Speckle Noise Removal In Medical Ultrasound Image

Speckle is a multiplicative noise that is an inherent phenomenon in medical ultrasound images. It reduces images'contrast, blurs details and produces local pseudo features, which largely limits clinical utilities of ultrasound images. A number of methods were therefore proposed for the speckle noise reduction. All these methods try to smooth the intensity variation due to the speckle noise and meanwhile preserve the one caused by the image features. Each group of method used different techniques, which have their own advantages as well as limitations.In order to optimize ultrasound image clinical utilities and improve the existed de-speckled methods performance, a set of novel frameworks and algorithms are proposed in this paper. Different techniques are employed for ultrasound I/Q image and B-mode image, respectively.For ultrasound I/Q image, two novel de-speckle frameworks are presented. The first one is based on Wiener filter and noise Gaussianization. The system point spread function (PSF) is blindly estimated from I/Q image, which is further imported into Wiener filter to eliminate the spatial correlation of speckle. The homomorphic transform is then performed on the de-convoluted I/Q data set in order to convert the multiplicative noise to the additive one. The additive noise is finally Gaussianized for the purpose of making full usage of existed classic de-noise algorithm.The second one takes the advantage of the expectation maximum (EM) algorithm. After the PSF is extract from I/Q image, both of them are involved in the EM algorithm:the E-step is implemented by a matrix-form Wiener filter, which tends to reduce the image's correlation; the M-step utilizes the anisotropic diffusion to generate noise-free tissue echogenicity, which further servers as a feedback into E-step.After processed with the above frameworks, the speckle noise is converted from spatical correlated into random distributed. Thus the existed de-speckled algorithm can suppress it more efficiently. The comparison experiment results indicate that, compared to process the I/Q image directly, the proposed frameworks optimize the performance of the existed de-speckle algorithms.For ultrasound B-mode image, two novel de-speckle algorithms are proposed. The first one is a spatially adaptive version of the maximum-likelihood (ML) technique. A binary edge mask is firstly estimated to indicate the possible edges in the speckled image. An ML estimation approach is then utilized, whose shape parameter and window size are adaptively controlled by the edge mask. Finally, a noise-free tissue echogenicity map is obtained.The second method extends the dual-tree complex wavelet transform (DTCWT) and bivariate shrinkage function from natural image to ultrasound one. One the one hand, the DTCWT preserves the image directional and detailed information. On the other hand, the bivariate shrinkage function differential noise wavelet coefficients from useful image ones more effectively.The quantitative assessment from both simulation and in vivo experiment indicate that the proposed two algorithms strike an optimal balance between noise suppression and edge preservation.