Unmasking imaging biomarkers with deep learning Applications in acute ischemic stroke

Stroke is one of the leading causes of death and disability worldwide. An ischemic stroke is caused by a thrombus that blocks a cerebral artery and cuts off blood flow to parts of the brain. If blood flow is not restored quickly, the brain tissue in the affected area starts to get damaged. When stroke patients arrive at the hospital, they typically undergo Computed Tomography (CT) scans of the brain. These scans contain important information which can help guide treatment decisions. This thesis develops machine learning methods to automatically extract imaging biomarkers from CT scans.
In this thesis, we develop methods to automatically detect and segment thrombi that block blood vessels in the brain. From these segmentations, we build a fully automated pipeline that can measure features of the thrombus, such as its size, density, and structure, and study how these features are linked to treatment success and recovery. We also create deep learning models that can identify the brain tissue affected in the acute phase of stroke, providing estimates of the infarct core and the surrounding hypoperfused region directly from CT scans. Finally, we design a model to predict which areas of the brain are likely to become irreversibly damaged over time. Together, these tools aim to facilitate the integration of these biomarkers into clinical practice, potentially improving care for stroke patients.