Artificial intelligence for tissue-based treatment decisions in acute ischemic stroke

Acute ischemic stroke treatment is highly time-sensitive, primarily intravenous thrombolysis. The current time-based treatment criterion has two key limitations: symptom onset time is unknown in approximately 25% of patients, and the rate of tissue damage varies considerably between individuals. Advanced imaging modalities such as CT perfusion and MRI can directly assess tissue status but are not universally available. This thesis investigates whether AI-derived biomarkers from routinely acquired CT scans (non-contrast CT and CT angiography) can support personalized treatment decisions.
Estimating stroke onset time: A scoping review confirmed that early ischemic signs on non-contrast CT carry temporal information. A deep learning model was then developed to segment infarct core and hypoperfused regions without requiring CT perfusion, enabling automated calculation of net water uptake. Net water uptake is a measure of tissue edema with which we can classify whether onset occurred within or beyond treatment time windows. Radiomics-based machine learning models significantly outperformed simpler single metrics like net water uptake for this classification task.
Tissue-based treatment selection: Analysis of six randomized trials showed that patients with low net water uptake benefit significantly from intravenous thrombolysis, while this benefit diminished at higher values. White matter lesion burden did not modify the thrombolysis treatment effect, though it was associated with worse outcomes and unexpectedly modified the hemorrhage risk of periprocedural aspirin during endovascular treatment.
Conclusion: AI-based analysis of routine CT imaging can yield valuable biomarkers for individualized stroke treatment decisions, without reliance on advanced imaging. Validation in broader patient populations remains necessary.