Lung Cancer Hybrid Model

Hybrid 3D lung nodule classification pipeline with a fusion architecture and deployment-ready inference app.

Live Links

• Hugging Face Model: https://huggingface.co/ShiroOnigami23/lung-cancer-hybrid-model
• Hugging Face Space (Web App): https://huggingface.co/spaces/ShiroOnigami23/lung-cancer-hybrid-app

Latest Results

• Test Accuracy: 84.19%
• Test AUC: 0.9202
• Test F1: 0.6918
• Dataset used for benchmark run: MedMNIST NoduleMNIST3D (derived from LIDC-IDRI)

Project Structure

app.py
Dockerfile
requirements.txt
requirements-app.txt
data/
models/
training/
evaluation/
inference/
scripts/

Recommended Framework

For fastest iteration + shareable demo, use:

• Gradio + Hugging Face Spaces (already set up)

Reason: easiest deployment flow, direct model hosting, and instant public/private demos with zero custom backend maintenance.

Run App Locally

pip install -r requirements-app.txt
python app.py

Then open http://localhost:7860.

Run With Docker

docker build -t lung-cancer-hybrid .
docker run -p 7860:7860 lung-cancer-hybrid

Train

python train.py --epochs 2 --batch-size 2 --num-workers 0 --mock-data

For real-dataset training in Colab, use scripts under scripts/.

Notes

• The app expects a .npy 3D nodule patch input.
• app.py loads weights directly from the Hugging Face model repo.