Modified U-Net Architecture for Ischemic Stroke Lesion Segmentation and Detection

In this paper, to improve the accuracy of detection and segmentation, we modify U-Net architecture to address ischemic stroke segmentation and detection, from ISLES 2018 dataset. In this dataset, CT images (in five modalities) and corresponding ground truth created by combining manual annotations are provided. We use shortcut connections in the architecture, which performs as a residual block. In the meantime, to reduce the overfitting caused by the scarcity of training data, we use elementwise-sum and concatenation in the network. We also use the dice coefficient and the Jaccard index to assess our model. Our architecture can be applied to ischemic segmentation and detection of CT images easily by choosing suitable hyperparameters. Experiment results show that our model can segment ischemic stroke accurately, with the dice coefficient between the segmentation given by our network and ground truth is about 0.77 while the dice coefficient of U-Net is about 0.74.