A web application that uses deep learning to detect pelvic tilt posture from side-profile photos. Built at MedHacks 2017.
Note: Built on TensorFlow 1.x. The Heroku deployment is deprecated, but the model and code remain as a portfolio piece.
Pelvic tilt is a common posture issue that can lead to:
- Chronic lower back pain
- Hip and knee problems
- Neck and shoulder tension
Most people don't know they have it, and professional assessment typically requires a physical therapist visit.
Upload a side-profile photo → Get an instant classification:
| Classification | Description |
|---|---|
| Anterior Pelvic Tilt | Pelvis tilted forward (most common) |
| Posterior Pelvic Tilt | Pelvis tilted backward |
| Neutral | Normal alignment |
We used TensorFlow's image retraining approach:
- Base model: Pre-trained Inception network (trained on ImageNet's 1M+ images)
- Transfer learning: Retrained the final classification layer on our posture dataset
- Custom dataset: ~60 labeled images across 3 categories
- Output: 5.5MB frozen graph (
retrained_graph.pb)
This approach allows training a custom classifier in minutes on a laptop, leveraging the visual feature representations already learned by Inception.
User uploads photo
↓
Flask web server (server.py)
↓
TensorFlow inference (label_image.py)
↓
Inception-based classifier (retrained_graph.pb)
↓
Classification result → displayed to user
As noted on Devpost:
"Worked on all of the Python (Flask/TensorFlow) and basic HTML for the project"
- Built the Flask web application (
server.py) - Integrated TensorFlow inference pipeline (
label_image.py) - Deployed to Heroku (now deprecated)
- Collaborated with 4 teammates on the 24-hour hackathon build
├── server.py # Flask web server
├── label_image.py # TensorFlow inference
├── retrained_graph.pb # Trained model (5.5MB)
├── retrained_labels.txt # Class labels
├── templates/ # HTML templates
├── static/ # CSS and uploaded images
└── images/ # Training dataset
├── ant/ # Anterior tilt examples
├── post/ # Posterior tilt examples
└── neutral/ # Neutral examples
- Devpost: devpost.com/software/medhack-pelvic-tilt
- Slides: Google Slides Presentation
- TensorFlow Image Retraining Tutorial
- Szegedy et al., "Rethinking the Inception Architecture" (CVPR 2016)
MedHacks 2017 — Wolfram Award Winner