Deep learning-based stroke segmentation and patient outcome prediction

We compared the ability of 2D and 3D U-Net Convolutional Neural Network (CNN) architectures to segment ischemic stroke lesions and predict patient outcome using single-contrast (DWI) and dual-contrast images (T2w FLAIR and DWI). The predicted lesion segmentation metrics and location relative to corticospinal tract correlated with post-stroke patient outcome measured by National Institutes of Health Stroke Scale (NIHSS). The 2D multi-modal CNN achieved the best results with mean Dice of 0.74. The highest correlation was for weighted-lesion load with both baseline and 90-days NIHSS (80%, p<0.001). Our results support that multi-contrast MR helps automate lesion segmentation and predict post-stroke outcomes.