ReXGroundingCT

Segmentation of Findings from Free-Text Reports

About ReXGroundingCT

ReXGroundingCT is a large-scale 3D chest CT dataset linking free-text radiology findings to pixel-level segmentations in volumetric imaging. It comprises 3,142 non-contrast chest CT scans with 8,028 annotated findings (16,301 entities) from the CT-RATE dataset. ReXGroundingCT enables sentence-level grounding for both focal and non-focal lung and pleural abnormalities across 14 categories. On ReXrank, we are hosting ReXGroundingCT's testset, which contains 100 CT scans with exhaustive radiologist annotations for all visible findings.


Performance Metrics

Global Dice: Average Dice per finding per case

Global HIT Rate: Proportion of findings that have Dice >= 0.1

Instance Precision: TP / (TP + FP), where True Positives are instances that have a Dice >= 0.2

Instance Recall: TP / (TP + FN)


Model Performance

Model Global Dice Global HIT Rate Instance Precision Instance Recall Instance F1
DAGG

Submission 2
Submission 1
Submission 2
0.253 0.517 0.235 0.292 0.26
SAT-FT

SJTU

0.209 0.473 0.074 0.369 0.123
BiomedParseV2

Microsoft

0.025 0.066 0.007 0.072 0.012
SegVol

SJTU & BAAI

0 0 0 0 0
SAT

SJTU

0 0 0 0 0