Abstract
We developed a deep learning model for classifying dermoscopic images into 7 types of skin lesions. Our method combines transfer learning from VGG-16, pre-trained using the ImageNet dataset, and diagnostic data which consists of age, gender, and location of the lesion in the body.
The best model achieved validation and test accuracies of 77.08% and 78.23%, respectively. Moreover, the model results to an accuracy of 94% on detecting Nevus and 0% accuracy for Dermatofibroma. Both cases can be explained in part by the class imbalance issue of the model.
