LAUSR

Management of stroke highly depends on information from imaging studies. Noncontrast computed tomography (CT) and magnetic resonance imaging (MRI) can both be used to distinguish between ischemic and hemorrhagic stroke, which is difficult based on clinical features. Hypodensity on CT and DWI hyperintensity on MRI identifies irreversibly damaged tissue, although the sensitivity of MRI is higher in the acute setting. Angiographic and perfusion imaging sequences can identify a large vessel occlusion and, along with perfusion imaging, can select patients for endovascular therapy. The FLAIR-DWI mismatch yields information about patients with unknown time of onset (including wake-up strokes). Stroke imaging also gives insight into prognosis, with current methods aiming to give a picture of the short-term consequences of successful reperfusion or continued large vessel occlusion. One important caveat about stroke imaging is that it must be done quickly, as faster treatment leads to better outcomes.1 However, most steps in the stroke imaging triage pathway require the presence of human radiologists and neurologists, and this is often the time-limiting step. The expertise required for these tasks may not be available at all sites or at all times. Therefore, there is interest in automated methods for stroke imaging evaluation. Artificial intelligence (AI) is a broad term reflecting the use of computers to perform tasks that humans may find difficult, often in ways that are hard to pinpoint. For example, although humans find high-level computation difficult, calculator technology is not considered AI because we know how to break this down into discrete steps and feel we understand it. However, facial recognition is a task that humans perform well, but an algorithm to identify faces is usually considered AI since we cannot articulate precisely how this is done. Machine learning (ML) is a subset of AI in which algorithms learn from the data itself without explicit programming. ML methods reflect a broad range of statistical techniques ranging from linear regression to more complex methods such as support vector machines and decision trees. ML methods can be further broken into supervised and unsupervised learning, which differ from one another in that the former requires access to gold standard labels although the latter attempts to find the answers implicitly in the data itself. While ML methods have grown more popular over recent years, the advent of a specific supervised ML method based on architectures resembling human neural networks over the past decade has led to a quantum leap in performance.2 This method, called deep learning (DL) because of many multiple internal layers, can be considered a transformative technology. Compared with previous methods that required humans to identify image features, a deep neural network trained on a dataset with known outputs can learn the best features for organizing the data. In this review, we will discuss ML methods applied to stroke imaging with an emphasis on DL applications. We refer to Figure for a graphical overview of the applications discussed in this review.