A production-ready deep learning system for classifying lumbar spine degeneration severity from MRI scans (RSNA 2024).
Deployed as an interactive FastAPI web application for clinical exploration and model demonstration.
Live Demo: https://lumbar-degeneration-classifier.onrender.com
Note: This web app is hosted on Render’s free tier.
The site may take up to 30 seconds to load if it has been inactive for a while, as the server spins back up.
Once active, all features will function normally.
This project implements a multi-class deep learning model to assess lumbar spine degeneration across three MRI sequences and multiple spinal regions.
The trained models were deployed using FastAPI with a lightweight web interface designed for clinician usability.
Highlights
- Trained convolutional architectures (CNN, EfficientNetV2, ConvNeXt) on 147,000+ MRI slices from ~2,000 patients.
- Classified degeneration severity (Normal/Mild, Moderate, Severe) for three MRI sequences: Sagittal T1, Sagittal T2/STIR, and Axial T2.
- Achieved 0.95 micro-average AUC and 87% classification accuracy on validation data.
- Deployed via Render as a FastAPI web application with real-time inference and visualization.
| Layer | Technologies |
|---|---|
| Backend | FastAPI, Uvicorn, Python |
| Machine Learning | PyTorch, Torchvision |
| Data Handling | NumPy, Pandas, PyDICOM |
| Frontend | HTML, CSS, Vanilla JavaScript |
| Deployment | Render (CPU environment) |