Accurate segmentation of pediatric brain tumors, especially midline and brainstem gliomas, is crucial for neurosurgery and radiotherapy planning, but is hindered by sparse enhancing tumor (ET), infiltrative non-enhancing tumor core (NET), cystic components (CC), and extensive peritumoral edema (ED). We propose an attention-guided 3D U-Net with Multi-Modal and Channel-Wise Attention to enhance feature extraction from multi-parametric MRI (T1, T1-Gd, T2, T2-FLAIR), paired with a novel multi-step tumor-aware compositional augmentation pipeline to simulate tumor variability. Evaluated on the BraTS-PEDs 2025 training set (256 cases) using 3-fold cross-validation, our model achieves robust lesion-wise mean Dice scores: 0.81 ± 0.04 (ET), 0.87 ± 0.05 (NET), 0.77 ± 0.05 (CC), 0.95 ± 0.02 (ED), 0.90 ± 0.03 (WT), and 0.91 ± 0.03 (TC). Hausdorff Distance (95th percentile) values are 57.8 ± 0.6 mm (ET), 13.5 ± 0.7 mm (NET), 63.7 ± 0.8 mm (CC), 6.3 ± 0.4 mm (ED), 8.2 ± 0.5 mm (WT), and 7.1 ± 0.5 mm (TC). Compared to nnU-Net, our model improves WT and TC Dice by 2% and 6%, respectively, driven by attention mechanisms and augmentation. Ablation studies show a 6–7% Dice drop without augmentation, highlighting its role in generalizability. Increased WT and TC Dice performance directly supports more accurate radiotherapy margin definition and prognostic modeling for survival prediction, highlighting the clinical utility of our approach. Despite remaining challenges with sparse ET and irregular cystic components, our framework demonstrates robustness and scalability, paving the way for translation into clinical neuro-oncology workflows.