|dc.description.abstract||Retinal imaging provides an opportunity to detect pathological and natural age-related
physiological changes in the interior of the eye. Diagnosis of retinal abnormality requires an image that is sharp, clear and free of noise and artifacts. However, to prevent tissue damage, retinal imaging instruments use low illumination radiation, hence, the signal-to-noise ratio (SNR) is reduced which means the total noise power is increased. Furthermore, noise is inherent in some imaging techniques. For example, in Optical Coherence Tomography (OCT) speckle noise is produced due to the coherence between the unwanted backscattered light. Improving OCT image quality by reducing speckle noise increases the accuracy of analyses and hence the diagnostic sensitivity. However, the challenge is to preserve image features while reducing speckle noise. There is a clear trade-off between image feature preservation and speckle noise reduction in OCT.
Averaging multiple OCT images taken from a unique position provides a high SNR
image, but it drastically increases the scanning time. In this thesis, we develop a multi-frame image denoising method for Spectral Domain OCT (SD-OCT) images extracted from a very close locations of a SD-OCT volume. The proposed denoising method was tested using two dictionaries: nonlinear (NL) and KSVD-based adaptive dictionary. The NL dictionary was constructed by adding phases, polynomial, exponential and boxcar functions to the conventional Discrete Cosine Transform (DCT) dictionary. The proposed denoising method denoises nearby frames of SD-OCT volume using a sparse representation method and combines them by selecting median intensity pixels from the denoised nearby frames. The result showed that both dictionaries reduced the speckle noise from the OCT images; however, the adaptive dictionary showed slightly better results at the cost of a higher computational complexity. The NL dictionary was also used for fundus and OCT image reconstruction. The performance of the NL dictionary was always better than that of other analytical-based dictionaries, such as DCT and Haar.
The adaptive dictionary involves a lengthy dictionary learning process, and therefore
cannot be used in real situations. We dealt this problem by utilizing a low-rank approximation. In this approach SD-OCT frames were divided into a group of noisy matrices that consist of non-local similar patches. A noise-free patch matrix was obtained from a noisy patch matrix utilizing a low-rank approximation. The noise-free patches from nearby frames were averaged to enhance the denoising. The denoised image obtained from the proposed approach was better than those obtained by several state-of-the-art methods. The proposed approach was extended to jointly denoise and interpolate SD-OCT image. The results show that joint denoising and interpolation method outperforms several existing state-of-the-art denoising methods plus bicubic interpolation.||en