TEFA-UNet: a texture-aware and efficient fusion-attention U-shaped network

Brain tumor segmentation from multi-modal MRI is clinically important but remains difficult because tumor sub-regions are heterogeneous, nested, and often have weak boundaries. Accurate delineation therefore requires both local texture sensitivity and longer-range anatomical context. We propose TEFA-UNet, a nested U-shaped architecture that couples a local-global aggregation (LGA) block in the encoder with a multi-scale refinement (MSR) block in the decoder. LGA modulates convolutional texture features with compressed global context, whereas MSR refines decoder features through parallel receptive fields before three sigmoid output heads for the overlapping BraTS regions: whole tumor (WT), tumor core (TC), and enhancing tumor (ET). In our experiments, the labeled BraTS 2018 training cohort is split at the subject level into 228 training cases and 57 internal validation cases to avoid slice-level leakage. The selected weighted BCE-Dice checkpoint achieved case-level Dice scores of 0.897, 0.809, and 0.697 for WT, TC, and ET, respectively, with corresponding HD95 values of 4.95, 6.87, and 5.97 pixels. A direct 2D nnU-Net baseline trained for 50 epochs on the same split achieved Dice scores of 0.891, 0.787, and 0.698, respectively. The model has 13.07M parameters, 30.64 GFLOPs, 7.50 ms single-slice inference, and 116.7 MB peak CUDA memory, indicating a moderate complexity with a latency trade-off introduced by the fusion modules. Ablation results support the complementary roles of LGA and MSR.