Abstract WP81: Initial Experiences With Artificial Neural Networks in Detection of CT Perfusion Deficits

Abstract

Background: Head CT with perfusion imaging has become crucial in the selection of patients for treatment for mechanical thrombectomy. In recent years machine learning has rapidly evolved and found applications in a wide variety of health care tasks. Objective: We report our initial experiences with training a neural network pipeline to predict the presence and sidedness of a perfusion deficit in patients with acute ischemic stroke, based on imaging alone. Methods: CT perfusion images of patients with suspicion for acute ischemic stroke were retrospectively obtained. Each study was preprocessed into 3 dimensional arrays of Hounsfeld Units stacked through time, segmented to exclude bone, and then normalized. Labels were derived from the attending neuroradiologists’ read. The data was split into training and validation sets. A long term, recurrent convolutional (LRCN) network was constructed consisting of a convolutional neural network stacked on top of a long short term (LSTM) layer. Results: 139 (35.1%) patients had a right sided perfusion deficit, while 199 (50.3%) had a left sided deficit, and 58 (14.6%) had no evidence of a deficit. The best model was able to achieve an accuracy of 85.8% on validation data. Receiver Operating Characteristic (ROC) curves were generated for each class and an Area Under the Curve (AUC) was calculated for each class. For a right sided deficit, the AUC was 0.90, for left sided deficit 0.96, and for no deficit the AUC was 0.93. Conclusion: The field of machine learning, powered by convolutional neural networks for the task of image recognition and processing, has quickly developed in recent years. We have constructed an artificial neural network that can identify and classify the presence and sidedness of a perfusion deficit CT perfusion imaging.