Skin Cancer Diagnotsic Using Machine Learning

Inspiration

Having studied AI theory in the news and in research papers, I thought BostonHacks was a great opportunity to advance my skills in the application of machine learning. As I contemplated what algorithm to use, I figured transfer learning was a great approach to the subject in the limited time of one day. After listening to the sponsors' talks, I came up with the idea to apply machine learning to healthcare. My first idea was to create an app that could diagnose any skin disease by taking a picture. But after surfing through datasets of diseases, I realized that there were not enough images to do so. Then, I came across melanoma, a skin cancer that can become deadly if not diagnosed early enough. Realizing this, I couldn't back down from the challenge.

What it does

The machine learning algorithm takes in input images of moles and then classifies whether or not a malignant form of melanoma exists. After training the model, it can be used to detect if a mole has developed, or may develop, into melanoma.

How I built it

I used the machine learning library, PyTorch, to develop a deep neural network on a Linux virtual machine. Because images have high dimensionality, it is extremely power and time consuming to train a deep neural net solely on such. So I used transfer-learning, which is a method of using a pre-trained model on millions of images, to speed up the process and to allocate more to develop more specific differentiation. The neural network is

Challenges I ran into

I tried installing Pytorch on windows, however, the unofficial build caused many problems and broke other dependencies. So I sacrificed the benefit of GPU acceleration to build a more stable application on a slower virtual machine.

Accomplishments that I'm proud of

I am proud that I have gotten my first real-world application of machine learning working. I had previously struggled a lot with deep learning libraries because of their rapidly changing architecture.

What I learned

I strengthened my knowledge in Python throughout my development in machine learning. And boosted my confidence to apply machine learning to a broader spectrum of things.

What's next?

It could definitely gain more accuracy by training on a larger dataset and improve by adding the identification of more skin conditions. It would be great to build an app that automatically identifies the diseases with only a picture. It could also suggest ways to prevent, mitigate, or reach help about a specific skin condition.

Built With

• machine-learning
• python