RadGraph graph inference for Italian radiology reports, with output in the canonical RadGraph-XL format.

This repository hosts the v1 Italian DyGIE model used by the RadGraph-IT package. It jointly extracts named entities and relations from Italian radiology reports using a shared encoder, span extractor, NER head, and relation head. The encoder is IVN-RIN/medBIT-r3-plus.

Daniel Rabottini, Edoardo Avenia, and Rocco Angelella.

The training corpus contains 2,850 manually revised Italian projections of RadGraph-XL and RadGraph 1.0 reports. Its provenance, annotation schema, data format, and 50-report excerpt are documented in the radgraphIT/RadGraph-IT dataset card.

model.predict(text) applies the word-level preprocessing used during training (regex cleanup and nltk.wordpunct_tokenize). If you already have a list of words tokenized with that exact scheme, use model.predict_words(words) instead.

The bundled tokenizer (AutoTokenizer.from_pretrained("radgraphIT/Radgraph-IT")) is the encoder's subword tokenizer; it is not the word-level splitter used by predict().

The model requires transformers, torch, numpy, and nltk (for predict()). The radgraphit package provides the supported Python and CLI interface, including RadGraph-XL-compatible serialization.

Word-level embeddings are produced by mean-pooling medBIT-r3-plus subwords back to one vector per word. Documents longer than 512 tokens are split into word-aligned chunks. Span representations concatenate start/end word embeddings with a span-width embedding. The NER head scores candidate spans of width up to 12; the relation head prunes to the top spans-per-word and scores each retained pair.

Entity labels are Anatomy and Observation, each with definitely present, definitely absent, or uncertain status. Relation labels are modify, located_at, and suggestive_of.

Generalization was estimated through report-level five-fold cross-validation over all 2,850 reports. In each outer fold, 570 reports were reserved for testing, while the remaining 2,280 were divided into 2,052 training and 228 development reports. Iterative multilabel stratification balanced the data source and the document-level presence of the six entity classes and three relation types. The development partition was used exclusively for early stopping and checkpoint selection through JointF1, defined as the unweighted mean of entity and relation micro-F1. Each report therefore received exactly one out-of-fold prediction.

Results are reported as fold-level means ± sample standard deviations. The released checkpoint was then trained on all 2,850 reports for 13 epochs, the median validation-optimal duration across folds. training_metrics.json contains training diagnostics only.

RadGraph-IT is a research software model for automatic information extraction. It is not a medical device; its output is not a diagnosis, report, or clinical opinion. Validate outputs for the intended setting and keep human clinical judgement in the loop.

Please cite both RadGraph-IT and RadGraph-XL when using this model:

For complete methodological details, per-class results, and comparisons with RadGraph 1.0 and RadGraph-XL, please refer to the full RadGraph-IT paper.