Deep learning-based obstructive coronary artery disease prediction from myocardial perfusion SPECT.
Authors
Affiliations (13)
Affiliations (13)
- Biomedical Imaging Laboratory (BIG), Department of Electrical and Computer Engineering, Faculty of Science and Technology, University of Macau, Taipa, Macau SAR, China.
- Center for Cognitive and Brain Sciences, Institute of Collaborative Innovation, University of Macau, Taipa, Macau SAR, China.
- Department of Nuclear Medicine, The First Affiliated Hospital of Army Medical University, Chongqing, China.
- Lauterbur Research Center for Biomedical Imaging, Shenzhen Institute of Advanced Technology, Chinese Academy of Sciences, Shenzhen, China.
- Department of Nuclear Medicine, Show Chwan Memorial Hospital, Changhua, Taiwan.
- Department of Nuclear Medicine and PET/CT-MRI Centre, The First Affiliated Hospital of Jinan University, Guangzhou, Guangdong, China.
- Department of Biomedical Imaging and Radiological Sciences, National Yang Ming Chiao Tung University, Taipei, Taiwan.
- Department of Nuclear Medicine, Taipei Veterans General Hospital, Taipei, Taiwan.
- Department of Nuclear Medicine, Taipei Veterans General Hospital, Taipei, Taiwan. [email protected].
- School of Medicine, National Yang Ming Chiao Tung University, Taipei, Taiwan. [email protected].
- Department of Nuclear Medicine, Chang Bing Show Chwan Memorial Hospital, Changhua, Taiwan. [email protected].
- Biomedical Imaging Laboratory (BIG), Department of Electrical and Computer Engineering, Faculty of Science and Technology, University of Macau, Taipa, Macau SAR, China. [email protected].
- Center for Cognitive and Brain Sciences, Institute of Collaborative Innovation, University of Macau, Taipa, Macau SAR, China. [email protected].
Abstract
We aim to apply the deep learning (DL) technique to predict the gold-standard invasive coronary angiography (ICA) for coronary artery disease (CAD) diagnosis, from non-invasive myocardial perfusion SPECT (MP-SPECT). A total of 515 anonymized patients from 3 clinical centers (primary: 212; external-1:108; external-2:195) underwent standard one-day Tc-99m-sestamibi or Tl-201 stress/rest MP-SPECT protocol were retrospectively recruited. DL models have been proposed for DL-based attenuation correction (DLAC), per-patient and per-vessel obstructive CAD prediction respectively. DL prediction models were trained with no AC (NAC), DLAC and CT-based AC (CTAC) data, as well as stress, combined stress and rest data (stress/rest) and 2-channel stress + rest input. Clinical factors were incorporated into the DL models to improve the prediction outcome. The accuracy and area under the receiver-operating characteristic curve (AUC) were analyzed. For per-patient analysis, the AUC was 0.57 for TPD-based diagnosis, and was 0.59, 0.77, and 0.84 for the stress-only, stress/rest and stress + rest input with CTAC in the primary dataset. The AUC of the CTAC-based stress + rest was further increased to 0.92 by clinical factors of gender, age and hypertension diagnosis. For per-vessel analysis, the AUC with the same clinical factors was 0.80. DLAC has improved AUC for different input as compared to NAC in both primary and external datasets and for both per-patient and per-vessel analysis. The DL-based AC, CTAC, combination of stress and rest polar plots and the incorporation of clinical information can enhance prediction performance of CAD.