Efficient U-Net CNN with Data Augmentation for MRI Ischemic Stroke Brain Segmentation
Fathia Aboudi, C. Drissi, Tarek Kraiem
Abstract
Ischemic stroke brains are common cerebrovascular diseases and one of the major issues of long-term disabilities and mortalities in the world. The Detection of ischemic stroke lesions has a significant role in the diagnosis process. In terms of timing and accuracy, automated biomedical segmentation from a magnetic resonance imaging modalities has shown to be quite beneficial. In this work, we suggest a deep convolutional neural network (CNN) approach based on U-Net architecture that is able to separate ischemic stroke lesions from normal tissue. Fine-tuning technique has been applied to adopt our aims with U-Net architecture. We used a public dataset ISLES 2015 to evaluate the proposed approach. Experimentally, our network achieved an average Dice Coefficient (DC), and accuracy is 55.77%, 99.96% respectively. Quantitative measures show that U-Net CNN provides significant evaluation metrics. Our proposed network could be used to create an automated tool to segment ischemic stroke brain lesions.