SAMed-2 is a new foundation model for medical image segmentation built upon the SAM-2 architecture. Specifically, we introduce a temporal adapter into the image encoder to capture image correlations and a confidence-driven memory mechanism to store high-certainty features for later retrieval. This memory-based strategy counters the pervasive noise in large-scale medical datasets and mitigates catastrophic forgetting when encountering new tasks or modalities.

[06/2025] 🎉 SAMed-2 is accepted by MICCAI 2025!
[06/2025] 🚀 Initial release of SAMed-2!

This project is released under the Apache 2.0 license.

Linux Environment

Clone this repository and navigate to the folder:

git clone https://github.com/ZhilingYan/Medical-SAM-Bench.git
cd Medical-SAM-Bench

# Create a new conda environment
conda create -n samed2 python=3.10 -y
conda activate samed2

# Install PyTorch (adjust cuda version as needed)
pip install torch==2.5.1 torchvision==0.20.1 --index-url https://download.pytorch.org/whl/cu118

# Install other requirements
pip install -r requirements.txt

📁 Place downloaded weights in ./checkpoints/

Download the pre-trained memory bank: memory_bank_list_640.pkl
Place it in the root directory of this repository.

🚀 Simple Python API

from predict import MedicalSegmenter

# Initialize
segmenter = MedicalSegmenter(
    model_type='samed2',
    checkpoint_path='checkpoints/latest_epoch_0217.pth'
)

# Segment
result = segmenter.predict(
    'medical_image.png', 
    box=[100, 100, 900, 900]
)

# Visualize
segmenter.visualize(
    'medical_image.png', 
    result['mask'], 
    'result.jpg'
)

📊 Benchmark Medical SAM Models

# 🏆 SAMed-2 (Ours)
python main.py -net samed2 -sam_ckpt checkpoints/latest_epoch_0217.pth -sam_config sam2_hiera_s

# 🔬 MedSAM2
python main.py -net medsam2 -sam_ckpt checkpoints/MedSAM2_pretrain.pth -sam_config sam2_hiera_t_original

# 🏥 MedSAM
python main.py -net medsam -sam_ckpt checkpoints/medsam_vit_b.pth

# 🎯 SAM2
python main.py -net sam2 -sam_ckpt checkpoints/sam2_hiera_small.pt -sam_config sam2_hiera_s_original

# 🔷 SAM
python main.py -net sam -sam_ckpt checkpoints/sam_vit_b_01ec64.pth

💡 Common: -exp_name ${DATASET} -image_size 1024 -data_path /path/to/data -val_file_dir /path/to/test.txt

Prepare Your Own Dataset

For custom datasets, organize your data as follows:

your_dataset/
├── image/
│   ├── case_idx_slice_001.png
│   ├── case_idx_slice_002.png
│   └── ...
├── mask/
│   ├── case_idx_slice_001.png
│   ├── case_idx_slice_002.png
│   └── ...
└── test.txt  # List of test image names

Run Evaluation

# Full evaluation on the dataset
bash run.sh

Parameters Explanation:

• -net: Model type (samed2, medsam2, sam2, medsam, sam)
• -exp_name: Dataset name for logging
• -sam_ckpt: Path to model checkpoint
• -sam_config: Configuration file
• -image_size: Input image size (default: 1024)
• -out_size: Output size (default: 1024)
• -b: Batch size
• -data_path: Root path to datasets
• -train_file_dir: Path to training file list
• -val_file_dir: Path to validation file list
• -memory_bank_size: Memory bank size for SAMed2 (default: 640)
• -lr: Learning rate (default: 1e-4)
• -epoch: Number of epochs (default: 100)

📈 Performance Comparison

If you find SAMed-2 useful in your research, please consider citing:

@article{yan2025samed,
  title={SAMed-2: Selective Memory Enhanced Medical Segment Anything Model},
  author={Yan, Zhiling and Song, Sifan and Song, Dingjie and Li, Yiwei and Zhou, Rong and Sun, Weixiang and Chen, Zhennong and Kim, Sekeun and Ren, Hui and Liu, Tianming and others},
  journal={arXiv preprint arXiv:2507.03698},
  year={2025}
}

Contributors:

Zhiling Yan¹, Sifan Song², Dingjie Song¹, Yiwei Li³, Rong Zhou¹, Weixiang Sun⁴, Zhennong Chen², Sekeun Kim², Hui Ren², Tianming Liu³, Quanzheng Li², Xiang Li², Lifang He¹, Lichao Sun¹*

¹Lehigh University
²Massachusetts General Hospital and Harvard Medical School
³University of Georgia, Athens
⁴University of Notre Dame

We gratefully acknowledge:

• SAM2 - The foundation model we build upon
• MedSAM - Inspiration for medical adaptation