Resources

• 📄 Paper: https://arxiv.org/abs/2601.15891
• 🤗 Hugging Face: https://huggingface.co/AIDELab-IITBombay/RadJEPA
• 💻 Code (GitHub): https://github.com/aidelab-iitbombay/RadJEPA

RadJEPA

RadJEPA is a self-supervised vision encoder for chest X-ray images based on a Joint Embedding Predictive Architecture (JEPA).
The model learns visual representations by predicting latent features of masked image regions, without text supervision or pixel-level reconstruction.

RadJEPA is intended as a general-purpose radiology image backbone for downstream tasks.

Overview

• Model type: Vision Transformer–based JEPA encoder
• Training: Self-supervised latent prediction
• Input: Chest X-ray images
• Finetuned from model: ijepa

Intended use

The model is a vision backbone that can be plugged to other models for downstream tasks. Typical downstream applications include:

• Multi-label classification
• Semantic segmentation using patch embeddings
• Image retrieval and clustering
• Report generation, with a language model to decode text

Load RadJEPA

from transformers import AutoModel
model = AutoModel.from_pretrained(
    "AIDElab-IITBombay/RadJEPA",
    trust_remote_code=True
)
print(model)

Dependency note (timm)

If you encounter issues with newer versions of timm, install the known working version explicitly:

pip install timm==1.0.24

Training details

Training data

We used images from five public, deidentified chest X-ray datasets to train this checkpoint of RAD-DINO.

Biases, risks, and limitations

RAD-DINO was trained with data from three countries; therefore, it might be biased towards the population in the training data. Underlying biases of the training datasets may not be well characterized.

Training procedure

We refer to the manuscript for a detailed description of the training procedure.

Evaluation

Our evaluation is best described in the manuscript.

Baselines

We report results for a subset of consistently competitive baselines for clarity. Notably, RadJEPA uses a ViT-B/14 backbone (86M parameters), making it substantially smaller than I-JEPA (ViT-H/14, 0.6B parameters), yet it achieves superior performance across classification, segmentation, and report generation tasks. Furthermore, Rad-DINO and RadJEPA are the only methods pretrained on comparable chest X-ray datasets at similar scale, enabling a direct and fair comparison of self-supervised objectives under matched data and model capacity.

Classification

Segmentation

Report Generation

Software

We leveraged the code in ijepa for training. We used SimpleITK and Pydicom for processing of DICOM files.

Citation

@misc{khan2026radjeparadiologyencoderchest,
      title={RadJEPA: Radiology Encoder for Chest X-Rays via Joint Embedding Predictive Architecture}, 
      author={Anas Anwarul Haq Khan and Mariam Husain and Kshitij Jadhav},
      year={2026},
      eprint={2601.15891},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2601.15891}, 
}

Acknowledgements

RadJEPA builds upon the I-JEPA architecture.
We thank the authors for making their work publicly available.

Model Card Contact

Anas Anwarul Haq Khan
Department of Computer Science and Engineering, IIT Bombay
📧 anaskhan@cse.iitb.ac.in

Mariam Husain
Department of Biomedical Engineering, Johns Hopkins University, USA
📧 mhusai10@jh.edu

Kshitij Jadhav
Koita Centre for Digital Health, IIT Bombay
📧 kshitij.jadhav@iitb.ac.in