HarMoE: Multi-Source Chest Radiograph Pretraining with Dataset-Disentangled Experts

Researchers introduced HarMoE, a dataset-aware pretraining framework for chest X-ray vision-language models.

HarMoE combines multiple labeled radiology datasets with a mixture-of-experts design to separate shared disease meaning from source-specific variation.

Using a unified disease vocabulary and masked multi-dataset supervision, the model reduces cross-dataset conflicts and better uses complementary annotations.

Experiments showed stronger zero-shot classification, out-of-distribution transfer, and grounding than competitive baselines.

The team plans to release the code and a harmonized 873K-image chest X-ray dataset.