3D U-Net for brain stroke lesion segmentation on ISLES 2018 dataset

Abstract

Brain stroke is one of the global problems today. An image such as a CT scan helps to visually see the whole picture of the brain. Segmentation of the affected brain regions requires a qualified specialist. However, manual segmentation requires a lot of time and a good expert. The role and support of trained neural networks for segmentation tasks is considered as one of the best assistants. Among the neural network models, the models based on U-Net are recognized as the leading ones. The U-Net architecture can work with a small number of datasets and is considered advanced for the segmentation method. In this paper, we use the classical U-Net architecture for the experiment. As datasets, we use 3D computed tomography images of the brain taken from ISLES 2018 the public domain. Using the classical U-Net architecture, we found that U-Net is considered the best architecture for segmentation methods. This study presents experiment results of 3D U-Net model for brain stroke lesion segmentation, and gives future perspectives for researchers who is going to segment brain strokes and create modified U-Net model for improvement. The developed model is useful for brain stroke segmentation when there is little number of images for train and testing the model.