PhD defence L. (Lennard) Wolff

Below is a brief summary of the dissertation:

Stroke is a major cause of death and long-term disability, with ischemic stroke accounting for most cases. Rapid diagnosis and treatment, particularly in large vessel occlusion (LVO), are crucial, and imaging plays a central role. Non-contrast CT and CT angiography (CTA) identify thrombus location, early ischemic changes, and collateral circulation, which guide treatment decisions. However, visual assessment is time-consuming and variable. This thesis evaluates automated imaging biomarkers in acute ischemic stroke.

First, a novel AI algorithm for LVO detection on CTA showed high sensitivity for proximal occlusions but lower sensitivity for distal (M2) occlusions, especially in real-world datasets, highlighting both its potential and limitations. Second, automated ASPECTS scoring on non-contrast CT demonstrated moderate agreement with expert radiologists and consistent performance across scanner vendors.

Third, a deep learning model for automated collateral scoring achieved performance comparable to expert readers. Collateral scores were influenced by CTA timing, with peak arterial phase imaging showing the strongest clinical associations. Despite timing-related variability in score values, outcome prediction remained stable. Training reduced inter-rater variability among human readers, and automated tools performed similarly to trained raters and commercial software. Collateral status was not associated with time from symptom onset within 6.5 hours and did not modify treatment effects.

Finally, replacing visual ASPECTS and collateral scores with automated equivalents in the MR PREDICTS outcome model resulted in comparable predictive performance. Overall, automated imaging biomarkers perform similarly to radiologists and can support standardized, efficient, and data-driven decision-making in acute stroke care.