Hierarchical Probabilistic Ultrasound Image Inpainting via Variational Inference

Alex Ling Yu Hung, Zhiqing Sun, Wanwen Chen, John M. Galeotti

Abstract: Image inpainting is a well established problem in computer vision which aims to fill in missing regions in images. In medical applications, it can be combined with other tools to remove artifacts or fill in occluded regions, allowing better understanding of the images from doctors or downstream algorithms. However, current methods that solve the problem usually pay no attention to the underlying pixel-intensity distributions in the missing input regions, and most approaches deterministically provide only one result per input region instead of several plausible results. We estimate the intensity distributions within each masked region using a novel Variational Autoencoder (VAE) based hierarchical probabilistic network. Our approach then generates a diverse set of inpainted images, all of which appear visually appropriate.