Fast automated artificial intelligence evaluation system for acute ischemic stroke

Abstract

Acute ischemic stroke (AIS) is a leading cause of morbidity and mortality worldwide, and it is usually due to a focal interruption of cerebral blood flow caused by occlusion of a cerebral artery. Large vessel occlusions (LVO) cause approximately one-third of acutely presenting AIS but are responsible for three-fifths of dependency and more than nine-tenths of mortality after AIS. Non-contrast CT (NCCT) is the first-line imaging modality for AIS due to its widespread availability, speed of imaging, and low cost. The hyperdense middle cerebral artery sign (HMCAS) representing thromboembolism has been declared as a vital CT finding for intravascular thrombus in the diagnosis of large vessel occlusion. Fast diagnosis of LVO and early recognition of HMCAS can assist in patient triage and subsequent thrombolysis or thrombectomy treatment. Studies in this thesis were data-driven tasks conducted based on two phases of data that were retrospectively collected from all recorded AIS patients within the Hong Kong Hospital Authority’s clinical management system. The first phase contains 300 subjects while the second phase involves additional 324 cases. Based on the first phase of data, we established an automated evaluation system to enable the rapid identification of LVO. The evaluation system contains three hierarchical models implemented by multiple machine learning techniques. The first two levels of modeling utilized structured demographic and clinical data, while the third level model involved additional NCCT image features obtained from deep learning models. Besides, we also presented an approach to segment the HMCAS by using the U-Net architecture plus special designed Tversky loss and negative mining technique in order to overcome the data imbalance problem. Based on our previous work, the other two end-to-end deep learning models were proposed with extra data collected in the second phase. We presented a Bilateral Comparison Network (BCNet) in order to have a reliable classification of LVO and a Dissimilar-Siamese-U-Net (DSU-Net) to segment the HMCAS. Both BCNet and DSU-Net fully explore the feature representation of the discrepancies between the bilateral hemispheres, which is a rule-based knowledge commonly considered by radiologists. Compared with models developed in the first phase of the study, the novel BCNet and DSU-Net witness obvious improvement in the evaluation performance for their respective tasks. Overall, we proposed several approaches capable of fast and reliable evaluations of large vessel occlusion utilizing both structural clinical data and non-structural image data. In the meanwhile, we also presented automated approaches to segment the HMCAS. The NCCT data involved in this study were acquired through a variety of CT scanners from multiple clinical institutions, demonstrating the proposed algorithm can endure the variation of image characters thereby further suggesting its possibility of broad clinical applicability. These findings are important and can support the clinical application of the algorithm as a diagnostic adjunct in the acute stroke pathway, especially in resource-limited settings when immediate expert neuroradiological interpretation is not readily available.