Tissue at Risk and Ischemic Core Estimation Using Deep Learning in Acute Stroke

Illustration of 3 approaches to define tissue at risk and ischemic core and detailed case distribution during 5-fold cross-validation. The pretraining deep learning approach (A), the separate deep learning approach (B), and the thresholding approach (C). For the pretraining, 94 patients with partial/unknown reperfusion were used in the training set, and 11 patients were in the validation set. D, Cases without 3- to 7-day follow-up were used only in training (A and B), while patients with 3- to 7-day follow-up were used in training, validation, and testing. Patients with 3- to 7-day follow-up were divided randomly into 5 sets and used in training, validation, and testing with a ratio of 3:1:1. For each fold of the tissue-at-risk model, 37 cases were used for fine-tuning on the pretrained model, 6 for validation, and 6–7 for testing. For each fold of the ischemic core model, 48 cases were used for the fine-tuning; 13, for validation; and 13–14, for testing.

Two representative cases of predictions from the pretraining, separate, and thresholding approaches. Upper row: A 71 -year-old man treated with IV tPA only, which achieved 0% reperfusion. Of note, only voxels with both diffusion and perfusion information are included in the model. His true lesion is used to define tissue at risk. Lower row: A 45-year-old man treated with IV tPA and thrombectomy, which achieved 100% reperfusion. His true lesion is used to define ischemic core. The pretraining approach had more accurate prediction than either the separate approach or the thresholding method, both visually, with DSC analysis, and volumetrically (Online Supplemental Data). Green areas overlaid on the FLAIR image represent true-positive, blue represents false-negative, and red represents false-positive.