Abstract WMP19: Prediction of Subacute Infarction in Acute Ischemic Stroke Using Baseline Multi-modal MRI and Deep Learning

Abstract

Objective: We aim to investigate if a deep learning algorithm can predict the location and size of subacute infarction in acute stroke patients using input data from baseline multimodal MRI. Methods: Acute ischemic stroke patients were reviewed and selected from Imaging Collaterals in Acute Stroke (iCAS) study from Apr 2014 to Aug 2017. We included patients who underwent baseline imaging including perfusion-weighted imaging, diffusion-weighted imaging, gradient echo, and arterial spin labeling; and follow up imaging with FLAIR performed 3-5 days after stroke onset. The ground truth was defined as the stroke lesions on follow-up T2-FLAIR, which were manually delineated by readers blinded to clinical information. Perfusion maps such as Tmax, cerebral blood flow, cerebral blood volume, and mean transient time were reconstructed by RAPID software. All images were co-registered to MNI template space. A U-Net model was trained with the above-mentioned 8 different contrasts as inputs (Figure A). 16 different rigid translation based augmentations were used to augment the training data size. The model was trained based on mixed loss function of cross-entropy and Dice score. Five-fold cross-validation was performed to evaluate the predictions and the results are reported as area under curve (AUC). Results: 38 patients were included (19 males, age 63±15, mRS 3 [IQR 1-4]). Eight patients underwent IV tPA only, 10 with IV tPA plus thrombectomy, 11 with thrombectomy only, and 9 without reperfusion therapy. The overall AUC of this model reached 0.942±0.053 in all slices, and 0.938±0.050 in slices that contain stroke lesions (Figure B&C), compared to 0.734±0.104 for Tmax map and 0.88±0.12 in a larger previous reported study (Nielsen et al., Stroke 2017). Conclusions: Deep learning algorithms can accurately predict established stroke lesions using the information provided in baseline multimodal MRI. This may have value for patient selection for stroke therapies.