FedFound: Federated Foundation Models for Gastrointestinal Endoscopy

This repository contains pretrained foundation models released as part of the paper:

FedFound: Federated Foundation Models for Gastrointestinal Endoscopy

The models were trained using self-supervised learning on gastrointestinal endoscopy images under centralized, local, and federated learning settings. Two pretraining paradigms are provided:

• Masked Autoencoder (MAE)
• Momentum Contrast (MoCo)

These checkpoints can be used as initialization for downstream gastrointestinal endoscopy tasks such as classification, segmentation, and representation learning.

---

Available Checkpoints

Checkpoint	Clients	Pretraining
lb_split1.pth	1	MAE
lb_split2.pth	1	MAE
lb_split10.pth	10	MAE
lb_split20.pth	20	MAE
ub_central.pth	Centralized	MAE
fedavg_split1.pth	6	MAE
fedavg_split2.pth	6	MAE
fedavg_split10.pth	10	MAE
fedavg_split20.pth	20	MAE
fedavgm_split1.pth	6	MAE
fedavgm_split2.pth	6	MAE
fedavgm_split10.pth	10	MAE
fedavgm_split20.pth	20	MAE
fedadam_split1.pth	6	MAE
fedadam_split2.pth	6	MAE
fedadam_split10.pth	10	MAE
fedadam_split20.pth	20	MAE
fedadagrad_split1.pth	6	MAE
fedadagrad_split2.pth	6	MAE
fedadagrad_split10.pth	10	MAE
fedadagrad_split20.pth	20	MAE
moco_lb_split1.pth	1	MoCo
moco_lb_split2.pth	1	MoCo
moco_ub_central.pth	Centralized	MoCo
moco_fedavg_split1.pth	6	MoCo
moco_fedavg_split2.pth	6	MoCo

---

Naming Convention

• lb: Lower Bound (single-client training)
• ub: Upper Bound (centralized training)
• fedavg: FedAvg aggregation
• fedavgm: FedAvgM aggregation
• fedadam: FedAdam aggregation
• fedadagrad: FedAdagrad aggregation
• moco: Momentum Contrast (MoCo) pretraining
• Models without the moco prefix use Masked Autoencoder (MAE) pretraining

---

Usage

import torch

checkpoint = torch.load("fedavg_split1.pth", map_location="cpu")

if isinstance(checkpoint, dict) and "model" in checkpoint:
    state_dict = checkpoint["model"]
else:
    state_dict = checkpoint

model.load_state_dict(state_dict, strict=False)

Repository Contents

This repository contains only pretrained model weights.

No patient images, labels, metadata, or clinical information are included.

---

Citation

If you use these models in your research, please cite:

@article{devkota2025federated,
  title={Federated foundation model for gi endoscopy images},
  author={Devkota, Alina and Amireskandari, Annahita and Palko, Joel and Thakkar, Shyam and Adjeroh, Donald and Jiang, Xiajun and Bhattarai, Binod and Gyawali, Prashnna K},
  journal={arXiv preprint arXiv:2505.24108},
  year={2025}
}

---

Contact

For questions regarding the models, datasets, or training procedures, please open an issue or contact the authors of the paper.