Speckle Noise Reduction In Retinal Optical Coherence Tomography Based On Deep Learning

Optical coherence tomography(OCT)technology can quickly acquire cross-sectional images of ocular biological tissue in micron resolution.It has become an important tool for imaging of retina and provides help for clinicians to diagnose and treat ophthalmic diseases.Speckle noise caused by multiple forward and backward scattering of light waves is a major factor in the degradation of OCT image quality.The presence of speckle noise often obscures subtle but important morphological details and thus is detrimental to clinical diagnosis.It also affects the performance of automatic processing and analysis methods.Although the imaging resolution,speed and depth of OCT have been greatly improved over the last two decades,speckle noise,as an intrinsic problem of the imaging technique,has not been well solved.The main work of this thesis is to study deep learning methods to remove speckle noise in retinal OCT images.We propose two different convolutional neural network architectures.Both methods are trained on the retinal images obtained by two types of OCT scanners,Topcon DRI-1 and Topcon 2000.Clean images for training in both methods are obtained by registering and averaging the OCT images repeatedly acquired from the same normal eye.Then we test 9 normal and pathological retinal OCT volumes acquired from different scanners and take 36 slices from these volumes for quantitative analysis.The performance of the two methods is evaluated by comparing both the visual quality and quantitative indices including signal-to-noise ratio(SNR),contrast-to-noise ratio(CSR),equivalent number of looks(ENL)and edge preservation index(EPI)of the denoised images.The first method is based on the densely connected residual learning.We adopt the hypothesis of additive noise and the whole model is divided into three modules:shallow feature extraction,multi-level feature fusion and noise estimation.The residual dense block(RDB)is introduced to optimize the model.The experimental results show that the method perform well on speckle noise reduction,especially in aspect of improving the overall contrast of images and the smoothness of homogeneous regions.The second method is based on the improved conditional generative adversarial networks(cGAN).We regard removing speckle noise as the task of image-to-image translation,and add the proposed edge loss to the objective function of the model to improve the sensitivity of the model to image edge details.The experimental results show that the method has better overall performance on denoising OCT images,and it is more effective in removing speckle noise and retaining edge details at the same time.The proposed speckle noise reduction methods based on deep learning have definite clinical values.Both methods are of great significance to improve the visual observation of OCT images and the performance of subsequent automated processing and analysis methods.