---
pretty_name: RadGraph-IT
language:
- it
task_categories:
- token-classification
size_categories:
- n<1K
tags:
- radiology
- clinical-nlp
- named-entity-recognition
- relation-extraction
- radgraph
- italian
configs:
- config_name: sample
  data_files:
  - split: train
    path: data/radgraph_it_50.jsonl
---

# RadGraph-IT Dataset

## Dataset summary

RadGraph-IT contains 2,850 Italian radiology reports with entity and relation
annotations in a DyGIE-compatible JSONL format. The corpus combines 2,300
reports from [RadGraph-XL](https://aclanthology.org/2024.findings-acl.765/)
with 550 reports derived from
[RadGraph 1.0](https://physionet.org/content/radgraph/1.0.0/).

This repository provides a 50-report excerpt of the corpus used for training.
The complete RadGraph-IT corpus is not currently available for publication.
The full corpus includes 1,050 chest X-ray reports, 600 chest CT reports, 600
abdomen/pelvis CT reports, and 600 brain MRI reports.

## Translation and manual review

All 2,850 English reports were localized with the English-to-Italian pipeline
developed for the thesis *AI-Driven Keyword Extraction and Structuralization
of Medical Reports* and released in the
[`EIDOSLAB/radiomicslab-keyword-extraction`](https://github.com/EIDOSLAB/radiomicslab-keyword-extraction)
repository.

After translation and span remapping, the localized reports and projected
annotations were manually reviewed. The Italian text, entity boundaries,
entity labels, and relations were corrected where necessary.

## Annotation schema

Each entity is a contiguous, inclusive token span assigned one of six labels:

- `Anatomy::definitely present`
- `Anatomy::uncertain`
- `Anatomy::definitely absent`
- `Observation::definitely present`
- `Observation::uncertain`
- `Observation::definitely absent`

Entities are connected by three directed relation types:

- `located_at`: an Observation and its anatomical site;
- `suggestive_of`: an Observation and another observation that it suggests;
- `modify`: two Observations or two Anatomy entities when one qualifies the
  other.

## Data format

The included [`data/radgraph_it_50.jsonl`](data/radgraph_it_50.jsonl)
uses the same representation consumed by the training stack. Each line is one
JSON object with:

- `dataset`: a provenance-aware RadGraph-IT source and modality label, such as
  `radgraph-it-stanford-chest-x-ray`;
- `doc_key`: the numeric index from the original source subset; the pair
  `(dataset, doc_key)` uniquely identifies and traces each report;
- `sentences`: a list containing the tokenized report;
- `ner`: entity tuples `[start, end, label]`, with inclusive token indices;
- `relations`: directed relation tuples
  `[source_start, source_end, target_start, target_end, label]`, with inclusive
  token indices.

The entire report is represented as one DyGIE "sentence" so relations can
connect entities anywhere in the report.

## Sample distribution

The 50-report excerpt is distributed across the four imaging modalities as follows:

| Modality | Sample reports |
|---|---:|
| Chest X-ray | 18 |
| Chest CT | 11 |
| Abdomen/pelvis CT | 11 |
| Brain MRI | 10 |
| **Total** | **50** |