raphael-r/bright-eds-histology

BRIGHT NER: EDS-NLP (CamemBERT + CRF) fine-tuned for histology

Description

This is a EDS-NLP (CamemBERT + CRF) architecture fine-tuned to extract clinical neuro-oncology entities related to the histology semantic group. It was trained on a synthetic dataset generated for the properly de-identified BRIGHT project dataset (see the generated_data folder in the primary repository).

This model repository was specifically designed to fit within the bright_db overarching namespace.

Fields

It extracts the following fields (described in French):

• histo_necrose: Présence nécrose
• histo_pec: Prolifération endothéliale/microvasculaire
• histo_mitoses: Nombre mitoses ou index mitotique
• aspect_cellulaire: Aspect cellulaire (astrocytaire, oligodendroglial)

Performance on Validation Set

Aggregates:

• Macro F1: 0.6069 (Precision: 0.5036, Recall: 0.9521)
• Micro F1: 0.6445 (Precision: 0.4913, Recall: 0.9365)

Per-Label Breakdowns:

Label	Precision	Recall	F1
histo_necrose	0.2732	1.0000	0.4292
histo_pec	0.2011	1.0000	0.3348
histo_mitoses	0.8424	1.0000	0.9145
aspect_cellulaire	0.6978	0.8083	0.7490

Usage

# Inference Code
import edsnlp

nlp = edsnlp.load("raphael-r/bright-eds-histology")
doc = nlp("Patient presenting with epileptic seizures...")

for ent in doc.ents:
    print(ent.text, "=>", ent.label_)