Background: Pediatric brain-tumor segmentation from multi-parametric MRI is challenging because of heterogeneous tumor characteristics, small or absent enhancing regions, and variations across imaging systems. Accurate delineation of the enhancing tumor (ET) region remains particularly difficult.

Objectives: This study aims to develop PED-SegNet, a modality-aware deep learning framework for improving pediatric brain-tumor segmentation, boundary delineation, calibration, and robustness.

Methods: PED-SegNet combines a transformer-based encoder with a UNet-style decoder to capture global context and fine spatial details. Squeeze-excitation gates and cross-modal attention enable adaptive fusion of MRI modalities. Class-balanced Dice loss, Focal Cross-Entropy, region-aware supervision for WT/TC/ET, and an ET absence-tolerant target were incorporated. The model was evaluated on the BraTS-PEDs 2023 dataset using subject-level five-fold cross-validation with Dice, HD95, and ECE metrics, along with ablation and robustness analyses.

Results: PED-SegNet achieved mean Dice scores of 0.908 (WT), 0.876 (TC), and 0.721 (ET), with HD95 values of 4.2, 5.5, and 7.8 mm, respectively. ECE values were 0.031, 0.039, and 0.057. Compared with the baseline, Dice improved by 0.035, 0.041, and 0.078 for WT, TC, and ET, respectively, with a 1.2 mm reduction in HD95. The model required approximately 49.9 M parameters, 141 G MACs, 26 s per subject, and 7.2 GB VRAM.

Conclusions: PED-SegNet effectively combines transformer-based contextual learning, adaptive multi-modal fusion, and region-aware optimization for pediatric brain-tumor segmentation. The results demonstrate improved segmentation and boundary accuracy with practical computational requirements, indicating potential for deployment-oriented neuroimaging applications.