Assessment and Analysis of Fidelity and Diversity for GAN-based Medical Image Generative Model

Recently, various researches on medical image generation have been suggested, and it becomes crucial to accurately evaluate the quality diversity of generated images. For this purpose, expert’s visual turing test, feature distribution visualization, quantitative evaluation through IS FID are evaluated. However, there few methods for quantitatively evaluating images in terms fidelity diversity. In paper, by learning a chest CT dataset non-small cell lung cancer patients DCGAN PGGAN generative models, performance two models evaluated The is FID, which one-dimensional score-based methods, Precision Recall, Improved two-dimensional characteristics limitations each method also analyzed imaging.