Inspiration

Skin cancer is the most common type of cancer. Trends show that one out of every five Americans will develop skin cancer over their lifetime.

We wanted to help people monitor their skin and detect cancer early via AI. Using visual clues, artificial intelligence can make faster (and more accurate, granted ample data) predictions than doctors. Collaboration between both AI and doctors can engender greater accuracy as well. Together, this can save time for medical professionals and allow them to focus on more important tasks.

Our model provides seamless implementation into, say, a mobile app or web app for self-diagnosis. Digital platforms also permit the democratization of medical tools (i.e. everyone has greater access to medical tools so say a patient doesn't need to necessarily wait days for an appointment).

What it does

Fleshed Out is primarily a Deep Learning model that classifies seven different types of skin cancers.

It also provides an interactive, educational diagram for people to learn more about the individual layers and parts of the skin.

How we built it

We built our deep learning model with a state-of-the-art architecture —EfficientNet— pretrained on "noisy-student" weights and with heavy data augmentation (e.g. horizontal and vertical flips, random brightness, rotations, shifts, random resized crops, etc).

Challenges we ran into

One challenge we ran into was playing with a brand new library! The PYPI library on Pytorch image segmentation models was completely foreign to me especially since I had just recently learned PyTorch.

Accomplishments that we're proud of

We're proud of how we learned a library so quickly! Usually it takes a while to grow accustomed to it, but with a much more hands-on approach, we found ourselves digging through documentation more curious than ever.

What we learned

For the majority of our team, it was our first ever hackathon. While we dreamed of a huge model pipeline with dozens of ensembles and datasets, we learned how to work with the time allotted for training a limited mode. Overall, we learned that hackathons are fun and competitive and we were able to reap a great deal of invaluable experience.

What's next for Fleshed Out

We plan to build out a much more robust and generalizable ML pipeline that can achieve even greater performance than our current model. This can be furthered in the development of the Streamlit and webapp!

Built With

Share this project:

Updates