🧠 About the Project
SkinSight AI is a machine learning-based tool developed to assist in the early detection and classification of skin cancer. It leverages computer vision and clinical data to differentiate between benign and malignant skin lesions. Our goal was to create a fast, accessible, and non-invasive solution that can support healthcare professionals and help raise awareness among the general public.
The project uses publicly available images from the NIH Skin Cancer Dataset, along with basic clinical metadata, to train a deep learning model capable of accurate classification. As a secondary aim, the system is designed to explore the potential for further subcategorization into specific types of skin cancer.
💡 What Inspired Us
Our inspiration came from the urgent need for early diagnosis of skin cancer, especially in resource-limited settings where access to dermatologists is scarce. We were also motivated by the power of AI in transforming healthcare and its growing use in diagnostic imaging.
The idea that a simple image of a skin lesion—when combined with AI—could help flag potential cancer cases early on was both powerful and exciting. We wanted to build something that could one day be used in real-world applications, saving lives through early intervention.
🧪 What We Learned
This project taught us valuable lessons in both technical and medical domains:
- The importance of balanced datasets, especially in healthcare where false negatives can be critical.
- How to process and integrate multimodal data (images + clinical info) into a cohesive AI pipeline.
- The effectiveness of data augmentation in improving model generalization.
- How AI interpretability (e.g., heatmaps, Grad-CAM) is vital for building trust in medical tools.
- The real-world responsibility that comes with developing AI for healthcare.
⚙️ How We Built It
We approached the project in the following stages:
Data Preparation
- Collected and cleaned the NIH skin cancer dataset.
- Preprocessed lesion images (resizing, normalization).
- Processed clinical metadata (age, sex, lesion location).
- Collected and cleaned the NIH skin cancer dataset.
Model Development
- Built a Convolutional Neural Network (CNN) using TensorFlow/Keras for image classification.
- Integrated a parallel model branch for clinical features using dense layers.
- Combined the outputs of both models into a final classification layer.
- Built a Convolutional Neural Network (CNN) using TensorFlow/Keras for image classification.
Training & Evaluation
- Used data augmentation to handle class imbalance.
- Applied techniques like early stopping, dropout, and validation splits for optimization.
- Evaluated performance with metrics like accuracy, precision, recall, and F1-score.
- Used data augmentation to handle class imbalance.
⚠️ Challenges We Faced
- Data Imbalance: Malignant cases were significantly fewer than benign ones, requiring careful balancing techniques.
- Multimodal Integration: Combining image and clinical data in one model architecture required experimentation and tuning.
- Image Quality Issues: Some lesion images were unclear, affecting the learning process.
- Time Constraint: Developing a healthcare-related AI model with meaningful accuracy in a short timeframe was challenging but pushed us to stay focused and efficient.
🌟 Final Thoughts
SkinSight AI reflects our belief in using technology to tackle real-world health challenges. While this model is still a prototype, it lays the foundation for future development into a more robust, clinically validated tool. We hope this project contributes to the broader mission of accessible and early cancer detection through AI.
Log in or sign up for Devpost to join the conversation.