Classification of skin lesions with generative adversarial networks and improved MobileNetV2

Abstract Malignant skin lesions pose a great threat to patients' health, and the use of computer algorithms for automatic skin medical image classification can effectively improve the efficiency of clinical diagnosis. However, the existing methods for skin classification have complex models and are greatly affected by the imbalance of the dataset. In this work, we propose a two‐stage framework called G‐DMN, it uses CycleGAN to expand the dataset and Dense‐MobileNetV2 (DMN) to achieve the automatic classification of skin lesion images. In the first stage, we use CycleGAN for data augmentation and propose a new image pairing strategy for training. Image pairs are formed from majority class images and minority class images, generators are trained for majority to minority class image conversion, and then minority class images are generated to balance the dataset. In the second stage, we propose a lightweight model called DMN by improving MobileNetV2, it enhances feature reuse by increasing the width of the network and allows the network to focus on focal areas from different scales. The original training set combined with the generated images is used to train DMN for skin lesion classification. We tested the proposed model on the HAM10000 dataset, and the G‐DMN achieved 87.07% classification accuracy, 80.13% precision, 75.28% sensitivity, 96.19% specificity, 77.26% F1‐Score and 0.952 AUC, which has a good classification effect, while the number of parameters of the model is only 5.33 M, which is much lower than other classical classification models. We demonstrate that the proposed method is lighter and more effective than classical classification methods, achieving significant performance improvements.