|
Optical coherence tomography (OCT) enables non-invasive imaging of biological tissue and has become one of the most effective tools for monitoring the retinal structures and detecting retinal diseases. However, the existence of speckle noise severely degrades the OCT image quality and makes it difficult to identify the retinal disorders accurately. In this work, a deep generative model, named as despeckling generative adversarial network (DSGAN), is proposed for retinal OCT image despeckling. The proposed DSGAN is composed of two components, i.e., a despeckling generator and a discriminator. The despeckling generator employs the residual-in-residual dense block-based U-shape network to learn how to map the noisy image to the clean image. The discriminator learns to accurately discriminate whether the real clean images are relatively more realistic than the image generated by the generator. To improve the structure preservation ability during speckle noise reduction, the structural similarity index measure (SSIM) loss is introduced into the objective function of DSGAN to achieve more structural constraints. The proposed DSGAN was evaluated and analyzed on two public OCT datasets. The qualitative and quantitative comparison results show that the proposed DSGAN can achieve higher image quality, and is more effective in both speckle noise reduction and structural information preservation than previous despeckling methods.
|
