MT-U2Net: Mixed Transformed Base U2Net for MRI Segmentation

In this paper, we propose a powerful while efficient net architecture for deep learning, MT-U2Net for Magnetic Resonance (MR) image segmentation or other fields of computer vision like semantic segmentation, Salient Object Detection (SOD). As U-Net has made a lot of contribution to computer vision tasks, it is obvious that the network architecture can still be improved. Thus, we mainly target two weaknesses: one is the weakness of explicitly modeling long-range-dependencies, the other is missing details and features on multi-scale. We took the strengthen of MT-UNet and U2-Net so that we can handle with both the weaknesses. Thus, it is named Mixed Transformed U2Net. We coordinated the net architecture and turned it to another configuration with fewer layers to Maintain the net structural stability. However, we used the novel Transformer module named Mixed Transformer Module (MTM) supported by Local-Global Gaussian-Weighted Self-Attention (LGG-SA) and External Attention (EA) to mine the inter-connections while calculate affinities to themselves efficiently, ReSidual U-blocks (RSU) to ensure the architecture can be deeper. We completed our network so that we can segmentation images accurately.