Brain Stroke Lesion Segmentation Using Computed Tomography Images based on Modified U-Net Model with ResNet Blocks

Abstract

Segmentation of brain regions affected by ischemic stroke helps to overcome the main obstacles in modern studies of stroke visualization. Unfortunately, contemporary methods of solving this problem using artificial intelligence methods are not optimal. Therefore, in the study we consider how to increase the efficiency of segmentation of the stroke focus using computer perfusion imaging using modifications based on UNet. The network was trained and tested using the ISLES 2018 dataset. The publication includes an analysis of the results obtained, as well as recommendations for future research. By choosing the appropriate model parameters, our approach can be easily applied to detect ischemic stroke. We present modified U-Net models with two ResNet blocks as U-Net+ ResNetblock 1 and U-Net +ResNetblock 2, as well as a modified UNet model.  Due to the small number of images for training the model, the best results were obtained by applying data preprocessing and object representation approaches, as well as data normalization methods to avoid overfitting. The results show that the modified UNet model is superior to other models in terms of average distance and recall, that are significant parameters for segmentation of the stroke.