Research question
How do ADS, an offline CPU-based model, DeepISLES, a cloud-based model, and a custom composite model compare in detecting and segmenting acute ischemic infarcts on MRI scans, and how does their performance change under controlled image-degradation conditions?
Summary
Stroke is a medical emergency in which rapid diagnosis can be the difference between recovery, permanent disability, or death. However, identifying a stroke and determining its type can be challenging, particularly when symptoms are subtle or when medical specialists and advanced imaging resources are not immediately available.
Advances in artificial intelligence are creating new opportunities to analyze medical imaging, patient data, and clinical symptoms with greater speed and accuracy. This project will explore whether AI can improve the early detection and diagnosis of stroke, potentially reducing time to treatment and improving patient outcomes, while also considering concerns surrounding reliability, bias, clinical integration, and the role of physicians in AI-assisted decision-making.
Why it matters
In stroke care, every minute without treatment matters, and many hospitals do not have a specialist reading scans around the clock. Knowing how different AI models hold up, including on lower-quality images and on hardware without cloud access, helps show where these tools could support clinicians and where they fall short.