This paper presents a semantic segmentation network for brain tumor subregion segmentation from 3D magnetic resonance images (MRIs). The network is designed to address the challenges of manual delineation, which is time-consuming, expensive, and prone to human error. The proposed method uses an encoder-decoder architecture with a variational auto-encoder (VAE) branch to reconstruct the input image, providing additional regularization and constraints for the shared decoder layers. This approach won 1st place in the BraTS 2018 challenge. The network architecture includes an encoder that extracts deep image features and a decoder that reconstructs dense segmentation masks. The VAE branch, which is used only during training, helps to regularize the shared encoder by reconstructing the input image. The loss function consists of three terms: dice loss for segmentation accuracy, L2 loss for image reconstruction, and KL divergence for VAE regularization. The method was evaluated on the BraTS 2018 dataset, which included 285 cases with four 3D MRI modalities (T1, T1c, T2, and FLAIR). The network achieved high Dice scores for all three tumor subregions (enhancing tumor core, whole tumor, and tumor core) in both the validation and testing datasets. The training process was efficient, with each epoch taking about 9 minutes on a single NVIDIA Tesla V100 GPU, and the final model was trained in 6 hours using an NVIDIA DGX-1 server with 8 V100 GPUs. The paper also discusses various experimental results and comparisons with other methods, highlighting the effectiveness of the proposed approach in brain tumor segmentation.This paper presents a semantic segmentation network for brain tumor subregion segmentation from 3D magnetic resonance images (MRIs). The network is designed to address the challenges of manual delineation, which is time-consuming, expensive, and prone to human error. The proposed method uses an encoder-decoder architecture with a variational auto-encoder (VAE) branch to reconstruct the input image, providing additional regularization and constraints for the shared decoder layers. This approach won 1st place in the BraTS 2018 challenge. The network architecture includes an encoder that extracts deep image features and a decoder that reconstructs dense segmentation masks. The VAE branch, which is used only during training, helps to regularize the shared encoder by reconstructing the input image. The loss function consists of three terms: dice loss for segmentation accuracy, L2 loss for image reconstruction, and KL divergence for VAE regularization. The method was evaluated on the BraTS 2018 dataset, which included 285 cases with four 3D MRI modalities (T1, T1c, T2, and FLAIR). The network achieved high Dice scores for all three tumor subregions (enhancing tumor core, whole tumor, and tumor core) in both the validation and testing datasets. The training process was efficient, with each epoch taking about 9 minutes on a single NVIDIA Tesla V100 GPU, and the final model was trained in 6 hours using an NVIDIA DGX-1 server with 8 V100 GPUs. The paper also discusses various experimental results and comparisons with other methods, highlighting the effectiveness of the proposed approach in brain tumor segmentation.