Prabhat9801/Skin-Lesion-Segmentation

🏥 Skin Lesion Segmentation - Deployment Guide

This folder contains everything needed to deploy the UNet++ model to Hugging Face Spaces.

📁 Files Needed

Required Files:

• —✅ app.py - Gradio web application
• —✅ unetpp.py - Model architecture
• —✅ requirements.txt - Python dependencies
• —⚠️ best_model.pt - YOU NEED TO DOWNLOAD THIS
• —⚠️ examples/ folder - YOU NEED TO ADD EXAMPLE IMAGES

📥 Download the Model

1. 1.Go to Google Drive: https://drive.google.com/drive/folders/14-wNH4hWoinkh1I1blsrmf_f9gXcwjyr
2. 2.Navigate to: unetpp_experiments/unetpp_100pct_20251206_183240/
3. 3.Download best_model.pt (~105MB)
4. 4.Place it in this deployment/ folder

🖼️ Add Example Images

1. 1.Create examples/ folder in this directory
2. 2.Add 3-5 sample dermoscopic images:
3. 3.example1.jpg
4. 4.example2.jpg
5. 5.example3.jpg
6. 6.You can get these from the ISIC dataset or use test images from your experiments

🧪 Test Locally

Before deploying, test the app locally:

cd deployment
python app.py

This will:

• —Load the model
• —Start Gradio server
• —Open in browser at http://localhost:7860

Test by uploading an image and checking if segmentation works!

🚀 Deploy to Hugging Face Spaces

Step 1: Create Account

1. 1.Go to https://huggingface.co/
2. 2.Sign up for free account

Step 2: Create New Space

1. 1.Click "New" → "Space"
2. 2.Name: skin-lesion-segmentation
3. 3.License: MIT
4. 4.SDK: Gradio
5. 5.Hardware: CPU basic (free)
6. 6.Click "Create Space"

Step 3: Upload Files

Upload these files to your Space:

• —app.py
• —unetpp.py
• —requirements.txt
• —best_model.pt
• —examples/ folder with images

Step 4: Wait for Build

• —Hugging Face will automatically install dependencies
• —Build takes ~5-10 minutes
• —Watch the logs for any errors

Step 5: Test & Share!

• —Your app will be live at: https://huggingface.co/spaces/YOUR_USERNAME/skin-lesion-segmentation
• —Test it with different images
• —Share the link!

🎨 Customization

Change UI Theme:

In app.py, line 132:

with gr.Blocks(theme=gr.themes.Soft(), ...):

Try: gr.themes.Base(), gr.themes.Monochrome(), gr.themes.Glass()

Add More Metrics:

In calculate_metrics() function, add:

• —Circularity
• —Asymmetry index
• —Border irregularity

Change Colors:

In create_overlay(), line 67:

colored_mask[mask > 0] = [255, 0, 0]  # Red

Try: [0, 255, 0] (Green), [255, 255, 0] (Yellow)

⚡ Performance Tips

For Faster Inference:

1. 1.Use GPU (if available on Hugging Face):

• —Upgrade to GPU hardware in Space settings
• —Free tier has limited GPU hours
• —Model Quantization:

   model = torch.quantization.quantize_dynamic(
       model, {torch.nn.Linear}, dtype=torch.qint8
   )

1. 1.Reduce Image Size:

• —Change resize from 256x256 to 128x128
• —Faster but slightly less accurate

🐛 Troubleshooting

"Model file not found"

• —Make sure best_model.pt is in the same folder as app.py
• —Check file name is exactly best_model.pt

"CUDA out of memory"

• —You're on CPU-only tier (this is normal)
• —Model will use CPU automatically
• —Inference takes 2-3 seconds instead of <1 second

"Module not found"

• —Check requirements.txt has all dependencies
• —Hugging Face will auto-install on build

"Image upload fails"

• —Check image format (JPG, PNG supported)
• —Max size: 10MB
• —Try resizing large images

📊 Expected Performance

🎯 Next Steps

After deployment:

1. 1.✅ Test with various images
2. 2.✅ Share link with friends/colleagues
3. 3.✅ Add to your portfolio/resume
4. 4.✅ Include in GitHub README
5. 5.✅ Get feedback and improve!

📝 Notes

• —Medical Disclaimer: Always included in the app
• —Privacy: Images are not stored, processed in memory only
• —Updates: Push new files to Space to update
• —Monitoring: Check Space logs for errors

🔗 Useful Links

• —Hugging Face Spaces Docs: https://huggingface.co/docs/hub/spaces
• —Gradio Docs: https://www.gradio.app/docs
• —Your GitHub Repo: https://github.com/Prabhat9801/Medical-Image-Segmentation

Ready to deploy? Follow the steps above and you'll have a live demo in ~1 hour! 🚀 Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference