PT - JOURNAL ARTICLE
AU - Chiu, Chuang-Hsin
AU - Cheng, Cheng-Yi
AU - Tseng, Ta-Wei
AU - Chen, Ing-jou
AU - Yin, You-Ping
TI - Deep Learning Analysis of Tl-201 Myocardial Perfusion Imaging for Improve Diagnostic Accuracy
DP - 2019 May 01
TA - Journal of Nuclear Medicine
PG - 169--169
VI - 60
IP - supplement 1
4099 - http://jnm.snmjournals.org/content/60/supplement_1/169.short
4100 - http://jnm.snmjournals.org/content/60/supplement_1/169.full
SO - J Nucl Med2019 May 01; 60
AB - 169Objectives: Tl-201 myocardial perfusion imaging (MPI) is widely used to diagnose coronary artery disease (CAD). Deep learning (DL) would improve diagnostic accuracy of CAD. The aim of this study was to predict CAD using Tl-201 MPI imaging combined with Deep learning. Methods: A total of 150 patients underwent clinically stress/rest Tl-201 perfusion scans. All patients underwent coronary angiography (CAG) within 6 months of MPI. Obstructive disease was defined asâ‰§70% stenosis of coronary artery. Myocardial perfusion defect scores were automated analysis using Emory Cardiac Toolbox. Scores were defined as normal=0, equivocal=1, moderate reduction=2, severe reduction=3, and absent=4 in 17-segment model. Summed stress/difference score (SSS/SDS) were calculate and compared with raw polar map combined with deep learning. Deep learning was trained using raw stress polar map (S+DL) and combine stress and rest polar map (S+R+DL) for prediction. Deep learning involved automatic variable selection, training and modeling. We divided the data into two parts, using 60% as the training set to build the model, and using 40% as the predictive test set. In the modeling of deep neural networks, grid search was used for hyperparameter adjustment, allowing us to test various parameter combinations of hyperparameters and find more accurate models without overfitting. Results: The area under the curve (AUC) of S+DL was higher than SSS (AUC= 0.91 vs. 0.77, p<0.001). AUC of S+R+DL was higher than SDS (AUC= 0.90 vs. 0.72, p<0.001). Preliminary results show that polar map combine with DL are highly improved diagnostic accuracy. Conclusions: Tl-201 MPI imaging combined with deep learning has potentially improve diagnostic accuracy compared with current clinical methods.Figure1: ROC curve (per-patient) for prediction from (A) sum stress score (SSS) and stress polar map combined with deep learning (S+DL), and (B) sum difference score (SDS) and stress/rest polar map combined with deep learning (S+R+DL).