Wavelet Convolution-Edge-enhanced Dual Encoder Network for Stroke Lesion Segmentation

Stroke (Cerebrovascular accident, CVA) is one of the leading causes of disability and mortality worldwide. According to the World Health Organization, over 12 million people die from stroke annually, and approximately 50 million people lose their ability to live independently due to stroke-related disabilities. China has the highest number of stroke patients globally. Accurate lesion segmentation is a critical technical foundation for stroke treatment and prognosis assessment. However, the similarity in grayscale features of tissue structures in medical images, along with the complex background and noise interference in ischemic stroke (accounting for 87% of stroke cases), often leads to blurred edges and inaccurate localization in lesion segmentation. To address these challenges, this paper proposes a segmentation method incorporating edge enhancement strategies. By introducing wavelet convolution skip connections, edge features are enhanced through frequency domain decomposition, enabling plug-and-play edge optimization. Additionally, an edge-enhanced auxiliary segmentation head is designed, combined with gated feature alignment and local enhancement modules, to further improve the model's adaptability to edge information. During the inference phase, the auxiliary head is discarded to avoid increasing model complexity and computational overhead. Experiments on the ATLAS R2.0 dataset demonstrate that the proposed method effectively preserves lesion edge information, significantly improves segmentation performance, and outperforms other state-of-the-art models.