OCT image speckle sparse noise reduction based on dictionary algorithm
Wang Fan, Chen Minghui, Gao Naijun, Zhang Chenxi, Zheng Gang     
Institute of Biomedical Optics and Optometry, Shanghai Institute for Minimally Invasive Therapy, University of Shanghai for Science and Technology, Shanghai 200093, China

Overview: As a new non-invasive high-resolution scanning method, optical coherence tomography (OCT) has been widely used in clinical practice. Since the OCT imaging system uses an interference technique, the use of tissue scattering properties of light will inevitably introduce speckle noise. These speckle noises reduce the signal-to-noise ratio and contrast of the image, and also destroy the edge features of the image. As a result, it seriously affects people's accurate acquisition of image information. Therefore, the processing of OCT image speckle noise is very important before making a clinical diagnosis. The dictionary algorithm was originally proposed for Gaussian additive noise. This paper improves two original dictionary noise reduction algorithms for multiplicative speckle noise in OCT. The improved algorithm is divided into four steps. The first step is to establish and solve the speckle noise model of the OCT image. Firstly, the sparse domain model of small-sized image blocks is established and its noise reduction problem is solved. Then, the ideas in the Markov random field are used to generalize to large-size images. In the second step, logarithmically transforming the OCT image and performing noise estimation; In the third step, overlapping blocks the noisy image, the size of the image block is 8 pixels×8 pixels, the dictionary algorithm requires sparse coding and noise reduction for each block. The orthogonal matching pursuit algorithm (OMP) is used to perform sparse coding of two dictionary algorithms. In the fixed dictionary algorithm, the dictionary selects the discrete cosine transform (DCT) dictionary. In the adaptive dictionary algorithm, the initial dictionary selects the DCT dictionary and the dictionary training is performed by itself, and the dictionary update is completed by the K singular value decomposition learning algorithm; In the fourth step, the overlapping image blocks in the sparse coding stage are weighted averaged and returned to the spatial domain by exponential transformation. Selecting a randomly OCT slice and reduce noise for it, compared with the two original dictionary noise reduction algorithms and the traditional filtering algorithms, the improved two dictionary algorithms preserve most of the image information and edge detail information while suppressing speckle noise. Furthermore, three random OCT slice images are selected to simulate the improved two dictionary denoising algorithms. The improved adaptive dictionary algorithm has better noise reduction performance through subjective visual effects and four objective evaluation indicators. The two improved dictionary noise reduction algorithms proposed in this paper can be flexibly applied to various OCT noisy images and serve for subsequent image processing.

Supported by the National Science Foundation for Young Scientists of China (61308115), Shanghai Natural Science Foundation (13ZR1457900), and Industrial Technology and Medical Research Funds of Shanghai (15DZ1940400)