[peterhan91]

Summary

• refresh the README around the Nature Biomedical Engineering publication, DOI, verified study numbers, artifact map, citation, intended use, and limitations
• add an official Hugging Face-backed load_pretrained() path with the published DINOv2-register architecture and CXR preprocessing
• make the released .pt and guarded .pickle concept formats loadable, including NumPy compatibility
• add a runnable backbone and top-concept retrieval example
• correct Python, dependency, and Apache-2.0 package metadata
• add CITATION.cff, unit tests, and GitHub Actions package validation

Why

The publication landing page still cited medRxiv, the package metadata prevented clean builds, and the documented loader/preprocessing did not match the released checkpoint or paper experiments. These changes give readers a verified path from the Nature article to a working model while clearly distinguishing the backbone demo from the full research pipeline.

Validation

• python3 -m pytest -q — 17 passed
• python3 -m build — wheel and sdist built successfully
• python3 -m twine check dist/* — passed
• git diff --check — passed
• downloaded and loaded the real public best_model.pt checkpoint on CPU
• verified image output shape (1, 768), paired text-prompt inference, and finite outputs
• verified the released checkpoint and loader have 328 matching keys with no missing, unexpected, or shape-mismatched tensors

Notes

This PR updates the GitHub repository only. The Hugging Face model card should be synchronized separately after review, particularly its Nature citation, patient count, and artifact descriptions.