A Dual-Encoder-Based Model for Small Lesion Segmentation in Stroke

Stroke, a major global health concern, with ischemic stroke being prevalent. Precise stroke lesion detection via medical image segmentation is vital. However, small lesion segmentation faces issues like size variation and class imbalance, and existing methods have limitations. To address inaccurate small ischemic stroke lesion segmentation, this paper presents a dual - branch algorithm integrating CNNs and Transformers. The SMP - CGLU encoder adaptively extracts local features by adjusting moving point positions and receptive fields, and CGLU modulates channel - wise attention. The Transformer branch, using Swin Transformer, captures global dependencies and fuses local - global information through window - splitting attention. Experiments on ATLAS R2.0 validate the algorithm. Comparative results show it outperforms in metrics like Dice, ASSD, HD, Precision, and Recall, achieving Dice scores of 71.34%. Ablation experiments confirm each module's contribution, highlighting the effectiveness of the proposed approach for small lesion segmentation.