Automated detection of ischemic strokes using Artificial Networks: a performance comparison of expert-trained and non-expert-trained networks

Stroke is the second-leading cause of death worldwide. Diagnosing a stroke from non-contrast-enhanced computed tomography images (NC-CT images) is often a challenge, even for experienced physicians. However, with Artificial Intelligence having become more and more present in healthcare in recent years, Neural Networks present a great opportunity to support physicians in detecting strokes early in NC-CT images. One of the biggest bottlenecks of using Neural Networks in this context is the limited availability of high-quality processed training data. The current conception is that only a few highly skilled experts have the expertise to generate data with a sufficient level of accuracy to train an effective Neural Network. These experts, however, usually lack the necessary time to do so in between their everyday clinical practice, leading to an overall shortage of useful training data. This study investigated the possibility of using non-experts (e.g., students) in addition to experts to contribute to the process of network training. Non-experts usually have more time and easier ways of integrating this task into their daily routine, which would accelerate the process tremendously. The question arises whether Neural Networks trained by non-experts are still performant enough to effectively detect strokes by analyzing NC-CT images.