Structure-Aware Noise Reduction Generative Adversarial Network for Optical Coherence Tomography Image

Optical coherence tomography (OCT) is a common imaging examination in ophthalmology, which can visualize cross-sectional retinal structures for diagnosis. However, image quality still suffers from speckle noise and other motion artifacts. An effective OCT denoising method is needed to ensure the image is interpreted correctly. However, lack of paired clean image restricts its development. Here, we propose an end-to-end structure-aware noise reduction generative adversarial network (SNR-GAN), trained with un-paired OCT images. The network is designed to translate images between noisy domain and clean domain. Besides adversarial and cycle consistence loss, structure-aware loss based on structural similarity index (SSIM) is added to the objective function, so as to achieve more structural constraints during image denoising. We evaluated our method on normal and pathological OCT datasets. Compared to the traditional methods, our proposed method achieved the best denoising performance and subtle structural preservation.