supplement_1/50.abstract N2 - 50Objectives: We sought to develop and evaluate a novel deep-learning network for the prediction of major adverse cardiac events (MACE) following single photon emission tomography (SPECT) myocardial perfusion imaging (MPI). To allow per-patient visual interpretation of the risk score by physicians, we created a mechanism to identify polar map image regions contributing to the risk estimation. Methods: Patients from the REgistry of Fast myocardial perfusion Imaging with NExt generation SPECT (REFINE SPECT) undergoing SPECT MPI at one of the 5 sites were included. Stress and ischemic TPD (difference of stress and rest TPD) were quantified automatically. The deep-learning network (MACE-DL) was developed with polar map image inputs of raw perfusion and gated derived maps of motion, thickening, phase angle and amplitude combined with age, sex, end-systolic, and end-diastolic volumes. Patients were randomly divided into 10 folds with repeated training, validation, and testing sets, using an 80%-10%-10% stratified split. To prevent overfitting due to class imbalance (low ratio of cases with MACE), the network was trained with polar map images of patients with MACE which were randomly rotated within a range (-10,10 degrees). The proposed network consisted of 2 convolution blocks, each with 3x3 convolution kernels, batch normalization, dropout and Leaky Rectified Linear Units layers, which were added to prevent overfitting. Age, sex, and cardiac volumes were introduced to the model in the first fully connected layer. The output of MACE-DL was the likelihood of a patient experiencing MACE during follow-up. An attention map, which highlights polar map image regions of importance for MACE-DL prediction, was generated to explain results to the physician. The prognostic accuracy of MACE-DL and quantitative analysis of perfusion were evaluated using area under receiver-operating characteristic curve (AUC) and compared using DeLong’s method. Results: In total, 20,401 patients were included with mean age 64 ± 12 years and 11,630 (57%) males. During median follow-up of 4.4 years (interquartile range, 3.4-5.7years), 3,538 patients experienced at least one MACE. The AUC [95% CI] for prediction of MACE by MACE-DL (0.75[0.74-0.76]) was significantly higher than stress TPD (0.70 [0.69-0.71], p<0.0001) or ischemic TPD (0.68[0.67-0.69]) p<0.0001). Patients in the highest quartile of MACE-DL score (≥0.50) had an annual MACE rate of 9.7% (95% CI:9.2-10.1), with a 10.2-fold increased risk compared to patients in the lowest quartile (DL score <0.13) who had an annual MACE rate of 0.9% (95% CI:0.8-1.0). The attention map highlighted regions that were associated with MACE risk as an overlay on polar map images, along with a continuous MACE probability with results generated in less than 1 second during the testing stage. In the 3,538 patients with at least one MACE, the output attention maps on an average covered 46.1%, 10.6% and 43.2% of the left anterior descending artery (LAD), left circumflex artery (LCx) and right coronary artery (RCA) pre-definied territories, respectively. Conclusions: Use of a deep learning network allows for prediction of MACE directly from polar map images with improved accuracy compared to automatic quantitation of perfusion. MACE-DL also incorporates a mechanism to explain to the physician, which polar map image regions contribute most to MACE prediction. ER -