Comparative Study on Stroke Lesion Core Segmentation in CTP Images

Automatic segmentation of ischemic stroke using Computed Tomography Perfusion (CTP) images becomes an essential tool for quantification of the extension of stroke in early stages because of its availability, speed, and low cost. In this paper, we combined non-contrast CT scans which help in showing the core of the lesion along with CTP perfusion maps that show the abnormality of perfusion to enhance the core segmentation task. We utilized the performance of Unet, ResUnet, and ResUnet++ architectures on the ISLES 2018 dataset to use a single architecture for direct segmentation. Data augmentation helped in increasing the training data, overcoming overfitting issues, and teaching the neural network the properties of invariance and robustness. Unet with data augmentation achieved a dice coefficient equal to 0.65 while ResUnet++ achieved a dice coefficient equal to 0.55 without using data augmentation.