AECXR: An Arabic/English Chest X-Ray Image Captioning Dataset with Physician-Validated Captions

AECXR (Arabic/English Chest X-Ray) is a captioning dataset of 1,101 chest X-ray images taken from the original ROCO dataset (Pelka et al., 2018). Each image has three captions in Arabic and three in English: 3,303 captions per language and 6,606 in total. The first caption of each image derives from the ROCO caption. For Arabic, it was machine-translated with the Google Translate API and then validated or corrected by a physician. The second and third captions, in both languages, were written by the same physician. When the ROCO caption was in another language or was a single word such as “An X-Ray”, the physician wrote all three captions. Five physicians contributed. The images are not redistributed. Each caption is keyed by the ROCO image file name (the prefix is the PubMed Central article ID), so users can retrieve the images from the original ROCO release. Files: AECXR_ar.txt (Arabic) and AECXR_en.txt (English). Both are UTF-8, tab-separated, one caption per line, in the format <image file name>#<k><TAB><caption>, where k is 0, 1 or 2. Both files list the same keys in the same order. No train/test split is provided. Intended for research on Arabic medical image captioning. Not validated for clinical use.