BRAIN TUMOR SEGMENTATION VIA MULTI-SCALE ATTENTION GUIDANCE AND DYNAMIC WEIGHT OPTIMIZATION

Accurate segmentation of brain tumors is crucial for clinical diagnosis and treatment planning. This paper proposes a brain tumor segmentation method based on multi-scale attention guidance and dynamic weight optimization, applied to the precise segmentation of meningiomas, gliomas, and pituitary tumors. The method integrates a lightweight Mix Vision Transformer [Chen J-N, Sun S, He J, Torr P, Yuille A, Bai S, TransMix: Attend to mix for vision transformers, 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), IEEE, New Orleans, pp. 12125–12134, 2021.] as the backbone network, implements multi-scale feature interaction through a Channel-Spatial Dual Attention (CSDA) module, and employs a Dynamic Weight Decoder (DWD) to address class imbalance issues [Oh C, Li Y, Song K, Yun S, Han D, DaWin: Training-free dynamic weight interpolation for robust adaptation, 2025.]. Experimental results demonstrate that the proposed method achieves a mean Dice coefficient of 0.9165 (91.65%), outperforming existing approaches including Zig-RiR (87.90%) and TransUNet (85.07%) in both Dice coefficient and Hausdorff distance [Cheng Z, Liu M, Yan C, Wang S, Dynamic domain generalization for medical image segmentation, Neural Netw 184:107073, 2025.] metrics while maintaining a low parameter count (18.81M parameters, 43% fewer than comparable Transformer models).