A Transfer Learning Based Approach for Skin Lesion Classification from Imbalanced Data

Skin cancer is the most common kind of cancer across the world and some types of skin cancer are deadly if not identified at the early stage. Therefore, it is crucial to detect the lesion class as early as possible. But skin lesion classification is a very challenging task and many automated systems have been developed so far based on different deep learning algorithms. In this study, we have used ResNet, Xception, and DenseNet three state-of-the-art deep learning pre-trained models to classify the skin lesions. For the training and evaluation of our models, we used the HAM10000 dataset and obtained balanced accuracy of 78%, 82%, and 82% for the three models respectively. We then combined the three models using the weighted ensemble technique without any further training and got 85.8% balanced accuracy and this improved other evaluation parameters as well by a significant amount.