You can deploy this model and start earning money today!

Lucario-K17/
Radiology_Med

Name: Radiology_Med
Rating: 2.5 (209 reviews)

safetensors

Views

No views yet

Overview

Lucario-K17/Radiology_Med takes structured inputs like diagnosis labels and study metadata, and generates a coherent, detailed radiology report.
It is built on top of Meditron, a medical LLM, and fine-tuned with paired radiology data from MIMIC-CXR.

Task: Text-to-text generation
Input: Diagnosis + Patient metadata (age, sex, view, etc.)
Output: Complete radiology report (findings + impressions)
Quantized 4-bit training on Kaggle (GPU efficient)
Improved fluency & coherence with loss < 0.44

Hardware Requirements

Task	GPU Memory Requirement
Training	~5 GB (Kaggle T4 GPU with 4-bit quantization)
Inference	~14 GB (Kaggle A100 / T4 High-RAM recommended)

Evaluation Metrics

Metric	Value
Training Loss	0.44 ↓ from 0.74
Validation Loss	0.43 ↓ from 0.73
BLEU Score	0.24
METEOR Score	0.54
Total Samples	20,000 reports

Installation

pip install transformers accelerate huggingface_hub

Model Loading & Inference

python

from transformers import AutoModelForSeq2SeqLM, AutoTokenizer

model = AutoModelForSeq2SeqLM.from_pretrained("Lucario-K17/Radiology_Med", device_map="auto", torch_dtype="auto")
tokenizer = AutoTokenizer.from_pretrained("Lucario-K17/Radiology_Med")

# Example input format
prompt = """
Patient Age: 52
Sex: Male
View: PA
Diagnosis: Atelectasis, Pleural Effusion

Generate a detailed radiology report based on this information.
"""

inputs = tokenizer(prompt, return_tensors="pt").to(model.device)

output = model.generate(
    **inputs,
    max_new_tokens=512,
    do_sample=True,
    top_k=50,
    top_p=0.95,
    temperature=0.7
)

report = tokenizer.decode(output[0], skip_special_tokens=True)
print(report)

Sample Output

Input:
Patient Age: 52 | Sex: Male | View: PA | Diagnosis: Atelectasis, Pleural Effusion

Generated Report:
The chest X-ray demonstrates evidence of left lower lobe atelectasis with associated blunting of the left costophrenic angle consistent with a small pleural effusion. Cardiac silhouette is within normal limits. No evidence of pneumothorax. Recommend clinical correlation and follow-up imaging if symptoms persist.

Training Details

Param	Value
Base Model	Meditron 7B
Quantization	4-bit (QLoRA)
Dataset	MIMIC-CXR (20,000 rows)
Optimizer	AdamW
LR	2e-5
Epochs	3
Batch Size	8 (gradient accumulation = 4)
Trainer Framework	Hugging Face Accelerate

Citation (MIT-GA Style)

If you use this model, please provide citation or link:

bibtex

@misc{lucario2025radiologymed,
  title        = {Radiology_Med: MIMIC-CXR Fine-tuned Meditron Model for Report Generation},
  author       = {Kishore},
  year         = {2025},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/Lucario-K17/Radiology_Med}},
  note         = {Licensed under MIT-GA; citation or link required for use}
}

License

MIT-GA License

Permission is granted to use, modify, and distribute this model for any purpose, including commercial, 
provided that appropriate attribution is given by citing or linking to this model: 
https://huggingface.co/Lucario-K17/Radiology_Med

The model and code are provided "AS IS", without warranty of any kind.

🙏 Acknowledgements

EPFL + Hugging Face: Meditron 7B
MIMIC-CXR Dataset
Hugging Face Transformers & Accelerate

🔗 Model Page

👉 Lucario-K17/Radiology_Med on Hugging Face