Accurate segmentation of brain tumor sub-regions is essential for effective diagnosis and treatment planning, particularly in radiation therapy. However, heterogeneity in tumor size, location, imaging protocols, and patient demographics leads to significant variability in appearance, making the task highly challenging. Deep learning-based methods have advanced this field by mitigating the limitations of manual segmentation, which is both time-consuming and subject to inter-observer variability. In this work, we leverage the Swin UNETR, a transformer-based model designed to capture both local and global dependencies, making it well-suited for segmenting complex and variable tumor structures. To address the challenge of limited labeled data and enhance generalizability across centers, we employ a multitask learning framework that jointly performs self-supervised reconstruction and supervised segmentation, enabling robust feature learning through combined task optimization. We evaluated our approach in the BraTS 2025 Challenge dataset, focusing on the segmentation of three key subregions: whole tumor (WT), tumor core (TC), and enhancing tumor (ET). Our method achieves an average Dice score of 0.72 (0.73,0.76,0.68 for WT, TC and ET respectively) in the validation set, demonstrating strong performance and robustness under varying clinical conditions. The Github code is available: https://github.com/mumuaktar/BraTS-Challenge-GoAT.