Artificial neural network retrained to detect myocardial ischemia using a Japanese multicenter database
- PDF / 2,774,083 Bytes
- 8 Pages / 595.276 x 790.866 pts Page_size
- 76 Downloads / 138 Views
ORIGINAL ARTICLE
Artificial neural network retrained to detect myocardial ischemia using a Japanese multicenter database Kenichi Nakajima1 · Koichi Okuda2 · Satoru Watanabe1 · Shinro Matsuo1 · Seigo Kinuya1 · Karin Toth3 · Lars Edenbrandt4 Received: 12 February 2018 / Accepted: 4 March 2018 © The Author(s) 2018. This article is an open access publication
Abstract Purpose An artificial neural network (ANN) has been applied to detect myocardial perfusion defects and ischemia. The present study compares the diagnostic accuracy of a more recent ANN version (1.1) with the initial version 1.0. Methods We examined 106 patients (age, 77 ± 10 years) with coronary angiographic findings, comprising multi-vessel disease (≥ 50% stenosis) (52%) or old myocardial infarction (27%), or who had undergone coronary revascularization (30%). The ANN versions 1.0 and 1.1 were trained in Sweden (n = 1051) and Japan (n = 1001), respectively, using 99mTc-methoxyisobutylisonitrile myocardial perfusion images. The ANN probabilities (from 0.0 to 1.0) of stress defects and ischemia were calculated in candidate regions of abnormalities. The diagnostic accuracy was compared using receiver-operating characteristics (ROC) analysis and the calculated area under the ROC curve (AUC) using expert interpretation as the gold standard. Results Although the AUC for stress defects was 0.95 and 0.93 (p = 0.27) for versions 1.1 and 1.0, respectively, that for detecting ischemia was significantly improved in version 1.1 (p = 0.0055): AUC 0.96 for version 1.1 (sensitivity 87%, specificity 96%) vs. 0.89 for version 1.0 (sensitivity 78%, specificity 97%). The improvement in the AUC shown by version 1.1 was also significant for patients with neither coronary revascularization nor old myocardial infarction (p = 0.0093): AUC = 0.98 for version 1.1 (sensitivity 88%, specificity 100%) and 0.88 for version 1.0 (sensitivity 76%, specificity 100%). Intermediate ANN probability between 0.1 and 0.7 was more often calculated by version 1.1 compared with version 1.0, which contributed to the improved diagnostic accuracy. The diagnostic accuracy of the new version was also improved in patients with either single-vessel disease or no stenosis (n = 47; AUC, 0.81 vs. 0.66 vs. p = 0.0060) when coronary stenosis was used as a gold standard. Conclusion The diagnostic ability of the ANN version 1.1 was improved by retraining using the Japanese database, particularly for identifying ischemia. Keywords Nuclear cardiology · Artificial intelligence · Myocardial perfusion imaging · Coronary artery disease Abbreviations ANN Artificial neural network AUC Area under the curve MIBI Methoxyisobutylisonitrile
* Kenichi Nakajima [email protected]‑u.ac.jp 1
Department of Nuclear Medicine, Kanazawa University Hospital, 13‑1 Takara‑machi, Kanazawa 920‑8641, Japan
2
Department of Physics, Kanazawa Medical University, Uchinada, Kahoku, Japan
3
EXINI Diagnostics, Lund, Sweden
4
Department of Clinical Physiology and Nuclear Medicine, University of Gothenburg, Gothenburg,
Data Loading...
