MerMED-FM
MerMED-FM is a self-supervised medical imaging foundation model pretrained across 7 imaging modalities and 12 medical specialties.
Model Details
- Architecture: ViT-B/16
- Parameters: 85.8M
- Input size: 224 × 224
- Embedding dimension: 768
- Pretraining: Teacher–student self-supervised learning with memory-based representation learning
- Released checkpoint: Teacher backbone from epoch 50
- Paper: MerMED-FM: Multimodal, Multi-Disease Medical Imaging Foundation Model, The Lancet Digital Health, 2026
- Code: https://github.com/yangzhou12/MerMED
Intended Use
MerMED-FM can be used as a visual backbone for:
- Medical image classification
- Feature and embedding extraction
- Transfer learning
- Data-efficient fine-tuning
The pretrained model does not directly produce disease predictions. A task-specific prediction head and labeled data are required for downstream classification.
Pretraining Data
MerMED-FM was pretrained on approximately 3.3 million unlabeled images from 53 public datasets.
Supported imaging modalities include:
- Computed tomography (CT)
- Chest X-ray (CXR)
- Colour fundus photography (CFP)
- Optical coherence tomography (OCT)
- Histopathology
- Ultrasound
- Dermatoscopy
Evaluation
The published study evaluated MerMED-FM on 31 downstream datasets, including 26 public and 5 private datasets, across different labeled-data fractions.
Please refer to the paper for detailed experimental settings and results.
Limitations
MerMED-FM is intended for research use. Performance may vary across diseases, populations, institutions, imaging devices, and acquisition protocols. Independent validation is required before clinical use.
Citation
@article{zhou2026mermedfm,
title = {MerMED-FM: Multimodal, Multi-Disease Medical Imaging Foundation Model},
author = {Zhou, Yang and Quek, Chrystie Wan Ning and Zhou, Jun and Wang, Yan and Bai, Yang and others},
journal = {The Lancet Digital Health},
volume = {8},
number = {7},
pages = {101007},
year = {2026},
doi = {10.1016/j.landig.2026.101007}
}