DENOISING AND SEGMENTATION OF ULTRASOUND MEDICAL IMAGES

Abstract

In the field of medical imaging technologies, ultrasound imaging (US) has an important role in the diagnosis of several diseases because of its safe application for patients, low cost and portability. In usual, the diagnostic procedures based on the perception of medical images are performed in a subconscious way which is based on the conclusion drawn upon how the clinicians understand and interpret them. It is only since the last couple of years that computers have been used to process the images digitally in the equipment to some extent and thus help the clinician in their diagnostic decision. However, it is a difficult task in the case of US images because of the presence of speckles and other artifacts. The image segmentation is also a key step in several computer aided diagnosis systems used to identify a particular region of interest such as the tumor, lesion and other abnormalities, to measure the growth of tumors and to help in treatment planning. Speckle considered as multiplicative noise is a prime factor that degrades the contrast resolution and masks the meaningful texture information present in the US images. This makes accurate identification of object boundaries and contours of anatomical structures a challenging task. Hence, the image denoising can be considered as a fundamental task to improve the quality of the US images by suppressing the speckle noise without affecting the edge information. Thus, in the above perspective, the image denoising algorithms should fulfil the following three principal criteria: 1. The algorithm must be capable of suppressing the maximum amount of noise from a particular region. 2. The true tissue information, including the edges and other fine details should also be preserved and if possible, may be enhanced. 3. The denoising algorithm must be computationally efficient, stable and robust. With the above background, the main objective of the present research work has been to design and develop the effective algorithms to achieve the improved performance in image denoising and segmentation. To obtain these objectives, it has been necessary to analyze and identify better approaches among the existing remarkable denoising and segmentation approaches and also it necessitated to improve the performance of the best identified approach, either by modifying the earlier algorithm or by suggesting a new algorithm. Accordingly, the entire research work has been planned and carried out in the following three different steps. 1. Comparison of several existing image denoising methods has been carried out under two major categories, i.e. spatial domain category and transform domain category, and an optimum approach suitable for the US medical images has been identified. 2. Based on the post-analysis of the approaches from two major categories mentioned above, new suitable denoising approaches have been designed and implemented to ii improve the quality of the US images by suppressing the noise and preserving the edges. 3. A comparative evaluation of several existing segmentation approaches used for the US medical images have been carried out and new segmentation algorithms for the US images have been designed and implemented. In order to achieve the first objective of the initial phase of work, the performance of various existing denoising methods namely, adaptive weighted median filter (AWMF), wiener filter (WF), maximum homogeneity over a pixel neighborhood filter (MHOPNF), anisotropic diffusion filter (ADF), speckle reducing anisotropic diffusion (SRAD), nonlinear complex diffusion filter (NCDF), total variation filter (TVF), nonlocal means filter (NLMF) in spatial domain, while wavelet, ridgelet, curvelet and shearlet based methods in the transform domain category, have been evaluated on the test images degraded by different types of noise such as speckle and Gaussian noise and on real US images. Their performance is not only analyzed and evaluated in terms of visual perception, but also in terms of different performance measures such as the peak signal to noise ratio (PSNR), signal to noise ratio (SNR), structural similarity index metric (SSIM), figure of merit (FOM) and edge keeping index (EKI). Besides the above parameters, mean to variance ratio (MVR) has also been used for the quantitative analysis of the US images because of the nonavailability of the reference US images. From the experimental results, it is observed for all noise levels that the TVF, NLMF, curvelet and shearlet based approaches are able to suppress good amount of noise with better edge preservation in terms of quantitative measures, but in curvelet based method, some visual distortion like oscillations occur in the denoised images. For some images, ridgelet based approach also provides a competitive performance than others, but further it leads to some visual distortion. Moreover, diffusion based and the TVF approaches suffer from the loss of edge information. It is also observed that the better perceptual quality is obtained by the TVF, NLMF and