About this deposit BI-RADS lexicon annotations for a subset of 400 images (100 cases) from the Emory BrEast Imaging Dataset (EMBED), released as a companion to: Corbetta, V., et al. "Beyond Clean Test Sets: Spurious Correlations in Medical Vision-language Models and the Role of Concept Supervision." Medical Image Computing and Computer Assisted Intervention (MICCAI), 2026. The deposit contains: Tabular annotations (CSV) with structured BI-RADS lexicon labels per image Pixel-wise segmentation masks (NRRD) registered to the original EMBED DICOMs Important: this is not a redistribution of EMBED These annotations are not usable on their own. The annotation files reference original EMBED image identifiers; no EMBED imaging data is redistributed here. To use these annotations you must have independently registered for and obtained the EMBED dataset via Emory's official access process (https://github.com/Emory-HITI/EMBED_Open_Data), agreed to the EMBED Research Use Agreement, and agreed to the terms governing this release. Release context This release is made with the written permission of the EMBED dataset creators at the Emory University School of Medicine (Health Innovation and Translational Informatics Lab). Code and documentation The code for the accompanying paper, a notebook demonstrating how to use the annotations and full documentation are available at: https://github.com/valecorbetta/framework_for_vlm_eval Access Access is gated by independent verification of EMBED access. Please email Valentina Corbetta (v.corbetta@nki.nk) with proof of access to EMBED. Expected response time: 5 working days. Citation If you use these annotations, please cite the following works: @article{jeong2023emory, title={The EMory BrEast imaging Dataset (EMBED): A racially diverse, granular dataset of 3.4 million screening and diagnostic mammographic images}, author={Jeong, Jiwoong J and Vey, Brianna L and Bhimireddy, Ananth and Kim, Thomas and Santos, Thiago and Correa, Ramon and Dutt, Raman and Mosunjac, Marina and Oprea-Ilies, Gabriela and Smith, Geoffrey and others}, journal={Radiology: Artificial Intelligence}, volume={5}, number={1}, pages={e220047}, year={2023}, publisher={Radiological Society of North America} } @inproceedings{corbetta2026beyond, title={Beyond Clean Test Sets: Spurious Correlations in Medical Vision-language Models and the Role of Concept Supervision}, author={Valentina Corbetta and Portaluri, Antonio and Ze, Muzhen and Boeke, Daniël and Beets-Tan, Regina and Lachi, Veronica and Wetzer, Elisabeth and Jenssen, Robert and Wilson Silva and Kristoffer Wickstr{\o}m}, booktitle={Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026}, year={2026} }