Improved Stroke Identification via Cerebral Edema Analysis Using SOR-UNET and HcH-SO-CNN

Stroke, or brain attack, occurs due to reduced blood supply to the brain. Existing stroke detection methods primarily focus on identifying stroke lesions for early diagnosis but do not account for cerebral edema, which is responsible for neurological deficits and complications; furthermore, these methods do not distinguish white matter hyperintensities from stroke lesions despite their similar imaging appearance but differing clinical implications, leading to reduced diagnostic accuracy. While prior research focused on detecting stroke lesions for early diagnosis, cerebral edema responsible for neurological deficits and complications has been less explored. This paper proposes a stroke identification model for cerebral edema analysis by detecting lesion boundaries using Second-order Regularization-U-shaped Network (SoR-UNet) and Hyperbolic cosine linearized squasHing function-Second Order similarity-based Convolutional Neural Network (HcH-SO-CNN). Magnetic Resonance Angiography (MRA) and Magnetic Resonance Imaging (MRI) brain data are first acquired and preprocessed. MRA images undergo augmentation via Neuron map-based Lyrebird Optimization–Generative Adversarial Network (NL-GAN), followed by pixel clustering using K-means. Blood vessels are segmented with SoR-UNet. For MRI data, clustering and SoR-UNet-based segmentation detect white matter hyperintensities. Lesion boundaries are then identified using X-AC to assess cerebral edema. Features from segmented regions are extracted, and Principal Component Analysis (PCA) reduces dimensionality. Finally, the proposed HcH-SO-CNN classifies stroke presence. Experimental results show the model achieves 98.67% accuracy and 98.38% sensitivity(recall), outperforming existing approaches in stroke lesion detection and cerebral edema analysis.