New radiology pipeline reconfigures labels 99% cheaper without relabeling
Reconfiguring 223K radiology reports in 196 seconds with no API cost.
Public chest-radiograph (CXR) datasets like CheXpert-14 come with fixed label schemas that miss many findings described in free-text reports. A team led by Jean-Benoit Delbrouck introduced a pipeline that converts free-text reports into cached annotation matrices. Once cached, label schemas can be reconfigured through dictionary edits — no new inference passes needed. Reconfiguring MIMIC-CXR (223K reports) takes only 196 seconds and costs nothing, compared to an estimated $6.6K using Claude Opus 4.7. This approach reveals that 43% of CXR studies contain at least one finding outside the standard CheXpert-14 schema, highlighting how much clinical detail is lost.
The pipeline uses a 58-label taxonomy and trains image probes that match CheXpert-14 performance on shared targets while also reaching 0.78 AUROC on expert-reviewed long-tail labels that CheXpert-14 cannot represent. These results suggest a paradigm shift: once reports are structured, the label schema becomes a configuration to edit rather than a corpus to relabel. For radiology AI teams, this means adapting to new clinical questions, sites, or reader preferences without incurring massive annotation costs. The work is published on arXiv (2607.06597) and opens the door to more flexible, cost-effective medical imaging datasets.
- Pipeline caches free-text annotation matrices from 223K MIMIC-CXR reports in a one-time pass, then reconfigures label schemas via dictionary edits without API costs.
- Reconfiguration takes 196 seconds versus $6.6K for an equivalent Claude Opus 4.7 relabeling pass — a 99.997% cost reduction.
- Using a 58-label taxonomy, 43% of studies have findings outside standard CheXpert-14; image probes hit 0.78 AUROC on expert-reviewed long-tail labels.
Why It Matters
Rad AI teams can now adapt label schemas in seconds, cutting costs and unlocking previously ignored clinical findings.