deepLEARNING786/ROCOv2-X-Ray-radiology_-cycle-1

ROCOv2 X-Ray, Report Generation Pilot (Cycle 1)

A 28-row pilot testing whether short ROCO captions can be expanded into report-shaped training pairs.

Why this exists

ROCOv2 captions are one or two clipped sentences, often written to make a teaching point rather than to read as a radiological finding. A vision-language model trained directly on them learns to produce clipped captions, not reports.

This pilot tested a different approach: take the caption and its UMLS concept codes, and expand them into a full-length report plus a synthetic clinical history, so the model sees both the kind of context a radiologist would have before reading a study and the kind of output expected afterwards.

It is deliberately small. 28 rows is a feasibility check, not a training set. It answers "does this expansion produce usable text?" and nothing more.

Columns

Inherited from the source dataset:

Added by this pilot:

How the text was generated

generated_report and generated_history were produced by [GENERATOR MODEL — fill this in], prompted with the source caption and its cui concept codes.

Both fields are synthetic. They are not written by a radiologist, and they were not reviewed by one.

Limitations

Read these before using the data.

• —The reports are model output, not ground truth. Anything the generator invented is now in the dataset. A model trained on this can learn to reproduce plausible-sounding findings that were never in the image. The original caption and cui columns remain the only trustworthy signal.
• —28 rows. Far too few to train on, and far too few to draw conclusions from. This is a pilot.
• —Not manually reviewed. No sample was checked against the source caption for factual drift.
• —No grounding. Nothing connects a sentence in generated_report to a region of the image, which is the limitation that stopped this line of work.
• —Histories are invented. generated_history is plausible clinical context, not a real patient record. It is there to shape the model's expected input format, not to convey fact.

If you are scaling this up

The obvious next step is to run the same expansion over all 4,254 rows. Two things worth doing first:

1. 1.Publish the generation prompt alongside the data, so the synthetic text is auditable.
2. 2.Manually review a sample against the source caption and cui codes, and report the drift rate. Without that number, a scaled dataset inherits an unknown hallucination rate.

⚕️ Not for clinical use

Synthetic research data. Not validated, not reviewed by a clinician, and not suitable for diagnosis, triage, or any clinical decision.

Attribution

Derived from ROCOv2; please cite the source dataset and respect its licence. UMLS concept codes are subject to the UMLS licence.