Collateral Circulation Classification Based on Cone Beam Computed Tomography Images using ResNet18 Convolutional Neural Network

Abstract

Collateral circulation is an arterial anastomotic channel that supply nutrient perfusion to areas of the brain. It happens when there is an existence of disruption of regular sources of flow due to an ischemic stroke. The most recent method, Cone Beam Computed Tomography (CBCT) neuroimaging is able to provide specific details regarding the extent and adequacy of collaterals. The current approaches for collateral circulation classification are based on manual observation and lead to inter and intra-rater inconsistency. This paper presented a 2-class automatic classification that is recently growing very fast in artificial intelligence disciplines. The two classes will differentiate between good and poor collateral circulation. A pre-trained convolutional neural network (CNN), namely ResNet18, has been used to learn features and train using 4368 CBCT images. Initially, the dataset is prepared, labeled and augmented. Then the images were transferred to be trained using the ResNet18 method with certain specifications. The algorithm performance was then evaluated using metrics in terms of accuracy, sensitivity, specificity, F1 score and precision on the CBCT images to classify collateral circulation accurately. The findings can automate collateral circulation classification to ease the limitations of standard clinical practice. It is a convincing method that supports neuroradiologists in assessing clinical scans and helps neuroradiologists in clinical decisions about stroke treatment.