🩻 ChestXray-BLIP Report Generator
AI-powered Chest X-ray → Radiology Report Generation
🚀 Overview
ChestXray-BLIP Report Generator is a Vision–Language deep learning model fine-tuned to automatically generate radiology-style chest X-ray reports from medical images.
The model is trained on the large-scale NIH ChestX-ray14 dataset (~45GB) and designed for medical AI research, education, and experimentation.
It is also integrated into an end-to-end medical application with an AI chatbot interface.
⚠️ Disclaimer: This model is not clinically approved and must not be used for real-world diagnosis.
🌟 Key Features
🩺 Automated chest X-ray report generation
🧠 Vision–Language architecture (BLIP-based)
🖼️ Image → Text medical understanding
💬 Compatible with AI medical chatbots (Qwen-based systems)
⚡ FP16 / mixed-precision support
Trained on kagggle 2 T4 GPUs
🧠 Model Details
| Attribute | Description |
|---|---|
| Task | Chest X-ray → Radiology Report Generation |
| Model Type | Vision-Language Model |
| Base Architecture | BLIP (Bootstrapped Language-Image Pretraining) |
| Framework | PyTorch + Hugging Face Transformers |
| Vision Encoder | ViT (BLIP) |
| Text Decoder | Transformer-based |
| Training Dataset | NIH ChestX-ray14 (~45GB) |
| Training Epochs | 3 |
| Precision | FP16 supported |
| Language | English |
🧬 Architecture Overview
Chest X-ray Image ↓ Vision Encoder (ViT / BLIP) ↓ Cross-Modal Attention ↓ Text Decoder ↓ Radiology-Style Report
📦 Model Files
best_model/ ├── config.json ├── model.safetensors ├── generation_config.json ├── preprocessor_config.json ├── tokenizer.json ├── tokenizer_config.json ├── vocab.txt └── special_tokens_map.json
📥 Usage
🔧 Load the Model
from transformers import BlipProcessor, BlipForConditionalGeneration
processor = BlipProcessor.from_pretrained(
"anassaifi8912/chestxray-blip-report-generator"
)
model = BlipForConditionalGeneration.from_pretrained(
"anassaifi8912/chestxray-blip-report-generator"
)
🖼️ Generate a Report
inputs = processor(image, return_tensors="pt")
outputs = model.generate(**inputs, max_length=128)
report = processor.decode(outputs[0], skip_special_tokens=True)
print(report)
📊 Evaluation Metrics
Evaluated on unseen test data only.
Metric Score
BLEU-1 0.1019
BLEU-2 0.0692
BLEU-3 0.0341
BLEU-4 0.0189
METEOR 0.1692
ROUGE-L 0.1803
🏥 Clinical Accuracy 0.3159
🧪 Intended Use
✅ Medical AI research
✅ Academic & educational projects
✅ Prototyping healthcare applications
❌ Not for clinical diagnosis or treatment
⚠️ Limitations
Generated reports may be synthetic or incomplete
No disease localization or bounding boxes
Single-image input only
Performance varies across disease categories
Requires expert human verification
🔐 Ethical Considerations
Outputs are AI-generated
May contain hallucinations or inaccuracies
Trained on publicly available medical data
Not evaluated for demographic or clinical bias
🏥 Medical Disclaimer
This model is not a medical device and is not FDA-approved.
It must not be used for real-world clinical diagnosis, treatment, or medical decision-making.
📚 Dataset
NIH ChestX-ray14 Dataset
Publicly available via NIH & Kaggle
Over 112,000 chest X-ray images
14 disease labels
👤 Author
Anas Saifi
AI / Data Scientist
🔗 GitHub: https://github.com/anassaifi775
🔗 Hugging Face: https://huggingface.co/anassaifi8912
🔗 LinkedIn : https://www.linkedin.com/in/mohd-anas-6570a6290/
⭐ Acknowledgements
NIH Clinical Center
Kaggle
Hugging Face 🤗
PyTorch Community
⭐ If you find this model useful, please give it a star on Hugging Face
- Downloads last month
- 100