Hui Liu1,2,3,4, Jing Wu5,6, Edward J Miller7,5,8, Chi Liu5, Yi-Hwa Liu9,10,11,12. 1. Department of Internal Medicine, Yale University School of Medicine, New Haven, CT, 06520, USA. liuhui2020@tsinghua.edu.cn. 2. Department of Engineering Physics, Tsinghua University, Beijing, People's Republic of China. liuhui2020@tsinghua.edu.cn. 3. Key Laboratory of Particle & Radiation Imaging (Tsinghua University), Ministry of Education, Beijing, People's Republic of China. liuhui2020@tsinghua.edu.cn. 4. Department of Radiology and Biomedical Imaging, Yale University, New Haven, CT, USA. liuhui2020@tsinghua.edu.cn. 5. Department of Radiology and Biomedical Imaging, Yale University, New Haven, CT, USA. 6. Center for Advanced Quantum Studies and Department of Physics, Beijing Normal University, Beijing, CT, People's Republic of China. 7. Department of Internal Medicine, Yale University School of Medicine, New Haven, CT, 06520, USA. 8. Department of Nuclear Cardiology, Heart and Vascular Center, Yale New Haven Hospital, New Haven, CT, USA. 9. Department of Internal Medicine, Yale University School of Medicine, New Haven, CT, 06520, USA. yi-hwa.liu@yale.edu. 10. Department of Nuclear Cardiology, Heart and Vascular Center, Yale New Haven Hospital, New Haven, CT, USA. yi-hwa.liu@yale.edu. 11. Department of Biomedical Engineering, Chung Yuan Christian University, Taoyuan, Taiwan. yi-hwa.liu@yale.edu. 12. Department of Biomedical Imaging and Radiological Sciences, National Yang-Ming University, Taipei, Taiwan. yi-hwa.liu@yale.edu.
Abstract
PURPOSE: Deep convolutional neural networks (CNN) for single photon emission computed tomography (SPECT) myocardial perfusion imaging (MPI) has been used to improve the diagnostic accuracy of coronary artery disease (CAD). This study was to design and evaluate a deep learning (DL) approach to automatic diagnosis of myocardial perfusion abnormalities from stress-only MPI. METHODS: The new DL approach developed for this study was compared to a conventional quantitative perfusion defect size (DS) method. A total of 37,243 patients (51.5% males) undergone stress 99mTc-Tetrofosmin or 99mTc-Sestamibi MPI were selected retrospectively from Yale New Haven Hospital. Patients were dichotomized as studies with normal (75.4%) or abnormal (24.6%) myocardial perfusion based on final diagnoses of clinical nuclear cardiologists. Stress myocardial perfusion defect size was calculated using Yale quantitative analytic software. A deep CNN was trained using the circumferential count profile maps derived from SPECT MPI and was evaluated for the diagnosis of perfusion abnormality with a 5-fold cross-validation approach. In each fold, 27,933, 1862 and 7448 patients were used as training, validation and testing datasets, respectively. The area under the receiver-operating characteristic curve (AUC) was calculated and analyzed for all patients as well as for the eight sub-groups classified based on patient genders, quantitative algorithms, radioactive tracers and SPECT cameras. RESULTS: The AUC value resulted from the DL method was significantly higher than that from the DS method (0.872 ± 0.002 vs. 0.838 ± 0.003, p < 0.01). Across the eight sub-groups, the DL method provided more consistent AUC values in terms of smaller standard deviation and higher diagnostic accuracy and specificity, but slightly lower sensitivity than the DS method (AUC: 0.865 ± 0.010 vs. 0.838 ± 0.019, Accuracy: 82.7% ± 2.5% vs. 78.5% ± 3.6%, Specificity: 84.9% ± 3.7% vs. 77.5% ± 6.5%, Sensitivity: 74.4% ± 4.2% vs. 79.8% ± 5.8%). CONCLUSIONS: The incorporation of deep learning for stress-only MPI has a considerable potential to improve the diagnostic accuracy and consistency in the detection of myocardial perfusion abnormalities.
PURPOSE: Deep convolutional neural networks (CNN) for single photon emission computed tomography (SPECT) myocardial perfusion imaging (MPI) has been used to improve the diagnostic accuracy of coronary artery disease (CAD). This study was to design and evaluate a deep learning (DL) approach to automatic diagnosis of myocardial perfusion abnormalities from stress-only MPI. METHODS: The new DL approach developed for this study was compared to a conventional quantitative perfusion defect size (DS) method. A total of 37,243 patients (51.5% males) undergone stress99mTc-Tetrofosmin or 99mTc-Sestamibi MPI were selected retrospectively from Yale New Haven Hospital. Patients were dichotomized as studies with normal (75.4%) or abnormal (24.6%) myocardial perfusion based on final diagnoses of clinical nuclear cardiologists. Stress myocardial perfusion defect size was calculated using Yale quantitative analytic software. A deep CNN was trained using the circumferential count profile maps derived from SPECT MPI and was evaluated for the diagnosis of perfusion abnormality with a 5-fold cross-validation approach. In each fold, 27,933, 1862 and 7448 patients were used as training, validation and testing datasets, respectively. The area under the receiver-operating characteristic curve (AUC) was calculated and analyzed for all patients as well as for the eight sub-groups classified based on patient genders, quantitative algorithms, radioactive tracers and SPECT cameras. RESULTS: The AUC value resulted from the DL method was significantly higher than that from the DS method (0.872 ± 0.002 vs. 0.838 ± 0.003, p < 0.01). Across the eight sub-groups, the DL method provided more consistent AUC values in terms of smaller standard deviation and higher diagnostic accuracy and specificity, but slightly lower sensitivity than the DS method (AUC: 0.865 ± 0.010 vs. 0.838 ± 0.019, Accuracy: 82.7% ± 2.5% vs. 78.5% ± 3.6%, Specificity: 84.9% ± 3.7% vs. 77.5% ± 6.5%, Sensitivity: 74.4% ± 4.2% vs. 79.8% ± 5.8%). CONCLUSIONS: The incorporation of deep learning for stress-only MPI has a considerable potential to improve the diagnostic accuracy and consistency in the detection of myocardial perfusion abnormalities.