RetiGAN: A Hybrid Image Enhancement Method for Medical Images
Zhen Yu Huang, Shiyuan Wang, Huijie Hu, Yiyao Xu · 2024
Image enhancement can improve image quality and visual effects well, and there are many outstanding applications in natural images. Medical image enhancement is relatively difficult due to its particularity, such as lack of unified standards, scarce annotation, small data, high accuracy requirements, complex patterns, insufficient image contrast, fuzzy details, and large noise interference. Due to the large gap between natural and medical images, there is a lack of effective universal methods for both types of data. This paper proposes a hybrid model based on the Retinex theory and adversarial generative network to implement the RetiGAN model. Firstly, the image is processed by Gaussian filtering, inverse number transformation and image stretching, and then the data is amplified by the generative adversarial network (GAN) to realize data enhancement. In terms of experiments, nearly a thousand X-ray data based on two public and one private medical image datasets were compared with baseline, the most common image enhancement method, and adequate comparative experiments were conducted with mainstream methods such as LL-Net, Zero-DCE, PELE, etc. Experiments show that the method not only improves the image quality indexes such as PSNR, LPIPS and SSIM but also has remarkable visualization effects and enhancement effects in texture details, which shows that our method has great potential and application space in practical clinical diagnosis.