ABSTRACT
Skin cancer is a severe form of cancer that can easily be curaed when treated at an early stage. However, skin lesion images usually have an intrinsic visual similarity that is difficult for dermatologists to differentiate and diagnose skin cancer. An automatic system to classify skin cancer can ease the workload of dermatologists and improve the efficiency of the skin cancer diagnosis. In this chapter, we propose an automatic system using deep learning techniques to classify multiple types of skin lesion images. Our system uses data augmentation to enhance the skin cancer dataset. Then, a pre-trained deep learning model, named Xception, is fine-tuned to extract image features of skin lesions. Transfer learning is utilized by freezing the first 36 layers and re-training the last 35 layers of the pre-trained model. We remove the pre-trained model's last layer and replace it with a dense layer to classify eight types of skin lesions. We evaluated our system on the publicly available ISIC 2019 dataset. The best performances are 95.96%, 71.98%, 70.50%, 96.64%, and 64.3% for accuracy, precision, recall, F1-score, and BMA score, respectively. Thanks to these results, we recommend using Xception model for skin cancer classification.
