NEURAL NETWORK-BASED SYSTEM of ACUTE STROKE NON-CONTRAST COMPUTED TOMOGRAPHY DIAGNOSTICS: A COMPARATIVE STUDY

Abstract

urpose. To evaluate the accuracy, sensitivity and specificity of acute stroke diagnostics with neural network-based system of non-contrast enhanced computed tomography analysis in comparison with CT-diagnostics specialists. Material and methods. A set of CT-images “without pathology” (absence of acute/subacute ischemic stroke or intracranial hemorrhage) and cases of ischemic stroke and intracranial hemorrhage (n=90) were selected and validated by three experts of CT-diagnostics in E.N. Meshalkin National medical research center. The selected anonymized cases were retrospectively assessed by four independent CT-diagnostics specialists and neural network-based system of acute stroke non-contrast computed tomography diagnostics. Results. The accuracy of acute stroke diagnostics by the neural network-based system (91.1%) didn’t statistically differ from all four CT-diagnostics specialists (91.1%, 94.4%, 88.9%, and 95.6%). During the ROC-analysis, the high level of sensitivity (98.1%) and specificity (80.6%,) as well as the AUC = 0.894 (p<0.001), were demonstrated. Discussion. According to current AHA/ASA guidelines for the early management of patients with acute ischemic stroke, computed tomography is a basic tool in acute stroke diagnostics. This article presents the first performance of newly developed neural network-based system of acute stroke non-contrast computed tomography diagnostics in comparison with CT-diagnostics specialists. Conclusion. The developed neural network-based system has demonstrated the comparable accuracy of acute stroke diagnostics. In terms of experiment, system showed high levels of acute stroke diagnostics sensitivity and specificity, however, further development and education of neural network to provide higher accuracy and better differential diagnostic are required.