Razi University

Ischemic stroke is one of the leading causes of disability and mortality worldwide. Accurate identification and segmentation of lesion regions in brain magnetic resonance imaging (MRI) play a crucial role in diagnosis, treatment planning, and assessment of brain damage severity. Despite the high capability of MRI in visualizing ischemic lesions, manual segmentation of these regions is a time-consuming process that depends on expert knowledge and is prone to inter-observer variability. Therefore, the development of automated deep learning-based methods can significantly improve the speed and accuracy of medical image analysis. In this study, a two-dimensional U-Net-based architecture, named Segmentation of ischemic stroke lesions in MRI images using a combination of attention and symmetry mechanisms in a deep network UNet, is proposed for the automatic segmentation of ischemic stroke lesions in MRI images. Furthermore, to investigate the model’s performance on a different imaging modality and evaluate its generalization capability under more diverse data conditions, a CT scan-based dataset was separately analyzed and evaluated. In the proposed model, the conventional U-Net architecture is enhanced with two main modules: the Contextual Feature Enhancement (CFE) module, designed to emphasize contextual features and reduce the impact of noise, and the Symmetry-Aware Attention (SAA) module, which exploits the relative symmetry of cerebral hemispheres to improve the detection of unilateral abnormalities. The input images were preprocessed by resizing to 256×256 pixels, intensity normalization, and preparation of binary lesion masks before being fed into the network. In addition, a combined loss function consisting of Dice Loss and Focal Loss was employed during training to maximize lesion-mask overlap and alleviate the >The performance of the proposed model was evaluated on the ISLES 2022 and ASID datasets and compared with existing approaches. Experimental results on the ISLES 2022 dataset demonstrated superior performance, achieving a Dice coefficient of 91.70%, an IoU of 84.67%, a Precision of 94.12%, a Recall of 89.39%, an Accuracy of 99.78%, and an HD95 value of 9.37%. Evaluation on the ASID dataset showed that the proposed model achieved a Dice coefficient of 80.90%, an IoU of 75.10%, a Precision of 88.02%, a Recall of 81.28%, and an Accuracy of 99.69%. The obtained results demonstrate that integrating the contextual feature enhancement module and the symmetry-aware attention mechanism into the U-Net architecture can improve segmentation accuracy, increase overlap with ground-truth masks, reduce false positive and false negative errors, and provide more precise reconstruction of ischemic lesion boundaries. Therefore, the proposed method can serve as an effective and extensible framework for intelligent computer-aided diagnosis systems in brain MRI analysis. Keywords: Ischemic stroke, Image segmentation, Magnetic resonance imaging, Attention mechanism, ISLES 2022 dataset, ASID dataset.