Abstract
arXiv:2601.15891v3 Announce Type: replace 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.