Deep learning for vessel occlusion classification using CT perfusion maps in acute ischemic stroke

Odysseas Papakyriakou, Charles B. Majoie, Henk Marquering

Keywords: CT perfusion, deep learning, stroke, occlusion location classification

TL;DR: We propose a 3D attention-based convolutional neural network for vessel occlusion classification using CTP maps in acute ischemic stroke.

Abstract: The assessment of the occlusion location in acute ischemic stroke is an essential step in treatment decisions. Occlusions in smaller vessels, like the M2, can be difficult to detect in routine clinical care. Computed tomography perfusion (CTP) maps improve clinicians’ accuracy and speed in locating occlusions compared to CT angiography alone. Deep learning (DL) could help automate this process. We propose an attention-based convolutional neural network to classify ICA-T, M1, and M2 occlusions using CTP maps. Our method shows an average accuracy of 79.2% and an F1 score of 86% for M2 occlusions, demonstrating the potential of DL utilizing CTP maps for occlusion location classification.

Submission Number: 172