GeMix is a new data augmentation technique proposed to overcome the limitations of existing Mixups in fields where realistic image generation is important, such as medical image classification. Instead of the simple pixel-by-pixel interpolation of existing Mixups, it performs learning-based interpolation that considers label information using a class-conditional GAN (StyleGAN2-ADA). Two label vectors are sampled from the Dirichlet distribution, mixed using the Beta distribution coefficient, and then given as a condition to StyleGAN2-ADA to generate visually consistent images. Experimental results using the COVIDx-CT-3 dataset show that it achieves a higher macro-F1 score than existing Mixups, and in particular, it reduces the false negative rate in COVID-19 detection. It can be easily applied to existing learning pipelines, and the code is also open to the public to increase reproducibility.
