DERMASCOPE

This project was our team's entry for the Philly Codefest 2025's Advanced category, and we also competed in the "For Social Good" category. We were only given 24 hours in total to work on and finish our project. Although we did not win, I am very proud of our efforts and our product.

Our Resnet-18 model was trained with PyTorch and CUDA using 10,000 images over 50 epochs to classify 7 different types of skin lesions. Using gradient descent and cosine annealing, we achieved an accuracy of around 88%. After training our model, we built a website using FastAPI to receive user images, and produce model images in real-time. Using OpenAI in combination with our model, we were also able to produce accurate diagnoses for patients and doctors. Above is the link to the GitHub page which also contains instructions for setting up and running the application, although it shouldn't work as it doesn't have a key to use OpenAI anymore. Below are images of our team, the pitch for our product, pictures from our finished website, and images for our model.

OUR TEAM

PITCH

Millions around the world lack access to dermatologists, leading to late diagnoses of serious skin conditions. Our AI-powered skin lesion detection app bridges this gap by providing instant, affordable, and accessible skin analysis. Simply upload a picture, and our technology predicts potential concerns—empowering early detection, especially in underserved communities. With this tool, we’re bringing life-saving dermatological insights to those who need them most.
Our model is trained on the HAM10000 dataset, which comprises data on seven different types of skin lesions, curated by medical professionals. We fine-tuned a PyTorch implementation of ResNet18 over 50 epochs, incorporating noise filtering during each training pass to overcome the challenges posed by the small dataset. These efforts have enabled us to achieve an accuracy rate of approximately 90% on our test evaluation.

WEBSITE PICTURES

MODEL IMAGES