ReXGroundingCT ReXGroundingCT is a dataset designed to link free text radiology findings with pixel level segmentations in 3D chest CT scans. Each sample consists of a volumetric CT scan, associated segmentation masks for one or more findings, and detailed textual descriptions. The dataset has segmentations for 8,028 findings across 14 different categories in 3,142 CT scans. There are 2,992 scans allocated for training, 50 for public validation, and 100 held privately to be hosted for a public leaderboard on ReXrank. Dataset Structure Each item in the dataset is represented as a JSON entry with the following structure: Field Descriptions name : The filename of the CT volume (e.g., train 1741 b 2.nii.gz ). findings : A dictionary mapping finding indices in the segmentation mask to free text radiology findings. entity counts : Number of segmented entities for each finding. shape : The shape of the CT volume in [H, W, D] (height, width, depth). pixels : Number of non zero pixels for each finding's segmentation mask. categories : The category code assigned to each finding (e.g., "2b" ). protocol : Annotation protocol used for the case (e.g., protocol1 , protocol2 ). Segmentation Masks…
Runs entirely in your browser via DuckDB-Wasm — this dataset's real data file is loaded once, then queried locally. Nothing is sent to a server.
We use cookies for essential functionality and analytics. You can accept or reject analytics cookies.Cookie policy