SC-GAN: Structure Consistent GAN for Modality Transfer with FFT and Multi-Scale Perception

The quality of the cornea endothelial microscopy image is critical for clinical analysis. Although the noncontact specular microscope is more user-friendly than the contact confocal microscope, the imaging quality of the specular microscope is lower. The modality transfer is a promising solution for image quality enhancement. This paper proposes a Structure Consistent Generative Adversarial Network (SC-GAN) to transfer the imaging style from the specular microscope to the confocal microscope. Specifically, we use the Fourier frequency domain consistency to preserve cell structure and propose a multi-scale perception discriminator to improve model robustness under cell size variation. Experiment results prove the effectiveness of our method.