Automated brain tumor segmentation must reconcile two competing needs: precise boundary localization driven by local image structure and context modeling over long spatial ranges. We introduce TransMorphNet, a hybrid architecture that (i) incorporates a learnable morphological-gradient prior (dilation–erosion contrast) as an attention gate for boundary emphasis before global mixing, and (ii) couples it to a lightweight Transformer encoder via a capacity-controlled bottleneck fusion module. On a publicly available, multi-source MRI cohort using a patient-level split and a 2D slice-based protocol, TransMorphNet achieves a mean Dice of \(0.618\pm 0.180\) , outperforming matched CNN baselines and contemporary strong 2D baselines trained under the same protocol (...). Because the Morphological Attention block primarily targets boundary refinement, we additionally report boundary-aware metrics (HD95 and ASSD), which better reflect its contribution beyond mean Dice (HD95 \(=6.12\) mm; ASSD \(=1.54\) mm on tumor-positive test slices). Voxel-wise discrimination is strong (ROC AUC \(=0.993\) ; average precision \(=0.789\) ). Under additive Gaussian noise, performance shows a modest decrease (Dice \(\downarrow 1.8\%\) at \(\sigma {=}0.10\) ). Nonparametric testing and effect-size analysis indicate consistent improvements over baselines. Grad-CAM and error maps provide qualitative evidence that model attention and errors are concentrated around lesion regions and boundaries, serving as proxy interpretability analyses. To support reproducibility, we release code, pretrained weights, and a containerized pipeline for preprocessing, training, and inference. These results are limited to a 2D, single-contrast setting on one public cohort; external multi-center validation and expert reader studies remain important next steps.
