Efficient S2D RefineFormer: Space-to-Depth Transformer with Orthogonal Multi-scale Attention for Precise 3-D MRI Glioma Segmentation

Precise delineation of glioma sub-regions in 3-D multiparametric MRI requires simultaneously preserving voxel-level boundaries and capturing whole-brain context across anisotropic planes-capabilities that conventional CNNs or windowed Transformers reconcile only partially. We introduce Efficient S2D RefineFormer, which folds every $2 \times 2 \times 2$ neighborhood into channels so that global self-attention spans the full volume with linear complexity, and appends an orthogonal multi-scale attention head to harmonize textures across sagittal, coronal and axial views. Evaluated on the BraTS 2020 validation set, the network attains a mean Dice of 0.852 and an HD95 of 3.61 mm, surpassing several state-of-the-art CNN- and Transformer-based baselines. Despite its accuracy, the model contains only 23 M parameters and requires 124 GFLOPs per scan-approximately half the compute reported for recent hybrid Transformer designs under identical settings. These results demonstrate that Efficient S2D RefineFormer combines fine-grained accuracy, global context modelling and practical computational efficiency for clinical 3-D MRI glioma segmentation.