Predicting stroke lesion and clinical outcome with random forests

The treatment of ischemic stroke requires fast decisions for which the potentially fatal risks of an intervention have to be weighted against the presumed benefits. Ideally, the treating physician could predict the outcome under different circumstances beforehand and thus make an informed treatment decision. To this end, this article presents two new methods: one for lesion outcome and one for clinical outcome prediction from multispectral magnetic resonance sequences. After extracting tailored image features, a random forest classifier respectively regressor is trained. Both approaches were submitted to the Ischemic Stroke Lesion Segmentation (ISLES) 2017 challenge and obtained a first and third place. The outcome underlines the robustness of our designed features and stresses the approach’s resilience against overfitting when faced with small training datasets.