Inspiration
We came up with SkinScan, because skin cancer is the most common type of cancer globally. In places like Miami, with high sun exposure, the risk is even greater. It's more common than you might think.
Many people find themselves looking at a new mole or spot and wondering, 'Is this something serious?'. But a trip to the dermatologist can be expensive, and there are often long wait times, especially for those in underserved communities.
That's why we created SkinScan, an AI-powered tool designed to help address this problem.
What it does and How we built it
This project utilizes deep learning to classify skin lesions from images in the HAM10000 dataset. We preprocessed the data to handle missing values and created distinct training, validation, and test sets. A pre-trained ResNet-50 model was fine-tuned, leveraging a WeightedRandomSampler to address class imbalance during training. The training loop incorporates an Adam optimizer, a learning rate scheduler, and early stopping. Performance is evaluated using a confusion matrix and classification report, providing detailed insights into per-class accuracy.
Challenges we ran into
HAM 10000 is a imbalance dataset, that’s why expanding our dataset to improve accuracy across all skin types is necessary in the future
Accomplishments that we're proud of
We are very happy that we got the model to work with a overall accuracy of 90% that is deployed on Streamlit to use.
What's next for SkinScan
Improving the model accuracy especially for underrepresented skin lesions. This can be done by collecting more diverse data. Moreover, Implementing more front-end features that the use can benefit from. Lastly, we would like to Implement data security for all users.
Log in or sign up for Devpost to join the conversation.