SE-Net: Symmetry-Enhanced Deep Learning for Advanced Stroke-Lesion Segmentation Using Multi-Level Brain-Symmetry Priors

Acute ischemic stroke is a leading cause of death and disability worldwide, and accurate lesion segmentation is important for diagnosis and treatment planning. Current deep learning approaches for stroke-lesion segmentation often fail to use the inherent bilateral symmetry of the human brain and are thus missing a valuable anatomical prior that clinicians routinely use for lesion detection. This paper introduces SE-Net, a novel symmetry-enhanced neural network that systematically incorporates brain-symmetry priors at multiple levels within a deep learning framework. Our approach extends the 3D U-Net architecture with four specialized components: a symmetry-aware encoder with dual-stream processing, a symmetry difference module for highlighting pathological asymmetries, a perfusion-diffusion mismatch attention mechanism, and enhanced skip connections that preserve symmetry information throughout the network. Comprehensive evaluation using the Ischemic Stroke Lesion Segmentation (ISLES) 2022 dataset demonstrates exceptional performance with a Dice similarity coefficient of 96.72%, which represents an improvement of 13–39 percentage points over state-of-the-art methods. The Hausdorff distance was 14.21 mm, which indicates improved boundary precision by 35–66% improvement compared to previous approaches. Statistical analysis confirmed that highly significant performance gains were achieved (p1.7). Qualitative analysis was performed using representative segmentation results, small lesion detection examples, and symmetry difference visualizations, and the results validated the clinical applicability of our approach. Cross-dataset evaluation demonstrates robust generalization with 86-89% zero-shot performance across different imaging protocols. Clinical validation through expert evaluation confirmed the method’s potential for practical translation with an inference time of 3.4 seconds, which is suitable for acute stroke protocols. SE-Net addresses critical challenges in stroke diagnosis while establishing a framework for incorporating domain-specific anatomical knowledge into deep learning architectures for medical imaging applications.