ACStyleGAN for Skin Lesion Image Generation

Medical datasets often suffer from class imbalance due to the scarcity of obtaining data for certain classes. This can hinder performance when attempting to build a good classification model based on such imbalanced datasets. Data augmentation methods have been used to mitigate this issue. Generative Adversarial Network (GAN) based data augmentation methods have recently become popular in this aspect. The ISIC 2018 image dataset for skin lesions also suffers from class imbalance. Various data augmentation methods such as the GAN, Conditional Generative Adversarial Network (CGAN) and Skin Lesion Style-Based Generative Adversarial Network (SL-StyleGAN) have been used as data augmentation techniques. In this research synthetic images of skin lesions were generated using an Auxiliary Classifier Generative Adversarial Network (ACGAN) and Progressive ACGAN based approaches. The final model used elements of the Style-Based Generative Adversarial Network (StyleGAN) to build an Auxiliary Classifier Style-Based Generative Adversarial Network (ACStyleGAN) architecture. This approach differs from the original StyleGAN implementation by incorporating class conditioning to the generator and further in the order of applying the Adaptive Instance Normalization (AdaIN) and noise injection. The ACStyleGAN augmentation method enhances the overall performance of the ResNet50 classifier in all metrics namely accuracy, sensitivity, specificity and recall in comparison to using the original dataset.