Segmentation of meningioma tumors is important for effective radiotherapy planning. Despite recent advances in medical image segmentation, automatic and accurate segmentation of meningioma tumors on T1-contrast-enhanced (T1CE) MRI sequences remains a challenge, particularly across varied tumor shapes and sizes. This study proposed an automatic segmentation of meningioma on T1CE MRI, as a part of the BraTS-Lighthouse 2025 Meningioma Radiotherapy Segmentation Task, by leveraging state-of-the-art deep learning architectures. In the current study, two deep learning 3D segmentation networks, 3D U-Net and Swin UNETR, were trained on T1CE brain MRI volumes of 500 patients provided by BraTS-Lighthouse 2025, followed by fine-tuning of models. Further, a condition-based ensemble strategy is implemented for combining Swin UNETR and 3D U-Net to achieve an optimum Dice Similarity Coefficient (DSC). Individual model evaluation showed that Swin UNETR outperformed the 3D U-Net. On the local training dataset comprising 50 patients, the optimized Swin UNETR achieved a lesion-wise DSC of 0.81 ± 0.18, compared to 0.74 ± 0.23 for the optimized 3D U-Net. Similarly, on the external validation dataset of 70 patients, Swin UNETR achieved a DSC of 0.65 ± 0.29, while 3D U-Net obtained 0.61 ± 0.33. Furthermore, both models are condition-based ensembled for the final predictions, the ensemble model provides a lesion-wise DSC of 0.83 ± 0.13 on local training dataset of 50 patients and lesion-wise DSC of 0.70 ± 0.28 for the external validation dataset of 70 patients. Our study describes the effectiveness of condition-based ensemble using a transformer-based (Swin UNETR) and convolutional neural network-based (3D U-Net) models for Gross Tumor Volume (GTV) segmentation of meningioma tumors in MRI scans.