MultiTrans: Multi-branch transformer network for medical image segmentation
Related Experiment Video
Updated: Jun 23, 2025
Swin-PSAxialNet: An Efficient Multi-Organ Segmentation Technique
Published on: July 5, 2024
Yanhua Zhang1, Gabriella Balestra2, Ke Zhang3
1Department of Electronics and Telecommunications, Politecnico di Torino, Corso Duca degli Abruzzi 24, Turin, 10129, Italy; School of Astronautics, Northwestern Polytechnical University, 127 West Youyi Road, Xi'an, 710072, China.
Background And Objective:
Transformer, which is notable for its ability of global context modeling, has been used to remedy the shortcomings of Convolutional neural networks (CNN) and break its dominance in medical image segmentation. However, the self-attention module is both memory and computational inefficient, so many methods have to build their Transformer branch upon largely downsampled feature maps or adopt the tokenized image patches to fit their model into accessible GPUs. This patch-wise operation restricts the network in extracting pixel-level intrinsic structural or dependencies inside each patch, hurting the performance of pixel-level classification tasks.
Methods:
To tackle these issues, we propose a memory- and computation-efficient self-attention module to enable reasoning on relatively high-resolution features, promoting the efficiency of learning global information while effective grasping fine spatial details. Furthermore, we design a novel Multi-Branch Transformer (MultiTrans) architecture to provide hierarchical features for handling objects with variable shapes and sizes in medical images. By building four parallel Transformer branches on different levels of CNN, our hybrid network aggregates both multi-scale global contexts and multi-scale local features.
Results:
MultiTrans achieves the highest segmentation accuracy on three medical image datasets with different modalities: Synapse, ACDC and M&Ms. Compared to the Standard Self-Attention (SSA), the proposed Efficient Self-Attention (ESA) can largely reduce the training memory and computational complexity while even slightly improve the accuracy. Specifically, the training memory cost, FLOPs and Params of our ESA are 18.77%, 20.68% and 74.07% of the SSA.
Conclusions:
Experiments on three medical image datasets demonstrate the generality and robustness of the designed network. The ablation study shows the efficiency and effectiveness of our proposed ESA. Code is available at: https://github.com/Yanhua-Zhang/MultiTrans-extension.
