LEGAN: A Light and Effective Generative Adversarial Network for medical image synthesis

Jing Gao, Wenhan Zhao, Peng Li, Wei Huang, Zhikui Chen

Abstract: Highlights•The medical image synthesis is modeled as a coarse-to-fine paradigm, imitating painting process of humans.•A lightweight convolutional layer is introduced to reduce the redundancy of computation.•A multi-stage mutual information loss is designed to explore the dependencies of data distributions for generating high-fidelity medical images.•LE-GAN generates high-fidelity medical images in a low cost manner, evaluated in retinal fundus image and PD-weighted MR image synthesis.