Accurate segmentation of brain tumors is critical for diagnosis, treatment planning, and post-operative assessment. In this study, we present a deep learning-based segmentation framework developed for the BraTS 2025 challenge Task-1, addressing both pre-operative and post-operative brain tumor MRI data. Due to the distinct structural and intensity variations introduced by surgical interventions, including heterogeneous resection cavities, two separate models were trained for pre-operative and post-operative cases. A patch-based training strategy with a patch size of \(128 \times 128 \times 128\) was adopted using the Swin UNETR architecture, optimized for resource-constrained environments. Validation on a fixed 20% subset of the training data showed Dice Similarity Coefficients (DSC) of 0.905, 0.926, and 0.929 across enhancing tumor (ET), tumor core (TC) and whole tumor (WT) regions for the pre-operative model, and DSC of 0.764, 0.761, 0.903 and 0.761 across ET, TC, WT, and resection cavity (RC) regions, respectively for the post-operative model. The best-performing model was further evaluated on the unseen validation dataset provided via the Synapse platform, showing a combined DSC of 0.750, 0.761, 0.904, and 0.824, and lesion-wise DSC of 0.724, 0.725, 0.793 and 0.821 across ET, TC, WT, and RC regions, respectively. On the BraTS-GLI test set, the proposed model achieved lesion-wise DSC of 0.746, 0.735, 0.775 and 0.844 across the ET, TC, WT, and RC regions, respectively. Overall, the proposed methodology can automatically and accurately segment brain tumor on pre- as well as post-operative MRI images. The model provided superior performance on pre-operative compared to post-operative MRI images of Glioma patients.