Cluster-based filtering framework for removing speckles with structural protection in OCT images

Generally speaking, it is the essential core of image filtering to keep the texture features better while denoising the image. To some extent, optical coherence tomography retina images have speckle noise, which masks the texture features of the image, and thus causes misjudgment to the doctor’s diagnosis. In this paper, we first propose a cluster-based filtering framework for removing speckles with structural protection in OCT images. The overall process can be divided into preprocessing, structure extraction and structure denoising. First, in the preprocessing stage, we propose to use the shearlet (SHT) method for preliminary filtering and combine block search and matching to achieve structure protection. Then in the structure extraction stage, we propose to use the relative total variation algorithm to achieve structure extraction, combined with fuzzy C-means Clustering filters out the background noise to obtain the structure mask of the image. Finally, in the structure denoising stage, we propose a new variational Block matching 3D (BM3D)-L2 method, and the structure of the image and the noise are described in BM3D space and L2 space, respectively. By assigning appropriate values to the parameters, image noise can be better eliminated, and the structural texture of the image can be protected. We test the proposed method on seven large noisy OCT images, which include five human retinal OCT images and two mouse optic nerve OCT images. In addition, we also compare it with SHT, BM3D, TV-SHT and TV-BM3D methods, which were proved to be effective in denoising. The performances of these methods are quantitatively evaluated in terms of the signal-to-noise ratio (SNR), contrast-to-noise ratio (CNR) and the averaged equivalent number of looks (ENL) at the aspects of speckle reduction and structure texture protection. Vast experiments show that our proposed method can effectively reduce speckle noise in OCT images, protect important structural information and improve image quality. Here, we believe that our method will improve image segmentation, medical diagnosis, and can use this as training samples to improve the accuracy of machine learning.