Mole Checker

GIF
Flow of the web app
A bad mole
A good mole

INSPIRATION

Changes of moles are an important indicator for skin cancer. If a melanoma (malignant mole) is detected in the early "localized stage", it is well treatable and the five-year survival rate is as high as 98 %, whereas the survival rate decreases drastically to 15 % if the melanoma is detected in the later "distant stage".

However, is takes a lot of time, money and effort to let doctors frequently check moles of patients. Mole checker has the goal to automatize this analysis, without the need for a doctor, to maximize the chances of early detection.

WHAT IT DOES

Users can take a picture of their moles with their smartphone, and our tool then classifies the mole as "common" (low risk), "atypical" (medium risk), or "malignant" (high risk).

HOW WE BUILT IT

The classifier is implemented as a neural network (basd on IBM's AlchemyVision), which was trained on 256 images of common moles, 256 images of atypical moles, and 128 images of malignant moles. 25 % or the stimulus set were original image files and the rest was generated by rotating original images at different angles.

CHALLENGES WE RAN INTO

finding an appropriate stimulus set (with high quality images and accurate labels)
segmenting the image to isolate the mole (from the surrounding skin and other objects in the background)
automatically resizing and cropping images to focus the stimulus set on the mole
dealing with different light conditions

WHAT WE LEARNED

We learned about the challenges of an image processing pipeline. It was difficult to get the system working robustly. We learned that a good approach is to try simple methods first.

ACCOMPLISHMENTS

We managed to build an end-to-end system, which includes a Front-End app, a webserver and a model on IBM Watson. Also, we collected and preprocessed the available data.