SkinsAI

Early detection is key for a successful recovery; nonetheless, not all patients are diagnosed at the first stages of melanoma

Melanoma detected at early stages has a 5-year survival rate of 99% [1]; nonetheless, one-fifth of patients are diagnosed at later stages[2]. How can we decrease healthcare barriers and improve early detection?

Enter SkinsAI

SkinsAI stands for Skin Key Imaging Nursing System, and it is a free-access diagnosis tool for classifying moles as benign or malign. SkinsAI is based on a convolutional neural network model. Our purpose is to provide a free, easy-to-use, anonymous and fast tool to give patients a first diagnosis, as well as a website where useful and up-to-date information can be gathered.

How SkinsAI works?

A user uploads a picture of a suspicious mole into the website. The image is fed to our backend convolutional neural network model and returns a diagnosis of whether a mole is malign or benign (80% accurate for the testing data). The user is also referred to updated and relevant information on melanoma prevention and detection.

How was SkinsAI built?

The convolutional neural network model was trained based on the Skin Cancer: Malignant vs Belignant data set, available at Kaggle. We used PyTorch to build the model. The website was built with HTML and CSS using the Bootstrap framework so it would be responsive. Finally, integration of the PyTorch model with the website was performed with the Django web framework.

Challenges we ran into

One of the main challenges we had to face for the construction of this project was the integration of the website and the machine learning model. We had to work using Django, for which we had very little prior experience. We are proud to have a product that could be further developed to serve millions and to help save lives.

What's next for SkinsAI

Some features that could be added in future versions include:

Expansion of the model to also predict other possible skin-related diseases. This will require a more complete dataset, including other types of lesions.
Incorporate a user interface where insurance information could be added, so that patients can be referred to healthcare providers affiliated with their insurance companies. Additionally, free clinics located in a near radius can be suggested as well.
Keep working on the machine learning model to increase its accuracy.