A Strictly Bounded Deep Network for Unpaired Cyclic Translation of Medical Images

Abstract

Medical image translation is an ill-posed problem. Unlike existing networks, we consider unpaired medical images as input, and provide a strictly bound generative network that yields a stable cyclic (bidirectional) translation. It consists of two cyclically connected conditional GANs where both generators (32 layers each) are conditioned with concatenation of alternate unpaired patches from input and target images of the same organ. The key idea is to exploit cross-neighborhood contextual feature information to bound translation space and boost generalization. Further, the generators are equipped with adaptive dictionaries which are learned from the cross-contextual patches to reduce possible degradation. Discriminators are 15-layer deep networks which employ minimax function to validate the translated imagery. A combined loss function is formulated with adversarial, non-adversarial, forward-backward cyclic, and identity losses that further minimize variance of the proposed learning machine. Qualitative, quantitative, and ablation analysis show superior results on real CT and MRI datasets.