Hack USF SKinder

IMPORTANT FOR RUNNING APP

username: [email protected] password: password

Inspiration

Team members are directly affected by skin lesions and have family members who suffer from skin cancer. Early detection can be a crucial thing in their lives for prevention of skin cancer as it is a fast moving cancer that can metastasize before it can be stopped.

We also have aspirations to work with Moffitt cancer data analytics and feel that this would be a great step in the right direction to attaining internships/jobs with that company.

What it does

Skinder is an AI-powered tool that analyzes images of skin lesions and classifies them as either benign or malignant. It serves as a first-line screening tool that can alert users to potentially dangerous skin conditions that warrant professional medical attention, effectively democratizing preliminary skin cancer screening.

How we built it

We developed SKinder using a transfer learning approach combined with gradient boosting:

First, we used a pre-trained ResNet50 neural network as a powerful feature extractor. This network, already trained on millions of images, acts as the "eyes" of our system, identifying 2,048 distinct visual patterns in each skin lesion image. These extracted features were then normalized using StandardScaler to ensure all features contribute equally to the classification decision. For the classification stage, we implemented an XGBoost model with carefully tuned parameters to optimize performance. We addressed the class imbalance between benign and malignant samples by calculating and applying appropriate class weights. Finally, we used ROC curve analysis to determine the optimal classification threshold instead of the standard 0.5 cutoff, balancing sensitivity and specificity for better medical relevance.

Challenges we ran into

1. Implementation of the AI
2. Fine turning of the models that work in tandem to squeeze out more accuracy

Accomplishments that we're proud of

1. Achieving over 85% accuracy using a computationally efficient two-stage approach that could feasibly run on consumer hardware.
2. Successfully implementing a transfer learning solution that leverages the power of deep convolutional neural networks without requiring massive labeled datasets.
3. Creating a balanced classification system that considers both false positives and false negatives, crucial in medical applications.
4. Developing a model that works directly with image data with minimal preprocessing requirements.
5. Building a system with real potential to make a positive impact on healthcare accessibility.

What we learned

Machine Learning is fun!

What's next for Hack USF Skinder

Further research into machine learning and AI models for implementation of future projects. We can expand our dataset to include more diverse skin types and lesion varieties to improve model generalizability across different demographics. As well as explore multiclass classification to identify specific types of skin conditions beyond the binary benign/malignant distinction

Built With

• ai
• css
• html5
• javascript
• json
• machine-learning
• python
• typescript