A classifier written in native Keras and converted to TensorFlow.js that can be used to classify seven different types of skin lesions.
This model uses the MobileNet architecture and was retrained on the HAM10000 dataset which contains approximately 10,000 documented images of skin lesions. https://arxiv.org/abs/1803.10417
The goal of the project was to deploy the final model as a web application. The live version can be found at: https://alexyu.ca/projects/Skin-Lesion-Classifier/
The model obtains an accuracy of approximately 80%, with a top-3 accuracy of approximately 94%.