Unpaired medical image enhancement based on generative adversarial networks
deng kezhi, Bo Tao, Jiaxin Hu · 2024
Medical imaging technology has significantly aided clinical decision-making, providing essential diagnostic and treatment information for physicians. Current medical image enhancement methods, based on pix2pix/CycleGAN, can improve image quality but often struggle to maintain uniform illumination and preserve texture details, introducing boundary pseudo-noise. Medical image enhancement faces unique challenges due to most medical image datasets being unpaired. Higher standards are required for illumination and texture in lesion areas. To address these issues, we propose UMIEGAN (Unpaired Medical Image Enhancement Generative Adversarial Networks) for medical image enhancement. In the generator, we introduce the residual shrinkage building unit with the channel-shared thresholds module (DRSN-CS) to suppress artifacts and image noise and extract medical image information through three paths using different convolution kernels. In the discriminator, a dual-scale discriminator is employed to evaluate generated images, enhancing the ability to judge medical images authenticity. We also introduce content-aware loss to improve lesion area details and illumination loss to optimize overall illumination distribution, resulting in smoother medical images. Extensive experiments were conducted on two datasets, the fundus retina dataset and the endoscopic image dataset, to verify the feasibility and accuracy of the proposed method. Results demonstrate that the proposed UMIEGAN outperforms traditional methods and other advanced deep learning methods and shows superior performance in downstream segmentation tasks on the fundus retina dataset.