Predictive Deep Learning Model for Neural Vessel Occlusion

Abstract

Recent advances in medical applications for clot removal have enabled physicians to unblock blood vessel occlusions that occur in extremely narrow regions of the brain along the Horizontal (M1) segment of the middle cerebral artery. FDA approved, clinical trials have proven the significant cerebral recovery that can be achieved for victims of ischemic stroke when they undergo swift surgical intervention for clot removal. That said, carrying out such a delicate operation simultaneously requires extreme surgical expertise, and the ability to quickly identify the site of occlusion from 2D x-ray images of the brain. Fortunately, recent efforts in object detection and classification within the fields of deep learning and computer vision have dramatically improved the predictive power of the computer. Accordingly, the goal of this research project is to develop a deep learning model capable of accurately predicting the site of occlusion from frontal view cerebral angiograms in real time. Our current model utilizes YOLOv3 architecture and identifies the site of occlusion in 94.4% of all cases given a minimum 25% confidence threshold. Furthermore, in 83.97% of cases, the occlusion region is detected with at least 50% average intersection over union between the predicted region and the ground truth region. Finally, distributed over the entire validation set, the average intersection over union between the predicted region and the ground truth region was 74.29%.