Medical Image Super-Resolution Based on the Generative Adversarial Network

In order to assist doctors to read medical pathological images with low resolution, this paper proposes a medical image super-resolution (SR) reconstruction method based on generative adversarial network (GAN). Considering that the pathological image has large non-organized regions, we design a medical pathological image preprocessing system to extract tissue area image patches. And, we improve discriminator with small batch relative discrimination to enhance the quality of reconstructed images by learning more prior information. We use Huber loss instead of the original MSE which can keep the network training stable. We find the feature similarity (FSIM) is suitable as an image quality evaluation way for medical image reconstruction research. And, the experimental results show the advantages of our method in the restoration of color and intercellular texture details.