SpineySaver

Inspiration

I was checked up for scoliosis by a school nurse to be informed I did not have any for later in high school having to wear braces. It showed me that detecting scoliosis in smaller bodies is incredibly hard as curves are less notictable

What it does

It helps school nurses screen for possible scoliosis using a camera and simple hand tracing. The nurse uses their hand to mark points or trace along the student’s back, and the AI uses those coordinates to more accurately match the visible spine curve. The system then converts that screening into a clear, report-ready result that accurately describes a student's likelihood of developing scoliosis.

How we built it

We built it using OpenCV for live camera capture and MediaPipe to track hand movements so nurses can place points directly on the student’s back. Those hand-drawn coordinates are then used as input for our AI pipeline. To power the curve detection, we started with a scoliosis X-ray model and fine-tuned it to work on surface-level back images, allowing the system to estimate spinal curvature from skin-based visual cues instead of radiographs.

Challenges we ran into

One of our biggest challenges was collecting and preparing enough training data to adapt a scoliosis model from X-ray images to external back images. It was also difficult to make MediaPipe work smoothly for hand-based drawing, since we needed a reliable way for nurses to place and trace points accurately in real time. Balancing usability with model accuracy was a major part of the build process.

Accomplishments that we're proud of

What we learned

We learned that training the model was one of the hardest parts of the project. Fine-tuning an existing scoliosis model for skin-based back images required better data, careful labeling, and repeated testing to improve prediction accuracy. We learned how serious scoliosis is for school nurses, with a 30-40 percent failure rate.

What's next for SpineySaver

Next for SpineySaver is improving the model with more training data and further fine-tuning to increase accuracy on real back images. We also plan to test the MVP with actual school nurses to see how well it works in real screening settings and improve the tool based on their feedback.