Noise Removal for Optical Coherence Tomography Images: A Statistical Model-Driven Network

Current denoising methods are usually based on noise modeling or deep learning. In order to reduce the inherent noise in optical coherence tomography (OCT) images, a two-stage depth neural network is proposed, which combines the advantages of the two methods. The spatial noise distribution on an OCT image was represented by the gamma noise model. Res-Unet was used as the backbone network for both the noise estimation network and the denoising network. To better differentiate the foreground and background noise, a special context attention module was proposed to guide the network focusing onto the foreground retinal structure. A new edge-sensitive loss function was introduced to make the network preserve the sharp boundary between the retinal structure layers. The experimental results showed that the performance of our proposed method was superior to other methods. The noise was significantly reduced, with the retinal structure and the blood vessel details better preserved by our method. The proposed method can be greatly useful for the ophthalmic diagnosis by using OCT images.