UCLA

For patients diagnosed with acute ischemic stroke, treatments such as thrombolysis and thrombectomy aim to restore blood flow to areas experiencing ischemia. These treatments have vastly improved outcomes, but it is currently unknown why some patients experience unsuccessful reperfusion or hemorrhagic complications. Taking advantage of recent advances in deep learning vision transformers, we developed algorithms for classification and prediction tasks regarding a patient’s potential response to therapies using imaging taken at hospital admission. These models achieved higher generalization performance when identifying patients within the treatment window and those that will achieve successful recanalization. Our results illustrate that magnetic resonance (MR) and computed tomography (CT) imaging contains signal that can predict successful treatment response and that deep learning models can localize to salient regions within imaging without requiring time-intensive manual segmentation.