Malignant Melanoma Detector

Prediction result of uploading

Inspiration

Early skin cancer detection can be life-saving, but access is limited. We wanted to build an AI tool that makes diagnosis more accessible-starting with just a photo.

What it does

Our web application allows users to upload an image of a nevus to determine the probability it is either benign or malignant.

How we built it

We built the application as a Flask web application using AWS as the cloud server to host the ResNet18 model we trained.

Challenges we ran into

Throughout the development process, we encountered several challenges:

A highly imbalanced dataset, with significantly more benign cases than malignant ones.
Difficulty interpreting feature maps from the CNN, which complicated explainability.
The technical complexity of managing and preprocessing large-scale image and metadata pipelines.

Accomplishments that we're proud of

We are proud of:

Our model reached a validation accuracy of 96.75%.
Successfully completed full-stack development, integrating both model training and deployment with a user-friendly interface.
Our data exploration revealed clinically relevant insights that deepened our understanding of the problem space.

What we learned

Throughout this project, we learned:

Training and optimizing convolutional neural networks for medical image classification.
Leveraging metadata to complement image-based models and provide richer insights.
Building interpretable machine learning systems that enhance user trust in model predictions.

What's next for Malignant Melanoma Detector

Looking ahead, we plan to:

Scale the application to include the full dataset.
Incorporate multimodal learning by combining image and metadata features more holistically.
Improve fairness across demographic groups such as age and sex.
Develop enhanced interpretability tools to make the model’s decision process more transparent and clinically useful.