Generating unlimited skin lesion images with generative adversarial networks

Abstract

Melanoma is the deadliest form of Skin Cancer. Automated skin lesion analysis can play an important role for early detection, which is crucial for patient's treatment. Deep neural networks are the current state-of-the-art for skin lesion analysis, but the lack of annotated data limits classification performance. We started in 2017 to generate annotated data for our classification models with Generative Adversarial Networks (GANs). A GAN can learn the data distribution, allowing us to sample from it, enabling a complementary method for data augmentation. In our previous work, we were able to generate high-resolution, clinically-meaningful skin lesion images that when used to augment a classification network's training data boosted performance. However, our method had limitations. The amount of synthetics we can generate is limited by a special annotation that is available only for a small subset of our training data. This limitation is troublesome because ideally we want to generate infinite amounts of data to aid skin lesion classification. In this Ph.D. proposal, we explore methods for infinite generation: we propose to incorporate relevant clinical information into the GAN generation process, maintaining quality without killing variability and quantity. We believe our experience in generative models and skin lesion analysis qualify this work, enabling us to improve the current state-of-the-art for skin lesion generation and classification. We highlight that our group is at the forefront of such research worldwide being responsible for groundbreaking results associated with skin lesion analysis. (AU)