|M.Sc Student||Falik Adi|
|Subject||On Compression and Speckle Suppression for Ultrasound|
|Department||Department of Electrical Engineering||Supervisors||Professor Moshe Porat|
|Dr. Zvi Friedman|
|Full Thesis text|
Reducing speckle noise in ultrasound (US) imaging is an essential pre-processing step
for most computer aided diagnostics (CAD) algorithms, as well as for efficient storage and transmission of ultrasound data. However, state-of-the-art methods still suffer from high computational complexity and therefore are not suitable for real-time applications.
In this work, we use transform domain filtering (TDF), which is one of the simplest
denoising techniques yet very effective for natural images. Nevertheless, finding an
optimal threshold for the filtering operation is a challenging and crucial task, especially in cases of non-Gaussian noise such as speckle noise. Accordingly, we minimize a constrained sparsity-prior for the transform coefficients, thus converting the algorithm to be controlled by a single threshold, which is robust and easier to adjust. Based on statistical and empirical analysis of the speckle noise, we develop a method for estimating the optimal threshold by means of PSNR with respect to a de-speckled image that is obtained by established de-speckling algorithms, such as non-local means (NLM). Based on experiments performed over in-vivo ultrasound images, we show that the proposed method significantly reduces the computational complexity, while achieving a performance level that is visually indistinguishable with respect to the predetermined de-speckling method. The proposed method can efficiently denoise and compress ultrasound imaging for real-time applications.
In addition, some ultrasound imaging modalities are severely corrupted by speckle noise that has a spiky nature. We show that, in such cases, the performance of the established de-speckling algorithms is significantly worse. In order to overcome this difficulty, we combine the proposed method with a simple pre-processing stage, denoted as outlier shrinkage, so that the spiky components are removed without affecting the anatomical information and the noise in the US image becomes similar in its behavior to a Gaussian noise. Based on experiments carried over in-silico and in-vivo ultrasound images, we demonstrate that this pre-processing step significantly improves the performance of our proposed method, enabling a performance level which is competitive with established de-speckling algorithms.