Inspiration

We wanted to create a deep learning platform that could detect and diagnose burns of varying severity. This app has a smooth viewing experience on iPhone, Android, and any browser, making it easily accessible to nurses or doctors who may need to assess patient burn injuries. Firefighters may find a use for it as well, considering the recent climate-fueled wildfires.

What it does

We allow users to upload an image of their burn, we analyze the severity of the injury, and then return a diagnosis and treatment recommendations.

How we built it

We trained a neural network on datasets of burn injuries using Scikit and Keras on a Jupyter Notebook executing on IBM Z Mainframe so that it could assess, to a relative degree of accuracy, what kind of burn the patient has, based on three main criteria: color, blistering, and breadth. Using these factors, we say whether the burn is categorized as superficial dermal (first degree), partial thickness (second degree), or full thickness (third degree). We hosted our associated web app on IBM Cloud Foundry with a domain provided by Domain.com.

Challenges we ran into

Due to the lack of accessible burn wound datasets, we had to scrape the web manually to compile the datasets we used to train our models on. Additionally, due to constraints on the IBM Mainframe sandbox being airgapped, we had to figure out how to export our trained model to be able to use it in production.

Accomplishments that we're proud of

We are proud of the image processing modules we implemented to process the images before passing them into the neural network, because we were able to dramatically increase both the train and test accuracy of our predictions.

What we learned

Deep learning models with Scikit and Keras, web development with Bootstrap, Flask, Postman API requests, AJAX, and compiling datasets.

What's next for Unfeel the Burn

Using a convolutional neural network within the flow of the product.

Built With

Share this project:

Updates