SMAFormer-Medical-Image-Segmentation

SMAFormer: Synergistic Multi-Attention Transformer for Medical Image Segmentation

This repository contains the weights and official implementation for SMAFormer, presented at BIBM 2024.

[Paper] [Code] [Video Introduction]

🛎 Citation

If you find our work helpful for your research, please cite:

@inproceedings{zheng2024smaformer,
  title={Smaformer: Synergistic multi-attention transformer for medical image segmentation},
  author={Zheng, Fuchen and Chen, Xuhang and Liu, Weihuang and Li, Haolun and Lei, Yingtie and He, Jiahui and Pun, Chi-Man and Zhou, Shoujun},
  booktitle={2024 IEEE International Conference on Bioinformatics and Biomedicine (BIBM)},
  pages={4048--4053},
  year={2024},
  organization={IEEE}
}

@article{zheng2025hbformer,
  title={HBFormer: A Hybrid-Bridge Transformer for Microtumor and Miniature Organ Segmentation},
  author={Zheng, Fuchen and Chen, Xinyi and Li, Weixuan and Li, Quanjun and Zhou, Junhua and Guo, Xiaojiao and Chen, Xuhang and Pun, Chi-Man and Zhou, Shoujun},
  journal={arXiv preprint arXiv:2512.03597},
  year={2025}
}

📋 Overview

SMAFormer V2.3 is a medical image segmentation architecture based on the Swin Transformer. Compared to the initial version (ViT-Base, 138M parameters), V2.3 uses a Swin Transformer backbone to achieve superior performance (89% Dice on Synapse) with only 41.4M parameters. It features an innovative mirror Encoder-Decoder design, which allows the decoder to load pre-trained weights via reverse mapping, reaching a total pre-training coverage of 94.84%.

🌳 Version Evolution

                           SMAFormer
                               │
           ┌───────────────────┴───────────────────┐
           │                                       │
    [Branch: ViT-Base]                    [Branch: Swin-Tiny]
           │                                       │
    ┌──────┴──────┐                        ┌──────┴──────┐
    │  SMAFormer  │                        │ SMAFormerV2 │
    │  (ViT-Base) │                        │ (Swin-Tiny) │
    └──────┬──────┘                        └──────┬──────┘
           │                                       │
    v1.1 Dice: 74%                         v2.1 Dice: 78%
           │                                       │
    v1.2 Dice: 80%                         v2.2 Dice: 80%
           ▼                                       │
                                           v2.3 ★ Mirror Architecture
                                           └─ Dice: 89%

📈 Performance (Synapse Dataset)

🚀 Quick Start

Dependencies

torch>=1.10.0
timm>=0.6.0
numpy>=1.21.0
scipy>=1.7.0
h5py>=3.1.0

Training

python train_synapse.py --max_epochs 300 --batch_size 40 --model SMAFormerV2 --config configs/config_setting_synapse.py

Evaluation

python test_synapse.py --model_path checkpoints/best.pth

🧧 Acknowledgement

This work was supported in part by the National Key R&D Project of China, the Natural Science Foundation of Guangdong Province, the Shenzhen Technology Innovation Commission, the Shenzhen Development and Reform Commission, and the Science and Technology Development Fund, Macau SAR.