Development of a deep learning model to identify hyperdense MCA sign in patients with acute ischemic stroke

Abstract

Purpose The aim of this study was to develop an interactive deep learning-assisted identification of the hyperdense middle cerebral artery (MCA) sign (HMCAS) on non-contrast computed tomography (CT) among patients with acute ischemic stroke. Materials and methods 35 HMCAS-positive and 39 HMCAS-negative samples extracted by 50-pixel-diameter circular regions of interest were obtained as training and validation datasets according to the consensus decisions of two experienced neuroradiologists. Data augmentation was performed to increase the number of training samples. A deep convolutional neural network (DCNN) (Xception) was used to classify input images as HMCAS-positive or -negative. Leave-one-case-out cross-validation was achieved to estimate sensitivity, specificity, and accuracy of the deep learning-based training model for identifying HMCAS. Results In terms of diagnostic performance, DCNN for HMCAS offered 82.9% sensitivity, 89.7% specificity, and 86.5% accuracy in leave-one-case-out cross-validation. Area under the receiver operating characteristic curve for HMCAS was 0.947 (95% confidence interval 0.895–0.998; P\u2009 <\u20090.05). Conclusion The deep learning method appears potentially beneficial for identifying HMCAS on non-contrast CT in patients with acute ischemic stroke.