Skin lesion classification based on multi-model ensemble with generated levels-of-detail images

Abstract

Skin cancer has been the top one of global cancer incidence. With regard to that the reliable registration of these cancers has not been achieved yet, the temporal trends of the incidence of skin cancers are difficult to determine. Therefore, using deep learning has become the essential element of the skin lesion classification. Even though, the benign skin lesion is common, patients with malignant skin lesions are reluctant to provide the information, which leads to extremely unbalanced skin lesion datasets, and the deep learning networks tend to classify the testing data into the category with a larger number of images, which is the benign lesion.We propose data balanced methods to solve this data imbalanced problem, including data augmentation, data balancing, and the multi-model ensemble. First of all, we use a generator of the Generative Adversarial Network (GANs) to generate images. Second, we use the generated images to balance the number of images in each category. Finally, we adopt the Deep Neural Networks (DNN) to train the model with the balanced data of different resolutions. Moreover, the ensemble of these model’s prediction value can improve the performance of the mean recall and accuracy. With our proposed method, we are able to achieve the mean recall of 82.1% and 62.5% on ISIC-2018 and ISIC-2019 test sets.