Stella X. Yu : Papers / Google Scholar

Improve Image-based Skin Cancer Diagnosis with Generative Self-Supervised Learning
Zhihang Ren and Yunhui Guo and Stella X. Yu and David Whitney
IEEE/ACM Conference on Connected Health Applications, Systems, and Engineering Technologies, Washington D.C., 16-18 December 2021

Skin cancer is the most common malignancy in developing countries. This is largely due to the lack of early detection. The best method for early detection of skin cancer is to track the changes in skin lesions. But, it is hard to implement in developing countries due to the scarcity of experts and their availability in remote areas. Teledermatology provides a promising technology for monitoring skin cancer. Currently, with the involvement of deep learning, teledermatology has become more efficient. However, deep learning, and in par- ticular supervised learning, requires a large amount of data, while collecting and labeling skin lesion images is tedious and requires a high degree of expertise. It is thus expensive to collect enough labeled data to train deep neural networks for skin cancer analysis. Recently, self-supervised learning has proven itself useful for learning representations directly from unlabeled images. Yet, for some rare diseases, e.g. Actinic Keratosis, it is also infeasible to collect enough unlabeled im- ages. In this paper, we utilize Generative Adversarial Network (GAN) to generate synthetic unlabeled images which have high semantic similarity with existing unlabeled medical data. In particular, we evaluate the use of StyleGAN for the data augmentation of skin cancer image self-supervised learning. We utilized StyleGAN to generate new training samples which have the same semantics as the original unlabeled training images. We then combined the new GAN-generated samples with the original unlabeled images as the new training dataset for self-supervised learning. The self-supervised pre-trained network is used as a fixed feature backbone for supervised classification with a limited number of labeled skin cancer images. Quantitative results confirm that our GAN-based data augmentation can boost the accuracy of self-supervised skin cancer image classification by 11.17\% on BCN20000 and 3.07\% on HAM10000.

skin cancer; self-supervised learning; generative adversarial networks; data augmentation