Skin.AI

Inspiration

My grandfather passed from Oral Cancer, and we first did try to find datasets to look at Oral Cancer, however there didn't seem to be much open source data for us to target it. So we chose to conquer Skin Cancer, but more than that, we chose to classify multiple skin conditions as millions of people have low access to healthcare and our product can provide a preliminary care to people all over the world.

What it does

Our website allows individuals to upload pictures of themselves and our Residual Network would classify which skin disease is plaguing the person. Our Large Language Model takes the output of our Vision Model and gives recommendations to the persons.

How we built it

We built our product through Streamlit for our front end. Our LLM was built using the Google PaLm API and we selected the Text Bison model. This essentially acts as our backend for our web app.

Challenges we ran into

Some of the challenges we ran into for this project were API utilization, figuring LLM's to use, finding LLM's that fit in our compute capacity, machine hardware difficulties (for training/testing), and issues with figuring out how we would deploy our models.

Accomplishments that we're proud of

While we had some rough times throughout the life cycle of this project, it was an awesome experience to code twenty hours straight even when the computer wasn't the most agreeable at times.

What we learned

We learned how to use Git effectively and how to ensemble heavy duty models together to deliver a product that can make a change in the world. We also learned how to persevere through the late nights when you feel like you can hear your bed speaking to you, but you resist the urge to sleep in order to quench your thirst for exploration.

What's next for Skin.AI

We plan to transfer our Vision Model to a Vision Transformer like we originally intended to, sadly we did not have enough time. We would also like to use a larger LLM like LLaMa-2 to experiment on different LLMs.