Modelpublic

suitch/radbert-german-ctrate-classifier

sourceHugging Facemitupdated 5mo agoView on Hugging Face

0likes13downloads

Model Card

RadBERT German CTRate Classifier

A RadBERT-based multi-label classifier for predicting 18 pathology labels from German-language radiology reports. The training data consists of German-translated reports from the CTRate dataset, translated using Qwen 3.5 9B.

Model Details

Property	Value
Base model	RadBERT (RoBERTa-base architecture, pre-trained on radiology text)
Task	Multi-label text classification (18 labels)
Language	German (`de`)
Framework	🤗 Transformers + PyTorch
Problem type	`multi_label_classification`

Labels (18 pathologies)

ID	Label
0	Medical material
1	Arterial wall calcification
2	Cardiomegaly
3	Pericardial effusion
4	Coronary artery wall calcification
5	Hiatal hernia
6	Lymphadenopathy
7	Emphysema
8	Atelectasis
9	Lung nodule
10	Lung opacity
11	Pulmonary fibrotic sequela
12	Pleural effusion
13	Mosaic attenuation pattern
14	Peribronchial thickening
15	Consolidation
16	Bronchiectasis
17	Interlobular septal thickening

Quick Start

Installation

bash

pip install transformers torch

Loading the model

python

from transformers import AutoTokenizer, AutoConfig
from modeling_radbert import RadBertForSequenceClassification
import torch

repo_id = "suitch/radbert-german-ctrate-classifier"

# Download the custom model class (or copy modeling_radbert.py locally)
from huggingface_hub import hf_hub_download
import sys, os

modeling_path = hf_hub_download(repo_id=repo_id, filename="modeling_radbert.py")
sys.path.insert(0, os.path.dirname(modeling_path))

# Load config, model, and tokenizer
config = AutoConfig.from_pretrained(repo_id)
model = RadBertForSequenceClassification.from_pretrained(repo_id, config=config)
tokenizer = AutoTokenizer.from_pretrained(repo_id)

model.eval()

Inference example

python

text = "Das Herz ist leicht vergrößert. Es zeigt sich ein kleiner Pleuraerguss links."

inputs = tokenizer(text, return_tensors="pt", truncation=True, padding=True, max_length=512)

with torch.no_grad():
    logits = model(**inputs)

probabilities = torch.sigmoid(logits).squeeze()
threshold = 0.5
predicted_labels = [
    config.id2label[i] for i, p in enumerate(probabilities) if p >= threshold
]

print("Predicted labels:", predicted_labels)
print("Probabilities:")
for i, p in enumerate(probabilities):
    print(f"  {config.id2label[i]}: {p:.4f}")

Training Details

—Base checkpoint: RadBERT (RoBERTa-base weights pre-trained on radiology corpora)
—Training data: German translations of CTRate radiology reports (translated with Qwen 2.5 9B)
—Classification head: Linear layer on top of the [CLS] / pooler output
—Loss: Binary Cross-Entropy with Logits (per-label sigmoid)

Limitations

—This model is trained for label inference from report text only — it does not process images.
—It should not be treated as a clinical decision support system.
—Performance is limited by the quality of the machine-translated training data.

Citation

If you use this model, please cite the CTRate dataset and RadBERT:

bibtex

@article{hamamci2024ctrate,
  title={CT-RATE: A Large-Scale Computed Tomography Report-Image Dataset for AI in Radiology},
  author={Hamamci, Ibrahim Ethem and others},
  journal={arXiv preprint},
  year={2024}
}

@article{yan2022radbert,
  title={RadBERT: Adapting Transformer-based Language Models to Radiology},
  author={Yan, Di and others},
  journal={Radiology: Artificial Intelligence},
  year={2022}
}

License

MIT