We are focusing on the difficult task of predicting final lesion in stroke, a complex disease that leads to divergent imaging patterns related to the occluded artery level and the geometry of the patient’s vascular tree. We propose a framework in which convolutional neural networks are trained only from synthetic perfusion MRI - obtained from an existing physical simulator - and tested on real patients. We incorporate new levels of realism into this simulator, allowing to simulate the vascular tree of a given patient. We demonstrate that our approach is able to predict the final infarct of the tested patients only from simulated data. Among the various simulated databases generated, we show that simulations taking into account the vascular tree information give the best classification performances on the tested patients.