MelanomIQ

Our Inspiration

“Imagine if a doctor can get all the information she needs about a patient in 2 minutes and then spend the next 13 minutes of a 15-minute office visit talking with the patient, instead of spending 13 minutes looking for information and 2 minutes talking with the patient.” Lynda Chin, Founder of University of Texas MD Anderson Cancer Center

Skin cancer is the most common cancer in the United States and worldwide -- 1 in 5 Americans will develop skin cancer by the age of 70. However, when detected early, the 5-year survival rate for melanoma is 99 percent. We aim to help American diagnose whether or not they have melanoma with our simple app.

What it does

Our app allows users to upload a photo of their skin lesion or mole, which then gets sent to a pre-trained model that detects whether it is cancerous or benign.

So far, we have accounted for eight difference classes: Melanoma (MEL) Melanocytic nevus (NV) Basal cell carcinoma (BCC) Actinic keratosis (AK) Benign keratosis (BKL) Dermatofibroma (DF) Vascular lesion (VASC) Squamous cell carcinoma (SCC)

Users also upload their symptoms, location, demographic data, and location of the lesion. Based on the sentiment of the symptom description, the app either suggests urgent care or dermatology facilities.

How we built it

We build the cancer detection model with Google's AutoML Computer Vision models, feeding in 25,000 images of labelled cancerous lesions. We used AFINN for symptom sentiment analysis, which specializes in word-list based sentiment analysis for small chunks of text. We also used the google maps API to find dermatologists and urgent care closest to us. The app itself was build using Vue.js, with a Flask app interfacing between Vue.js and external APIs.

Challenges we ran into

We had a challenge collecting additional samples of healthy, non-cancerous skin to supplement our dataset. Also, we had to level out data by reducing unbalanced classes and using data augmentation to boost other classes. (Methods involved: Rotated Images, Flipped Images, Changed Image Brightness, Filled blank spaces as a result of augmentation). It was challenging to find appropriate sentiment analysis for small chunks of text. Keeping the UI/UX pleasant while garnering extremely specific data.