---

🔥 News

• [2026.03] 💻 Code is now publicly available.
• [2026.03] 📄 Paper released on arXiv.
• [2026.02] 🎉 Our paper has been accepted to CVPR 2026!

---

🧠 Method Overview

The VEMamba pipeline consists of four key stages:

1. Shallow Feature Extraction
2. Degradation Representation Extraction (via MoCo)
3. Deep Feature Extraction (using Residual Volume Mamba Groups)
4. Reconstruction Module

The core component is the VEMamba Module (VEMM), which features:

• ALCSSM: Multi-directional 3D dependency scanning.
• SSM: Global dependency modeling.
• DWAM: Dynamic feature aggregation.

---

📂 Datasets

We evaluate our method on two public Volume Electron Microscopy (VEM) datasets:

1. EPFL Dataset

• Content: FIB-SEM hippocampus dataset with annotated mitochondria segmentation labels.
• Resolution: 5 × 5 × 5 nm

2. CREMI Dataset

• Content: ssTEM dataset of Drosophila melanogaster brain (contains three training volumes: A, B, C).
• Resolution: 4 × 4 × 40 nm

---

⚙️ Environment

Recommended Requirements:

• python == 3.10
• torch == 2.4.0
• causal_conv1d == 1.5.2
• mamba_ssm == 2.2.5

Standard Installation:

pip install -r requirements.txt

⚠️ Note: To avoid environment conflicts, we highly recommend installing the following core packages manually from their official releases:

• causal-conv1d releases
• mamba-ssm releases

---

🚀 Training

Training is divided into two stages to ensure optimal representation learning and reconstruction.

Stage 1: Degradation Learning (MoCo)

Train the MoCo encoder to effectively learn degradation representations.

python train_moco.py

Stage 2: Reconstruction Training

Freeze the MoCo encoder and train the main reconstruction backbone.

python train.py

---

🧪 Testing

To reconstruct the full isotropic volume, run the following command. The output will be saved in your specified output directory.

python test.py

---

🙏 Acknowledgements

This project is built upon the excellent work from the following open-source repositories. We sincerely thank the authors for making their code publicly available:

• SCST
• IsoVEM
• MambaIR
• CDFormer

---

📚 Citation

If you find this work or code useful for your research, please cite our paper:

@inproceedings{gao2026vemamba,
  title={VEMamba: Efficient Isotropic Reconstruction of Volume Electron Microscopy with Axial-Lateral Consistent Mamba},
  author={Longmi, Gao and Pan, Gao},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
  year={2026}
}