What it is
Probabilistic LLM-derived labels for all 4,407 RSNA knee training studies, improving report-based targets over regex baselines. 4,407 knee MRI studies with one row per study and 12 finding probabilities from radiology reports, including blended LLM labels and explicit 0.5 'not addressed' values.
Key results
- 4,407 studies with 12 finding probabilities each
- One row per study keyed by StudyInstanceUID
- Includes v1, v2, and blended v4 CSV label sets
Useful for training and benchmarking knee MRI classifiers with stronger report-derived targets.