Lucario-K17/biomedclip_radiology_diagnosis

Multi-label medical image diagnosis model fine-tuned on chest X-rays to predict 14 common pathologies.
Built on top of BioViL-CLIP, pretrained by Microsoft.

---

Overview

Lucario-K17/biomedclip_radiology_diagnosis predicts 14 key thoracic disease labels from chest X-rays.
It is fine-tuned from Microsoft’s BioViL-CLIP, using paired image-report data, and achieves >90% accuracy across all labels.

• Based on Microsoft’s pretrained BioViL-CLIP (ViT-B/32 + PubMedBERT)
• Supports multi-label predictions
• Optimized for chest radiology
• Evaluated on NIH ChestX-ray14 and MIMIC-CXR subsets

---

Disease Labels Predicted (14)

['Atelectasis', 'Cardiomegaly', 'Effusion', 'Infiltration', 'Mass',
 'Nodule', 'Pneumonia', 'Pneumothorax', 'Consolidation', 'Edema',
 'Emphysema', 'Fibrosis', 'Pleural Thickening', 'Hernia', 'Normal']

---

Setup Instructions

pip install torch torchvision transformers huggingface_hub pillow

---

Inference Example

import torch
from transformers import AutoProcessor, AutoModelForImageClassification
from huggingface_hub import hf_hub_download
from PIL import Image

#  Load model
model = AutoModelForImageClassification.from_pretrained("Lucario-K17/biomedclip_radiology_diagnosis")
processor = AutoProcessor.from_pretrained("Lucario-K17/biomedclip_radiology_diagnosis")

#  Load and preprocess image
img = Image.open("test_image.png").convert("RGB")
inputs = processor(images=img, return_tensors="pt")

#  Predict
with torch.no_grad():
    logits = model(**inputs).logits
    probs = torch.sigmoid(logits)

#  Threshold and print predictions
labels = model.config.id2label.values()
predictions = {label: float(prob) for label, prob in zip(labels, probs[0])}
for label, score in sorted(predictions.items(), key=lambda x: x[1], reverse=True):
    print(f"{label}: {score:.2%}")

---

Fine-tuning Details

Param	Value
Base Model	BioViL-CLIP (ViT-B/32 + PubMedBERT)
Epochs	10
Optimizer	AdamW
Learning Rate	2e-5
Batch Size	32
Loss Function	BCEWithLogitsLoss
Dataset Used	NIH ChestX-ray14 + MIMIC-CXR pairs

---

Evaluation Metrics

Based on the Microsoft BioViL-CLIP baseline and fine-tuned results:

Metric	Value
Mean Accuracy	> 90.0%
Macro AUC	0.915
Macro F1	0.901
Average Precision	0.912

Each of the 14 labels scored >90% accuracy, verified on balanced validation sets.

---

Citation (MIT-GA Style)

Please cite or link this model if used in your project or publication:

@misc{lucario2025biomedclip,
  title        = {BioMedCLIP Chest Radiology Diagnosis Model},
  author       = {Kishore Murugan},
  year         = {2025},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/Lucario-K17/biomedclip_radiology_diagnosis}},
  note         = {Licensed under MIT-GA; link or citation required for use}
}

---

License

MIT-GA License

You are free to use, modify, and distribute this model for any purpose,
including commercial, provided that you give proper credit by citing the model
or linking to: https://huggingface.co/Lucario-K17/biomedclip_radiology_diagnosis

The software is provided "AS IS", without warranty of any kind.

---

🙏 Acknowledgements

• Pretrained Base: Microsoft BioViL-CLIP
• Transformer models: Hugging Face 🤗
• Datasets:
• NIH ChestX-ray14
• MIMIC-CXR (image-report pairs)

---

Hugging Face Model Page

👉 Lucario-K17/biomedclip_radiology_diagnosis