Peter C. Tay

Ultrasound Despeckling for Contrast Enhancement

Images produced by ultrasound systems are adversely hampered by a stochastic process known as speckle. A despeckling method based upon removing outlier is proposed. The method is developed to contrast enhance B-mode ultrasound images. The contrast enhancement is with respect to decreasing pixel variations in homogeneous regions while maintaining or improving differences in mean values of distinct regions. A comparison of the proposed despeckling filter is compared with the other well known despeckling filters. The evaluations of despeckling performance are based upon improvements to contrast enhancement, structural similarity, and segmentation results on a Field II simulated image and actual B-mode cardiac ultrasound images captured in vivo.

Ultrasound Despeckling Using an Adaptive Window Stochastic Approach

A novel stochastically driven filtering method to despeckle B mode ultrasound images is presented. This method is motivated by viewing the pixel values as a stochastic process and removing outliers, where outliers are defined by local extrema. These outliers are removed by local averaging. This produces another image with new outliers (local extrema) and the process is iterated. With each iteration homogeneous regions become smoother while edges that defined these regions are preserved. By allowing a dynamically varying window to determine the local mean, we achieve equivalent results with fewer iterations.

A stochastic approach to ultrasound despeckling

A novel stochastically driven filtering method to despeckle B mode ultrasound images is presented. This method is motivated by viewing the pixel values as a stochastic process and removing outliers, where outliers are defined by local extrema. These outliers are removed by local averaging. This produces another image with new outliers (local extrema) and the process is iteratively repeated. With each iteration homogeneous regions become smoother while edges that defined these regions remain preserved. To evaluate the performance of our proposed method in satisfying these two opposing goals we develop a modified Fisher discriminant contrast metric. Larger values of this metric indicate better performance in reducing each intraregion or intraclass variance and increasing the difference of interregion or interclass means

A wavelet thresholding method to reduce ultrasound artifacts

Artifacts due to enhancement, reverberation, and multi-path reflection are commonly encountered in medical ultrasound imaging. These artifacts can adversely affect an automated image quantification algorithm or interfere with a physicians assessment of a radiological image. This paper proposes a soft wavelet thresholding method to replace regions adversely affected by these artifacts with the texture due to the underlying tissue(s), which were originally obscured. Our proposed method soft thresholds the wavelet coefficients of affected regions to estimate the reflectivity values caused by these artifacts. By subtracting the estimated reflectivity values of the artifacts from the original reflectivity values, estimates of artifact reduced reflectivity values are attained. The improvements of our proposed method are substantiated by an evaluation of Field II simulated, in vivo mouse and human heart B mode images.

AM-FM Image Analysis Using the Hilbert Huang Transform

This paper explores the incorporation of the two dimensional (2D) empirical mode decomposition, which is used in the Hilbert-Huang Transform, into a meaningful AM-FM image model framework. A virtue of the empirical mode decomposition is that it decomposes a non-stationary signal into stationary intrinsic mode functions and a residue signal. The empirical mode decomposition attempts to produce intrinsic mode functions that are stationary and a residue function that is dominated by piecewise monotonic functions. When considering image pixel values as produced from a non-stationary process, the 2D empirical mode decomposition shows promise as a precursor step in determining AM-FM components where stationarity is needed.

A novel shape feature to classify microcalcifications

The cited references claim that microcalcifications from many benign regions are all round or oval. The detection of at least one roughly shaped microcalcification in a suspicious region could be an early sign of potentially developing malignant cancer. This paper proposes a shape analysis method to aid radiologists in classifying regions of interest that are difficult to diagnosis. A region growing and a gradient vector flow methods are used to obtain an ordered set of contour points of each microcalcification. A three level wavelet transform frequency analysis provides a band pass approximation of the normalized distance signature. A novel metric derived from the normalized distance signature is proposed to quantify the roughness of a microcalcification. An experiment using a large dataset is used to evaluate the robustness of the proposed roughness metric against several published shape features.

Unsupervised texture segmentation using dominant image modulations

We present an unsupervised modulation domain technique for segmenting textured images. A dominant component AM-FM analysis is performed on the image, and estimates of the locally dominant amplitude and frequency modulations are extracted at each pixel. Modulation domain density clustering is then applied to estimate the maximum number of textured regions that might be present in the image. The feature space is augmented with horizontal and vertical spatial information prior to the application of k-means clustering to arrive at an initial image segmentation. Connected components labeling with minor region removal and morphological smoothing are then applied to yield the final segmentation. We demonstrate the technique on several synthetic and natural images.

A novel translation and modulation invariant discrete-discrete uncertainty measure

The quantification of signal localization simultaneously in time and in frequency is fundamental to a variety of signal processing applications where time-frequency analysis is to be performed on nonstationary signals. In this paper, we develop novel joint localization measures defined on equivalence classes of finitely supported discrete-time signals. These measures bear strong analogies to the well-known continuous-time Heisenberg-Weyl inequality. In particular, they are invariant to signal translations and modulations and admit an intuitive interpretation in terms of the temporal and spectral variance of the signal energy. The new measures are used to design optimal wavelet quadrature mirror filter banks that exhibit improved localization relative to the Haar and Daubechies analysis filters.

Analysis of Stress in speech using adaptive Empirical Mode Decomposition

Stress in human speech can be detected by various methods know as Voice Stress Analysis (VSA). The detection is accomplished by measuring the frequency shift of a microtremor normally residing in the frequency range of 8 to 12 Hz when not stressed. Conventional detection methods include Fast Fourier Transform (FFT) or McQuiston-Ford algorithm. This paper presents a new method called Adaptive Empirical Mode Decomposition (AEMD) applied to voice stress detection. Because AEMD in essence is a time-frequency analysis method, it is possible to use this method for real-time voice stress detection.

Frequency implementation of discrete wavelet transforms

The paper implements the discrete wavelet transform in the discrete Fourier domain. The need for such an approach arose out of our desire to find a convenient means of realizing a new class of non-separable orientation selective 2D wavelet filter banks that are designed directly in the DFT domain. The filter bank design process begins with a conventional separable 2D perfect reconstruction parallel filter bank that is not orientation selective. In the DFT domain, each non-low pass channel is decomposed into the sum of two orientation selective frequency responses that are each supported on only two quadrants of the 2D frequency plane. The resulting filter bank possesses the good joint localization properties of orthogonal wavelet transforms and offers both perfect reconstruction and orientation selectivity. However, the orientation selective channels are non-separable - they cannot be implemented as iterated 1D convolutions according to the usual separable 2D wavelet transform paradigm. To overcome this difficulty, we develop straightforward techniques for implementing the DWT directly in the DFT domain.