12 Introduction: Hemorrhagic transformation is a major limitation of thrombolytic therapy for the treatment of acute ischemic stroke. A method to identify patients at high risk for hemorrhage would be of substantial value for selecting the best candidates for treatment. Methods: Diffusion-perfusion MRIs were obtained prior to treatment and at 7 days post-treatment in patients with large vessel anterior circulation occlusions treated with intra-arterial (IA) thrombolysis. Patients received either combined IV/IA tPA or only IA thrombolytics (tPA or urokinase) within 6 hours of symptom onset. CT scans were performed immediately following the angiographic procedure and at 24 hours. Regions of hemorrhagic transformation were outlined on the post-treatment scans, which were then co-registered to the pretreatment MRI. A stepwise discriminant analysis was performed using pretreatment apparent diffusion coefficient (ADC) and mean transit time (MTT) measures to identify voxels that developed hemorrhage. Results: Eighteen patients were studied with mean age 73, 12 females / 5 males, median NIHSS score 14. Five patients (29%) developed regions of hemorrhagic transformation following therapy. There were no significant differences in time to pretreatment MRI or time to completion of thrombolysis between patients who developed hemorrhage vs. those that did not. For regions that developed hemorrhage, mean ADC was 583 um2/sec and mean MTT was 38 secs, vs. 746 and 25 for regions that went on to infarction, vs. 874 and 17 for regions that were salvaged (p<0.05 for all mean comparisons by ANOVA). A discriminant model employing pretreatment ADC and MTT variables correctly classified tissue fate (hemorrhage vs. non-hemorrhage) with overall accuracy 85%. Conclusions: In humans undergoing intra-arterial thrombolysis, pretreatment MRI ADC values are significantly lower and MTT values significantly higher in regions that develop hemorrhagic transformation compared to regions that are salvaged or develop ischemic infarction. Using these variables, a discriminant model can correctly classify tissue fate with overall accuracy of 85%.