LAUSR

The current state of the art neuroimaging in patients with acute stroke presenting 6–24 hours after symptom onset includes CT or MRI perfusion sequences for delineating volumes of the core infarct and ischemic penumbra. These imaging derived volumes together with clinical neurological symptom severity are the parameters driving the decision for endovascular therapy in the 6–24 hour time window for large vessel occlusion. As per current AHA guidelines no consideration is taken for the eloquence of the tissue at risk. Here we show an original machine learning method using perfusion MRI for predicting the expected motor improvement of reperfusing the tissue at risk in acute stroke. The ISLES 2015 data set which included diffusion and perfusion MRI as well as expert delineated core infarct and ischemic penumbra in 30 patients with acute stroke was used to train a Convolutional Neuroal Network model. The model output label maps indicating core infarct and ischemic penumbra. These maps were subsequently transformed into standard MNI space and overlaid onto a probabilistic map of motor regions. Percent overlap with primary motor, premotor and supplementary motor areas were calculated for the core infarct and ischemic penumbra. External technical performance was evaluated using clinical acute stroke MRI exams from our institution. The model derived volumes closely resembled those of the commercial RAPID software for these patients. Visual examination of the standard MNI space maps showed good anatomical alignment and correspondance of the motor areas. A software prototype generating an automatic report was developed (figure 1). The results show good technical performance of the Convolutional Neural Network model on acute stroke MRI on an independent data set. The degree of core infarcts and penumbra involvement of anatomical motor areas could be rapidly calculated using regular commercial computer hardware. Further investigation with a prospective clinical study is required for testing the clinical efficacy and possible improvement in clinical outcome prediction using the model. Disclosures H. Ullman: None. G. Duckwiler: None.