Meningiomas are the most common primary central nervous system (CNS) tumors in adults, typically arising from arachnoid cells in the meninges. While often benign, their recurrence and treatment planning, particularly following surgical resection, require precise volumetric delineation. Accurate segmentation of postoperative gross tumor volume (GTV) in meningioma remains a largely unaddressed challenge in automated medical image analysis. This study proposes a novel hybrid architecture, DeSURVAE (Dual-encoder Swin UNETR VAE), designed to improve segmentation performance in this setting. DeSURVAE integrates a dual-encoder framework that combines convolutional and transformer-based encoders to jointly capture local and global contextual features. A variational autoencoder (VAE) branch is incorporated to regularize the latent space and improve generalization. The model was trained on a combination of the BraTS’24 meningioma radiotherapy dataset with 500 postoperative contrast-enhanced T1-weighted (T1W + C) MRIs, and the BraTS’23 meningioma dataset with 1,000 preoperative T1W + C MRIs with consistent single-label tumor delineations. DeSURVAE achieved an average lesion-wise Dice score of 0.779 ± 0.011 and a 95th percentile Hausdorff Distance (HD95) of 21.6 ± 2.51 mm. These results represent an improvement over baseline architectures, including SegResNet with 0.696 ± 0.012 Dice and 64.2 ± 2.70 mm HD95, as well as Swin UNETR with 0.711 ± 0.018 Dice and 57.3 ± 3.74 mm HD95. Stratified analysis across tumor sizes indicated consistent performance, with poor segmentation observed in only 8% of test cases. The results suggest that hybrid encoder designs combining convolutional and transformer-based representations, along with latent space regularization (VAE branch), can effectively address the variability inherent in postoperative imaging.