NeuroLG-Seg: Attention-Guided Local–Global Feature Learning for Brain Tumor Segmentation in Super-Resolved MRI

Accurate delineation of brain tumor sub-regions on magnetic resonance imaging (MRI) underpins surgical planning, radiotherapy targeting and response assessment, yet manual segmentation is slow and varies between raters. This study presents NeuroLG-Seg, an attention-guided network that jointly learns local and global tumor features from super resolved MRI. Recurrent residual convolutional blocks and UNet++-style dense nested skip pathways capture local boundary detail across scales, a multi-head self-attention bottleneck models global context, and attention gates guide the fusion of the two in the decoder. Input volumes are up-sampled from 1.0 mm to 0.5 mm isotropic resolution with a pretrained enhanced deep super-resolution (EDSR) network, and six hyperparameters are selected by a Hybrid Tuatara Numbat Algorithm (HTNA). On a 187-case held-out test split of the BraTS 2021 cohort (1,251 cases), NeuroLG-Seg achieved Dice similarity coefficients of 0.924, 0.891 and 0.873 for the whole tumor (WT), tumor core (TC) and enhancing tumor (ET), compared with 0.896, 0.855 and 0.827 for TransUNet, the strongest of five baselines trained under the same protocol. It also reduced the ET 95th-percentile Hausdorff distance from 4.83 mm to 3.21 mm (Wilcoxon signed-rank test with Bonferroni correction, p < 0.01). In a cumulative ablation, the local components accounted for about half of the 0.068 WT Dice gain over a U-Net baseline, while the global transformer bottleneck gave the largest single gain for the enhancing tumor. On 262 TCGA-GBM cases, evaluated as an additional cross-dataset test, the model retained 97.5–98.3% of its BraTS Dice scores. Comparison with self-configuring 3D pipelines and an explicit super-resolution ablation remain necessary next steps.