Inspiration
Everybody has already noticed a weird looking mole on their body but did not want to take the time to book a doctor appointment as it might be nothing and it is time consuming. We wanted to develop a quick and easy solution to this problem by giving the option of a quick check up in a matter of minutes.
What it does
Take a picture of the mole you are worried about, upload it and let our image processing model do the work ! It will tell you whether you should consult a doctor or if it is perfectly healthy.
How we built it
We use a pre-trained image processing model (VGG16) and fine tuned it to be able to analyze moles.
Challenges we ran into
For a good amount of time, the accuracy of the model was stagnating around 74%. We had to do extensive research on data augmentation and play around with how we were pre processing the training data in order to make the model as robust as possible.
Accomplishments that we're proud of
With a final accuracy of more than 86%, we overcame all the issues that we ran into one by one. It was a very rewarding journey !
What we learned
Basically everything about training a model, doing transfer learning, and playing around with the different choices of learning rates, optimizer, loss function, but also where to do the model training as image processing models take very long to train on the CPU!
What's next for No Mole Problem
We would like to make it as accessible as possible as well as implementing new features like a more detailed diagnostic and not just a binary output.
Log in or sign up for Devpost to join the conversation.