CoolFace
Datasetpublic

deepLEARNING786/ROCOv2-X-Ray-radiology

ROCOv2 X-Ray Subset Radiographs extracted from ROCOv2 (Radiology Objects in COntext, version 2), for training vision-language models on X-ray interpretation. What this is ROCOv2 spans many imaging modalities. This subset keeps only the X-ray studies, so that a model can be trained on a single modality rather than learning across CT, MRI, ultrasound and radiography at once. Rows 4,254 Split train Size ~977 MB Modality X-ray only… See the full description on the dataset page: https://huggingface.co/datasets/deepLEARNING786/ROCOv2-X-Ray-radiology.

sourceHugging Faceupdated 15d agoView on Hugging Face
2likes41downloads
Dataset Card

ROCOv2 X-Ray Subset

Radiographs extracted from ROCOv2 (Radiology Objects in COntext, version 2), for training vision-language models on X-ray interpretation.

What this is

ROCOv2 spans many imaging modalities. This subset keeps only the X-ray studies, so that a model can be trained on a single modality rather than learning across CT, MRI, ultrasound and radiography at once.

Rows4,254
Splittrain
Size~977 MB
ModalityX-ray only

Columns

ColumnTypeDescription
imageimageThe radiograph
image_idstringROCOv2 identifier, for tracing a row back to the source dataset
captionstringThe original ROCOv2 caption, typically one or two clipped sentences
cuisequence of stringsUMLS Concept Unique Identifiers for the concepts the caption refers to

The cui column is the most useful part for evaluation. Because the concepts are codes rather than free text, they can be compared automatically: extract concepts from a model's generated report and score precision and recall against this set. That measures whether the model named the right findings, which is a different and harder question than whether its prose reads well.

Intended use

Training and evaluating vision-language models on radiograph interpretation. It was built as the training source for the vision stage of a radiology assistant pipeline.

Limitations

  • —Captions are not reports. ROCO captions are short and often describe a teaching point rather than giving a full radiological reading. A model trained directly on them will not produce report-shaped output. See the cycle-1 pilot for one attempt at bridging that gap.
  • —No region-level annotation. There are no bounding boxes or masks. Nothing links a phrase in a caption to a location in the image, so this dataset cannot on its own support grounded or localised report generation.
  • —Inherits ROCOv2's biases. The source is drawn from open-access literature, so the case mix reflects what gets published: unusual and teaching-worthy findings are over-represented relative to routine clinical practice.
  • —Single split. No validation or test split is defined. Make your own, and split by image_id so that studies do not leak between them.

⚕️ Not for clinical use

Research data. Not validated 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.