MelaScan

Inspiration

Melanoma, the deadliest type of skin cancer, is actually easily treatable if detected at an early stage. However, due to how easy it is to overlook on the skin and how difficult it is to recognize by non-specialized doctors, it often goes undetected until it's too late. So we built a decision support app to help doctors in the diagnosis of melanoma in a very easy to use and rapid fashion, <20 seconds from launch to analysis result.

What it does

The MelaScan app takes a photo of the mole on the patient's skin. With the help of our AI model, it instantly returns a probability of this mole being a Melanoma.

How we built it

We used Flutter to create a lightweight cross-platform app, which then sends the image data to the locally hosted server (in the clinic, for privacy concerns. Could opt-in to global central server). The python back-end runs the inference on the image and sends back the prediction via a simple RESTful API.

Challenges we ran into

One of the biggest challenges we ran into was the extremely imbalanced dataset. Only 1.76% of the whole dataset contains positive examples (melanoma). Another problem also related to the data is the lack of variety, especially little data on people of color, potentially leading to poor performance and bias against minorities. We need more metadata in the future to increase accuracy: such as exposure for UV-Radiation, usage of skin care products, a marker for the size of the mole, hereditary and nodule mole type.

Accomplishments that we're proud of

Intuitive app with low hardware requirements (smartphone with built-in camera) and get the result with just 3 taps within seconds after launching the app. Very high accuracy model. Automatic size measurement and growth history tracking feature coming soon™.

What's next for MelaScan

Since change/growth over time is a major feature to diagnose melanoma, a simple paper marker in the image around the mole to automatically measure the size of the pigment and keep track of its change over weeks/months to actively monitor it. Thanks to the high accessibility of the MelaScan, it can reach general practitioners from more diverse parts of the word. With the data gathered along the way (anonymized and opt-in), we can make a significant contribution to fight dataset imbalance against minorities and diversify the samples.