Residual neural network‐assisted one‐class classification algorithm for melanoma recognition with imbalanced data

Abstract Skin cancer, also known as melanoma, is a deadly form of skin cancer that can significantly improve survival rates when diagnosed at an early stage. It is usually diagnosed visually from dermoscopic images, and such visual assessment of skin cancer by the naked eye is a challenging and arduous task. Therefore, the detection of melanoma from dermoscopic images using trained artificial intelligence models is of great importance today. However, since melanoma is a rare disease, existing databases of skin lesions often contain highly unbalanced numbers of benign and malignant samples. In this paper, we propose a new one‐class classification‐based skin lesion classification strategy for small and unbalanced datasets. One‐class classification (OCC) is a special case of multi‐classification. OCC aims to learn a descriptive paradigm from positive class data (true data) during training and reject pseudo data (fake data) that do not conform to the paradigm during inference. OCC has great potential for application in anomaly detection problems. We have analyzed several approaches to the OCC task in recent years and propose a new design paradigm for the OCC problem, taking into account the unbalanced data set of the melanoma classification task. We have designed an improved OCC network based on this design paradigm, where the network is based on the architecture of a residual neural network, combining the coding and decoding idea of variational self‐encoder and the adversarial training idea of an adversarial neural network, using binary cross‐entropy as the loss function and introducing the channel attention mechanism. Tests on several publicly available dermatology datasets show that this improved OCC network addresses the unbalanced dataset situation in melanoma image classification to some extent while having relatively excellent performance. Compared with some traditional networks, it can obtain more stable training results and perform more consistently on complex datasets.