MamNet-PT: A Mamba-enhanced hybrid architecture with selective state-space modeling for uncertainty-aware brain tumor segmentation.

Precise segmentation of brain tumors from MRI remains a challenging problem in medical image analysis because tumor regions exhibit substantial size variability, diffuse and infiltrative boundaries, and severe foreground-background imbalance. To address these challenges, we propose MamNet-PT, a hybrid segmentation architecture that integrates efficient long-range dependency modeling, multi-resolution feature aggregation, and uncertainty-aware prediction within a unified framework. First, a selective state-space model is embedded into the U-Net-based feature pathway to capture long-range spatial dependencies with linear computational complexity, which is particularly important for irregular and spatially extended tumor regions. Second, a pre-trained ResNet-50 encoder is used to improve feature robustness under limited annotated medical data. Third, a gated feature interaction mechanism adaptively balances Mamba-derived global contextual features and CNN-derived local boundary features, avoiding simple feature concatenation or uncontrolled module stacking. In addition, a multi-resolution pyramid fusion module strengthens scale-aware representation of small enhancing foci and extensive edema, while Monte Carlo Dropout-based uncertainty estimation provides spatial confidence maps for retrospective confidence characterization and failure-mode analysis. On the BraTS2020 benchmark, MamNet-PT achieves a Dice score of 96.7% and an Intersection over Union of 95.4%, outperforming representative CNN-Transformer and Mamba-based segmentation baselines. Ablation experiments further confirm that the performance gain is attributable to the complementary effects of selective state-space modeling, gated global-local fusion, multi-resolution aggregation, and uncertainty-aware inference. These results suggest that MamNet-PT is a promising research framework for accurate and efficient brain tumor segmentation under retrospective benchmark evaluation.