Ischemic Stroke Lesion Segmentation in MRI using Convolutional Neural Networks .0

Abstract

Introduction: Magnetic resonance imaging (MRI) is widely used to diagnose stroke. Stroke is the third leading cause of death and the most significant cause of disability worldwide. Determining the location of a brain lesion plays a vital role in deciding on optimal treatment intervention. Materials and Methods: This dissertation proposes a structure for automatically segmenting stroke lesions using FLAIR sequence images. The proposed network is based on the U-Net architecture with an encoder and decoder path; each has blocks consisting of five parallel layers. Results: The proposed model was evaluated on the ISLES2015 challenge data set for ischemic stroke lesion segmentation, and the mean accuracy of Dice coeficient achieved is 0.89 in total. Conclusion: One of the most critical steps in implementing methods based on deep learning is preprocessing and preparing data for network entry. In this study, the effect of the contrast limited adaptive histogram equalization (CLAHE) method was obvious as preprocessing. Also, in the architecture section, creating blocks with multi-layer convolution was effective in learning the features and increasing the architectural efficiency.