EfficientNet-Based Model With Test Time Augmentation for Cancer Detection

Skin cancer is the most common type of cancer and among various kinds of skin cancer, melanoma causes the most deaths. In clinical practice, contextual information from every one of a patient's moles help dermatologists make better judgments about whether a particular one is a lesion. In this paper, we proposed an EfficientNet based deep learning method to identify melanoma in skin lesion images. Our method takes into consideration all skin lesion images from a patient and employs effective data augmentation during training and test time augmentation during inference to improve classification accuracy. On the SIIM-ISIC Melanoma Classification dataset, our method achieved 0.901 Auc-Roc scores, outperforming other deep learning models such as VGG16 or Resnet50.