As a complicated ubiquitous phenomenon encountered in ultrasound imaging, speckle can be treated as either annoying noise that needs to be reduced or the source from which diagnostic information can be extracted to reveal the underlying properties of tissue. In this study, the application of Independent Component Analysis (ICA), a relatively new statistical signal processing tool appeared in recent years, to both the speckle texture analysis and despeckling problems of B-mode ultrasound images was investigated. It is believed that higher order statistics may provide extra information about the speckle texture beyond the information provided by first and second order statistics only. However, the higher order statistics of speckle texture is still not clearly understood and very difficult to model analytically. Any direct dealing with high order statistics is computationally forbidding. On the one hand, many conventional ultrasound speckle texture analysis algorithms use only first or second order statistics. On the other hand, many multichannel filtering approaches use pre-defined analytical filters which are not adaptive to the data. In this study, an ICA-based multichannel filtering texture analysis algorithm, which considers both higher order statistics and data adaptation, was proposed and tested on the numerically simulated homogeneous speckle textures. The ICA filters were learned directly from the training images. Histogram regularization was conducted to make the speckle images quasi-stationary in the wide sense so as to be adaptive to an ICA algorithm. Both Principal Component Analysis (PCA) and a greedy algorithm were used to reduce the dimension of feature space. Finally, Support Vector Machines (SVM) with Radial Basis Function (RBF) kernel were chosen as the classifier for achieving best classification accuracy. Several representative conventional methods, including both low and high order statistics based methods, and both filtering and non-filtering methods, have been chosen for comparison study. The numerical experiments have shown that the proposed ICA-based algorithm in many cases outperforms other algorithms for comparison. Two-component texture segmentation experiments were conducted and the proposed algorithm showed strong capability of segmenting two visually very similar yet different texture regions with rather fuzzy boundaries and almost the same mean and variance. Through simulating speckle with first order statistics approaching gradually to the Rayleigh model from different non-Rayleigh models, the experiments to some extent reveal how the behavior of higher order statistics changes with the underlying property of tissues. It has been demonstrated that when the speckle approaches the Rayleigh model, both the second and higher order statistics lose the texture differentiation capability. However, when the speckles tend to some non-Rayleigh models, methods based on higher order statistics show strong advantage over those solely based on first or second order statistics. The proposed algorithm may potentially find clinical application in the early detection of soft tissue disease, and also be helpful for better understanding ultrasound speckle phenomenon in the perspective of higher order statistics. For the despeckling problem, an algorithm was proposed which adapted the ICA Sparse Code Shrinkage (ICA-SCS) method for the ultrasound B-mode image despeckling problem by applying an appropriate preprocessing step proposed by other researchers. The preprocessing step makes the speckle noise much closer to the real white Gaussian noise (WGN) hence more amenable to a denoising algorithm such as ICS-SCS that has been strictly designed for additive WGN. A discussion is given on how to obtain the noise-free training image samples in various ways. The experimental results have shown that the proposed method outperforms several classical methods chosen for comparison, including first or second order statistics based methods (such as Wiener filter) and multichannel filtering methods (such as wavelet shrinkage), in the capability of both speckle reduction and edge preservation.
Library of Congress Subject Headings
Ultrasonic imaging--Data processing; Multivariate analysis; Imaging systems--Image quality; Diagnostic ultrasonic imaging; Imaging systems in medicine; Image processing
Department, Program, or Center
Chester F. Carlson Center for Imaging Science (COS)
Lai, Di, "Independent component analysis (ICA) applied to ultrasound image processing and tissue characterization" (2009). Thesis. Rochester Institute of Technology. Accessed from
RIT – Main Campus