While deep learning has achieved numerous results in brain tumor segmentation, most models still struggle with insufficient data, particularly for heterogeneous pediatric cases. This study develops a dedicated segmentation model for pediatric patients, a population where brain tumors represent the leading cause of cancer-related death despite their rarity. Standard AI models, often trained on adult data, typically fail when faced with the distinct biological heterogeneity and imaging characteristics of pediatric tumors. The real-world data impurities of the MICCAI BraTS-PEDs 2025 Challenge, specifically the non-skull-stripped images, further confound model training. To address this trifecta of challenges—biological heterogeneity, data scarcity, and data impurity—we propose a novel three-stage segmentation framework. Our core strategy is to decouple feature learning from noise adaptation: a high-capacity U-Net is first trained on a clean, algorithmically skull-stripped dataset to learn invariant tumor features; it is then fine-tuned on the original, non-skull-stripped data to enhance domain robustness; finally, a post-processing step refines the predictions through. Developed on a limited training set (n = 261), our resource-efficient approach achieved state-of-the-art performance on the officially scored, unseen validation set (n = 91), yielding mean LesionWise Dice scores of 0.945 for the whole tumor, 0.944 for the tumor core, and impressively, 0.917 for the highly challenging non-enhancing tumor (NET) sub-region—all accomplished on a limited training set using only a single consumer-grade GPU.