U-isles: Ischemic stroke lesion segmentation using u-net

Abstract

. An accurate automation of image segmentation has a great significance in enhancing the processing of biomedical images. This could avoid the tedious task of manually segmenting biomedical images and the data inconsistencies from the varying configurations in MRI scans from hospitals. In this study, a new method in segmenting ischemic stroke lesions is proposed. The dataset is downloaded from Ischemic Stroke Lesion Segmentation 2017 (ISLES 2017) Challenge. It consists of MRI scan images of the brain from the patients. The preprocessing of data involves the application of batch normalization, image resizing, color conversion, and contrast correction. The proposed deep learning technique is a type of a two-dimensional CNN known as U-Net. Experiments are performed to determine the combination of loss function and regularizer which returns the optimal result. The results of different contrast intensities are also compared among themselves. After experimentation and segmentation, the output is post-processed to eliminate isolated small regions and smooth lesion edges. The metric used to measure the accuracy of the model is dice score.