SLGMA-UNet: Comprehensive Feature Aggregation With Context-Sensitive Attention for Medical Image Segmentation

Medical image segmentation is essential for clinical diagnosis and treatment planning. Existing segmentation methods encounter challenges such as managing size variations, interpreting contextual relationships, and integrating multi-source data. This paper introduces SLGMA-UNet, an enhanced architecture that incorporates a multi-level attention system, adaptive scale processing Efficient Multi-Scale Attention (EMA), and a combined focus mechanism Spatial & Local-Global Attention (SLGA). The EMA module integrates multi-scale information, avoids channel dimensionality reduction, facilitates cross-space learning, and effectively establishes dependencies. The SLGA module balances local and global feature processing through an adaptive fusion strategy. Experiments conducted on five medical datasets, including ISIC2017, ISIC2018, CVC-ClinicDB, Kvasir-SEG, and GLAS, demonstrate that SLGMA-UNet outperforms existing models, achieving accuracy improvements of 4.2-7.8% (mean Intersection over Union, mIoU) and 3.6-6.4% (Dice Similarity Coefficient, DSC), along with enhanced edge detection and reduced errors in complex cases. Ablation studies confirm the positive and synergistic effects of the additional components. SLGMA-UNet offers a more effective solution for accurate medical image segmentation and shows promise for practical clinical applications.