EHD-Net: A Hybrid Approach to Brain Tumor Segmentation in MRI Using U-Net Transformer Architecture with Double Attention

Abstract

Accurate segmentation of brain tumors from magnetic resonance imaging (MRI) plays a crucial role in diagnosis, treatment planning, and follow-up. EHD-Net (Efficient Hybrid Double-Attention Network) is proposed as a hybrid architecture that combines the strengths of convolutional and Transformer-based approaches while maintaining computational efficiency. The model employs an optimized U-Net backbone with depth wise separable convolutions to reduce parameters, a 3D Contextual Transformer (CoT) at the bottleneck to capture long-range dependencies, and lightweight Double Attention (DA) modules in skip connections to refine local boundaries. Trained and evaluated on the BraTS 2021 dataset with multimodal MRI sequences (T1, T1ce, T2, FLAIR), EHD-Net achieves state-of-the-art performance with an average Dice Similarity Coefficient (DSC) of 93.2 % (95.1 % WT, 92.4 % TC, 92.0 % ET), IoU of 85.1 %, HD95 of 5.5 mm, and MSD of 1.2 mm. Importantly, EHD-Net delivers the fastest inference among the compared models, with an average processing time of 9 seconds per MRI volume, making it suitable for near real-time clinical deployment. These results highlight EHD-Net as an effective, robust, and efficient solution for brain tumor segmentation in multimodal MRI, bridging the gap between high accuracy and practical usability in clinical workflows.