Malignant melanoma is a skin cancer caused by integumentary cells called melanocytes and has a median survival rate of 5.3 months for patients post metastasis (when the cancer begins spawning in areas outside of the original mole). Identifying melanoma before metastasis increases the probability of successful removal via excision and other minimally invasive techniques. Our project aims to assist medical professionals with making biopsy referrals, which is why our neural network-based classification system is accessible in an app.
What it does
Existing solutions to detect skin cancer (malignant melanoma) are time-insensitive, expensive, and not accessible in underprivileged areas. To address this problem, we developed Melatect, an app that detects potential melanoma using machine learning, while being inexpensive and accessible to any dermatologist with a smartphone.
In addition to providing a diagnosis for a mole as malignant or benign, Melatect includes features such as a risk assessment tool, mole evolution tracker, and doctor search and contact tab- as well as a separate clinical trials interface.
How we built it/what we learned
Creating Melatect allowed us to self-learn new programming languages, softwares, and design values including: (1) applying the engineering and design process, (2) cloud deployments, (3) conducting lengthy and thorough research, (4) increasing our knowledge of complex machine learning (dataset selection/augmentation, algorithms/frameworks, and statistical analysis), (5) utilizing countless Swift libraries (AWS Amplify, relevant Amplify plugins, Realm for local storage, FSCalendar, SQLite, Lottie for animations and visual design), (6) increasing our overall proficiency in languages such as Python, Swift, Java, and HTML.
We also used design tools such as Sketch and Figma to design our prototype skeletons and app screenshots, Lottie Editor to create custom animations for use in Melatect (such as the loading screen), Adobe softwares (illustrate, photoshop, and animate), and flaticon for downloading open-source icons and symbols. Some low-expense ($15/course) asynchronous online courses we used for reference and self-teaching included The Complete iOS App Development Bootcamp and Machine Learning A-Z: Hands-On Python & R in Data Science on Udemy.
Challenges we ran into
Some challenges we ran into were having enough images to train our model with and reach our goal accuracy, so we emailed dermatologists for more mole images. We also want to make sure that our app has no bias, so we augmented our images to include dark skinned individuals, as there was a severe lack of such photos in our dataset. However, this is because darker skinned individuals are more likely to develop carcinoma, which is a skin cancer that is not visible and develops underneath the skin.
Accomplishments that we're proud of
We are very proud of our accuracy rate, being able to incorporate so many features into the app, and also designing it accessible to everyone, unlike existing solutions to detect melanoma.
What's next for Melatect: Identifying Malignant Melanoma in Skin Growths
We plan to add many more features including a freemium model where users can directly send a picture of their mole to a dermatologist for a professional diagnosis; the user would pay a small consultation fee that would go to the dermatologist. We also plan to expand to an Android app, and continually retrain our model to achieve higher accuracies. Finally, we would like to get FDA approval for our app to create a higher level of legitimacy and publish it to the app store.