# RadJEPA: Radiology Encoder for Chest X-Rays via Joint Embedding Predictive Architecture

Source: [RadJEPA: Radiology Encoder for Chest X-Rays via Joint Embedding Predictive Architecture](<https://www.sota2.com/research/paper/6973007cb2be0ed85a78c9d3>)

Collection index: [llms.txt](<https://www.sota2.com/research/llms.txt>)

Data updated: 2026-05-28T05:00:05.404Z

Authors: Anas Anwarul Haq Khan, Mariam Husain, Pratik Jalan, Kshitij Jadhav

Published: 2026-01-22T12:11:53.000Z

## Abstract

Vision-language pretraining has driven much of the recent progress in medical image representation learning, but this paradigm is constrained by the availability of paired image-text data and by the reporting bias of clinical narratives. We ask whether competitive radiology encoders can be learned without any language supervision. We introduce RadJEPA, a self-supervised framework built on a Joint Embedding Predictive Architecture and pretrained on approximately 840K unlabeled chest X-ray images. The model learns to predict latent representations of masked target regions from a visible context region, an objective that differs from both image-text contrastive pretraining and DINO-style self-distillation by explicitly modelling conditional structure in representation space. We evaluate RadJEPA primarily on radiology report generation with a frozen Vicuna-7B decoder, and additionally substitute its encoder into four widely used vision-language backbones (MedLLaVA, Qwen-2.5, BLIP-2, and Phi-4). For completeness we also report disease classification and semantic segmentation results. Across two datasets and four metrics, RadJEPA matches or exceeds the strongest image-only and vision-language baselines while using a ViT-B/14 backbone at 224 x 224 resolution.

## Recorded benchmark results

A research result applies to its stated dataset, split, metric and evaluation configuration. Compare like-for-like settings, preserve the metric's unit and direction, and distinguish paper-reported results from independent reproduction. The best result recorded here is not a claim of exhaustive, current global coverage.

- [Radiology Report Generation on MIMIC-CXR v1.5 (Vicuna-7B)](<https://www.sota2.com/research/sota/radiology-report-generation-on-mimic-cxr-v1-5-vicuna-7b.md>): Radiology Report Generation; MIMIC-CXR Findings generation v1.5 (Vicuna-7B); ROUGE-L: 26.1.

- [Radiology Report Generation on IU-Xray Findings generation v1.5 (Vicuna-7B)](<https://www.sota2.com/research/sota/radiology-report-generation-on-iu-xray-findings-generation-v1-5-vicuna-7b.md>): Radiology Report Generation; IU-Xray Findings generation v1.5 (Vicuna-7B); ROUGE-L: 28.4.

- [Image Classification on VinDr-CXR (test)](<https://www.sota2.com/research/sota/image-classification-on-vindr-cxr-test.md>): Image Classification; VinDr-CXR (test); AP (LO): 19.2.

- [Image Classification on RSNA-Pneumonia 5,337 images (test)](<https://www.sota2.com/research/sota/image-classification-on-rsna-pneumonia-5-337-images-test.md>): Image Classification; RSNA-Pneumonia 5,337 images (test); AP: 72.7.

## Original sources and implementation

[Original paper](<https://arxiv.org/abs/2601.15891>)

[PDF](<https://arxiv.org/pdf/2601.15891.pdf>)