Improved Image Enhancement and Acquisition for Ultrasound Images

Abstract

Compared to others imaging modalities such as computed tomography (CT), magnetic resonance imaging (MRI), positron emission tomography (PET), ultrasound (US) imaging is a quite popular and extensively used diagnostics tool due of its affordability, comparative safety, portability, adaptability, non-invasiveness and absence of ionizing radiation. However, US scans have considerably lower quality images compared to other imaging modalities, making its effectiveness in medical diagnosis quite constrained. But there is a strong demand for providing the viewer with high quality images, not only for better visualization but also for automated feature extraction and recognition tasks. The image quality of US imaging is severely affected by the presence of speckle noise and blur. Speckle noise, apart from Gaussian noise, is a signal dependent noise often called multiplicative noise. Therefore, quality enhancement by reducing speckle noise and blur is highly required in US imaging. Moreover, long acquisition times and very large data processing bandwidth are other limiting factors especially for 3D US imaging which is finding its way in many consultation rooms. Compressed sensing (CS) is one of the best candidates to overcome the limiting factors in 3D US imaging. It allows-under certain assumptions-to recover a signal sampled below the Nyquist-Shannon sampling rate. CS is a novel theory allowing to reduce the amount of data collected during the acquisition, which in turn reduces long acquisition times. However, CS is prone to introduce noise-like artefacts due to random under-sampling. This research dissertation aimed at developing a collection of US image processing algorithms with the purpose of quality enhancement and faster acquisition. The dissertation first describes the complex wavelet-based speckle reduction algorithm where speckle is iteratively reduced by the multi-scale nonlinear diffusion in the dual tree complex wavelet transform (DTCWT) domain. The proposed method exploits some useful features of the DTCWT and nonlinear diffusion. Rayleigh mixture model is used to model the overall wavelet modulus of the noisy input image, and the parameters of the mixture model are calculated using a genetic algorithm. Secondly, artificial neural networks (ANNs) and regularisation based restoration algorithms are proposed. In ANNs based algorithm, an intelligent estimator based on ANNs is used to estimate the noise and blur variances, which in turn, are used to obtain an image without discernible distortions. In the proposed algorithm, the inverse of the Rayleigh function is solved numerically using a look-up table for speckle reduction, and the Lucy-Richardson algorithm is used for de-blurring. In regularisation based restoration algorithms, the estimation of the prior model has the greatest impact on restoration results. In the proposed algorithm, a Laplacian scale mixture (LSM) model combined with a total variation (TV) constraint is used as a regularisation prior in the complex wavelet domain. Thirdly, a super resolution (SR) algorithm is proposed to obtain a high resolution image from a set of low resolution (LR) images acquired from the same scene. An SR algorithm which can improve the resolution and reduce the blur and speckle noise without perceivable distortions is highly necessary. In the proposed SR algorithm, a DTCWT based Rayleigh scale mixture (RSM) model regularisation prior is used, which could take the benefits of the heavy-tailed nature of the wavelet coefficients of US images. Finally, a compressed sensing (CS) method is proposed and also adapted to 3D US imaging. In the proposed method, an RSM model constraint combined with a TV constraint in the wavelet domain to create a new regularisation prior is used to reconstruct the image from under sampled data without degrading the image quality. The proposed algorithms mentioned above are validated using different set of synthetic, physical phantom, and clinical images. In every case the proposed algorithms have been observed to give better results, both quantitatively and qualitatively compared to state-of-the-art methods.