Lightweight 3D IDC-MambaUNet for Multi-Modal MRI Brain Tumor Segmentation Using Multi-Resolution Ensemble

Brain tumor segmentation on MRI is a critical task in clinical diagnosis and treatment planning. To address this issue, we propose a lightweight network based on the U-Net architecture—3D IDC-MambaUNet. This model integrates Inception Depthwise Convolution (IDC) blocks with a U-Mamba bottleneck and employs a multi-resolution fusion strategy to effectively capture multi-scale features. Specifically, we combine a low-resolution model designed to capture global features with a full-resolution model that extracts local details, thereby integrating broad and local information simultaneously. This approach significantly improves the detection and segmentation of brain tumor lesions. Experiments on a brain tumor MRI dataset demonstrate that, compared to conventional benchmark models, our method achieves higher lesion-level F1 scores and Dice coefficients while significantly reducing the model parameters and inference time. Overall, these results indicate that our model strikes an excellent balance between high segmentation accuracy and computational efficiency, making it highly suitable for real-time clinical deployment.