Swin-BSSeg: A Novel Swin Transformer-Enhanced Architecture for Accurate Ischemic Stroke Lesion Segmentation in MRI Images

Abstract

Ischemic stroke, caused by obstructed cerebral blood flow, remains a leading cause of mortality and disability, necessitating precise magnetic resonance imaging (MRI)-based lesion detection. This paper proposes Swin-BSSeg (brain symmetry segmentation) algorithm, an enhanced version of the brain symmetry segmentation network (BSSNet), for improved stroke lesion segmentation. Swin-BSSeg integrates Swin Transformers to capture global context and long-range dependencies through a hierarchical attention mechanism. The encoder replaces traditional convolutions with depth-wise separable convolution (DWSC) blocks comprising three cascaded depth-wise convolutional layers for efficient feature extraction and parameter reduction. Feature transfer to the decoder is accomplished via concatenation operations. The decoder also employs DWSC blocks to reduce computational demands and incorporates attention-guided connections (AGC) to refine the contextual diversity—defined as model’s ability to capture varied and spatially distributed lesion features. The model was evaluated on the public datasets (Anatomical Tracings of Lesions after Stroke) and ISLES (Ischemic Stroke Lesion Segmentation), which achieved a Dice Coefficient of 0.842 and an accuracy of 0.87 on the ATLAS dataset, while a Dice Coefficient of 0.8049 and an accuracy of 0.84 on ISLES, outperforming BSSNet in contextual understanding and boundary precision. The improvements on the ATLAS dataset were statistically significant (p < 0.05), confirming the reliability of the proposed enhancements.