Chest CT Scans with Radiology Findings and Segmentations

ReXGroundingCT links free-text radiology findings with pixel-level segmentations in 3D chest CT scans. The dataset contains 3,142 CT scans with 8,028 segmented findings across 14 categories. It was created by rajpurkarlab and updated in March 2026.

Use Cases

• Train models to generate segmentation masks from free-text radiology findings using the textual descriptions and associated masks.
• Develop visual grounding models that localize specific anatomical findings mentioned in radiology reports within 3D CT scan volumes.
• Benchmark multimodal models on the task of linking 14 different categories of medical findings across 2,992 training scans.
• Fine-tune large language models for generating descriptive radiology text conditioned on specific regions of interest highlighted in segmentation masks.

Strengths

• Contains 3,142 volumetric CT scans providing substantial 3D imaging data.
• Includes 8,028 pixel-level segmentation masks for findings across 14 distinct categories.
• Offers a dedicated training split of 2,992 scans for model development.

Limitations

• The validation set is very small with only 50 publicly available scans.
• A test set of 100 scans is held privately, limiting independent benchmarking.
• Specific details on scan provenance, patient demographics, and clinical context are not provided in the summary.

Provenance

Source

rajpurkarlab

Collection Method

Dataset designed to link radiology findings with 3D CT scan segmentations; specific collection method not detailed.

Freshness

Last updated in March 2026.

Geography

Region tag suggests 'us', but specific coverage is unknown.