Computational Image Analysis of Nonenhanced Computed Tomography for Acute Ischaemic Stroke: A Systematic Review

Background: Noncontrast enhanced computed tomography (NCCT) remains the most common method for brain imaging patients who present acutely with ischaemic stroke. Computational methods may improve NCCT analysis in this context. We systematically reviewed current research. Methods: We searched 7 medical and computer engineering databases for studies testing computational methods for analysing NCCT in acute ischaemic stroke. Two independent reviewers extracted the following data; computational method, imaging features investigated, test dataset, ground truth comparison, and performance. We critically evaluated studies for risk of bias and applicability using the Quality Assessment of Diagnostic Accuracy Studies tool (QUADAS-2). Results: From 11,235 nonduplicated articles, we full-text reviewed 200 and selected 68 for inclusion. We identified three dominant study types testing a large range of computational methods for: (1) identifying acute ischaemic stroke (n = 42); (2) ischaemic lesion segmentation (n = 6); and (3) automated Alberta Stroke Program Early CT Score (n = 20). Most articles presented small test datasets, poorly documented patient populations, and did not specify the acuity of the CT scans used in development. There was limited validation or clinical testing of computational methods. Automated Alberta Stroke Program Early CT Score methods were the only software systems presented in multiple publications. Critical evaluation was often limited by lack of data. Conclusions: Computational techniques for analysing NCCT in patients with acute ischaemic stroke have not been adequately clinically validated. Further research with larger and more relevant datasets, in addition to better collaboration between clinicians and researchers, is needed to aid more widespread clinical adoption and implementation.