PUC-CXR: A Multimodal Grounded Chest X-ray Dataset from Latin America

Artificial intelligence for chest radiography is developed and evaluated almost exclusively on datasets from North America and Europe, leaving Latin America essentially unrepresented and raising concerns about how current models generalize to underrepresented populations. We introduce PUC-CXR, the largest publicly released chest X-ray dataset from Latin America and the first with expert-validated sentence-level visual grounding and bilingual (Spanish-English) reports. PUC-CXR contains 76,173 studies and 107,856 images from 30,000 adult patients acquired between 2018 and 2023, spanning the pre-, intra- and post-COVID-19 periods, with report sentences linked to a 47-label finding taxonomy. A grounded subset, PUC-CXR-VG, adds radiologist-drawn bounding boxes for 1,454 studies and 2,298 images and is, to our knowledge, the first public chest X-ray resource with grounded lateral views. Benchmarks on phrase grounding, report generation and image-embedding separability show that PUC-CXR is a distinct, out-of-distribution population for current chest X-ray foundation models, and that in-domain fine-tuning on it yields large phrase-grounding gains.