shearlet based approaches. The diffusion, ridgelet and curvelet based approaches are also suitable to provide better denoising performance, but at the cost of blurring the edges and introducing some visual distortion. Under the second objective, six different algorithms are proposed in the present work. Based on the results obtained from the above comparative analysis, the first denoising approach based on M-band ridgelet transform is proposed in the present work. It utilizes the features of M-band wavelet transform in place of the 2-band wavelet transform (WT) used in the implementation of the ridgelet transform. The proposed approach utilizes the variation of the frequency resolution feature of the À trous algorithm by which a noisy image has been decomposed into different scales. NeighShrink (NS) thresholding is also utilized in the proposed approach to provide the approximated modified image coefficients that also iii improve the noise reduction efficiency. Based on the experimental results, it is observed that the proposed M-band ridgelet thresholding (MBRT) approach is able to produce better results by suppressing a sufficient amount of the speckle noise with more edges being preserved. It is observed on the basis of the findings obtained from the first objective, that the curvelet based approach also produces better denoising results. Curvelet transform (CVT) uses a parabolic scaling law to resolve the two dimensional singularities along �������� curves. It also overcomes the limitation of the wavelet and ridgelet, which are less efficient to represent the sharp transition like line and curve singularities available in the images. To represent the edges more efficiently, ripplet transform (RT) has been evolved by incorporating two new additional parameters. It also provides a new tight frame with a sparse representation for the source images with discontinuities along the �������� curves, where ���� ���� 2 refers to parabolic scaling same as the curvelets and for ���� ���� 3, ripplet has the cubic scaling and so forth. Based on the literature review and comparative evaluation performed earlier, the WT thresholding approach has improved its performance by incorporating some spatial domain techniques. Considering their merits, two different nonlinear filtering approaches in ripplet domain have been proposed here using the NS and BlockShrink (BS) thresholding approach that are named as the RTNLF-1 and the RTNLF-2, respectively. Nonlinear bilateral filtering (NLBF) is applied to the low frequency ripplet coefficients. The performance of these proposed denoising methods depend on the decomposition levels, different parameters of the RT and NLBF approach. The optimal values of these parameters have been decided by conducting the several experiments on the available test image datasets for the different levels of speckle noise with several combinations of these parameters. The results of the proposed RTNLF-1 and RTNLF-2 methods are compared with the bilateral, wavelet based NeighShrink (WT-NS), wavelet based NeighShrink using the NLBF (WT-NLBF-NS), wavelet based BlockShrink (WT-BS), linear homogeneous mask area filter (LHMAF), ADF, Fourth order PDE filter (FOPDEF), SRAD, NCDF, improved nonlinear complex diffusion (INCDF), wavelet based approach using generalized Gaussian distribution (WT-GGD), squeez box filter (SBF) and TVF approach. It is observed from their comparative results that proposed RTNLF-1 and RTNLF-2 methods provide better quality of images by suppressing more speckle noise as compared to others. Moreover, the RTNLF-2 approach performs better than the RTNLF-1. The higher SNR and PSNR values with larger EKI and SSIM values obtained by the proposed approaches indicate that the noise suppression is neither at the cost of blurring the edges nor at the loss of edge information. From the analysis of experimental results obtained earlier, it is observed that the TVF approach is quite good and has the ability to suppress the noise, but it performs noise reduction with the loss of edge information which means that some edge information is lost iv within the noise residual. Thus, a remnant approach for adaptive fusion based noise filtering (RBAF) is proposed using the TVF and shearlet thresholding using cycle spinning (CSST) approach. The proposed RBAF approach fuses three different images processed by the (a) TVF approach, (b) CSST approach and (c) extracted edges structured information (ESI) from the remnant of TVF approach and processed by the CSST approach. The proposed RBAF approach fuses these images based on the 3×3 block variance map evaluated for all three above processed images. The RBAF approach improves both the perceptual quality and the detectability of real US images and several test images corrupted with the speckle and Gaussian noise of different levels (characterized by their standard deviation and variance). To assess the performance of the proposed RBAF approach, the results of other methods such as the TVF, TI-WT, curvelet thresholding using cycle spinning (CSCVT) approach, CSST, CVT with the TVF approach (CVT-TVF) are considered. It is observed from the analysis of all the experimental results that the higher values of EKI and FOM with improved values of the SNR, PSNR and SSIM are obtained for the proposed RBAF approach. Further, it indicates that the improved noise suppression provided by the proposed RBAF approach does not produce the blurred edges. Besides this improved performance, the proposed RBAF approach also helps to suppress the staircase/blocking effects produced by the TVF method and the fuzzy edges introduced by the CVT and ST based methods. Based on the outcomes of the different experiments performed earlier and the literature, it is observed that anisotropic diffusion is widely used for denoising of the US image, but it suffers from the loss of edges information available in the images that are also very important for visual perception. To represent more edges, nonsubsampled shearlet transform (NSST) has been presented by providing both the multiscale and direction analysis of an image. Further, two different noise filtering approaches using the modified nonlinear adaptive anisotropic diffusion (NADF) equations in the NSST domain and thresholding approach (similar to the CVT and ST thresholding) have also been proposed. In the modified diffusion process, an adaptive gray variance is also incorporated with the gradient information of eight connected neighboring pixels to preserve the edges, effectively in the first proposed approach and named as the NSST-NADF. Motivated by the better noise reduction results of the NLMF presented in the first comparative analysis, the nonlocal pixel information is also incorporated to evaluate the gradient of eight connected neighboring pixels with an adaptive gray variance in the second proposed approach and is named as the NSST-NLNADF. Their denoising performance is also compared with all aforementioned existing methods, including the speckle reducing bilateral filter. The proposed methods are also adapted to both the speckle and Gaussian noise. Based on the experimental results, it is observed that the proposed NSST-NLNADF approach ensures an improvement in noise reduction and preservation of more edges by providing higher SNR, PSNR, FOM, SSIM and v EKI values than the NSST-NADF approach and others and thus providing the processed images with better visual quality. In order to achieve the next objective of the present work, two different segmentation methods are proposed to delineate the region of interest in the US medical images using clustering and level set methods. Based on the literature, the performance of the traditional active contour segmentation methods is subjected to appropriate and accurate contour initialization and optimal configuration of the contour propagation controlling parameters, which also require extensive manual intervention. Therefore, two different segmentation approaches, namely hybrid edge-based active contour model (EBACM) with the kernel fuzzy c-mean clustering (KFCM) and region-based active contour model (RBACM) with the Gaussian kernel fuzzy clustering (GKFCM) have been proposed here to segment the US medical images. In the first proposed segmentation approach, the features of both the edge-based active contour model using the distance regularized level set and the KFCM clustering are merged to detect the abnormalities from the US images. In the proposed approach, the fuzzy membership function from the variants of KFCM with spatial constraints, i.e. KFCM_S1 and KFCM_S2 clustering, is employed to initialize the curve that evolves to extract the desired object of interest. In addition to contour initialization, the results of the fuzzy clustering are also used for further evaluation of the contour propagation controlling parameters. Thus, two different segmentation approaches, namely the EBACM-KFCM_S1 and EBACM-KFCM_S2 are proposed, which start with the KFCM_S1 and KFCM_S2 methods, respectively, to initialize the curve and evaluate the curve evolution controlling parameters. The segmentation results of both the EBACM-KFCM_S1 and EBACM-KFCM_S2 methods are compared with the different variants of fuzzy c-means using spatial constraints (FCM_S1 and FCM_S2), KFCM_S1, KFCM_S2, geodesic active contour (GAC), region-based active contour driven by region scalable fitting (ACMRSF), edge-based active contour model (EBACM) applied on the large database of the US images. Their performances are estimated, quantitatively in terms of different performance measures such as true positive (TP), false positive (FP), accuracy (ACC), Jaccard similarity index (JSI), dice coefficient (DC) and Hausdorff distance (HD). The quantitative analysis shows that the results of the proposed segmentation methods provide higher segmentation accuracy compared to the others. It also provides better values of other performance metrics such as the TP, FP, JSI, DC and HD. For the proposed EBACM-KFCM_S2 method, when tested on the US images, the averaged performance measures such as the TP, ACC, JSI, and DC are higher than the EBACM-KFCM_S1 method. The proposed approaches also help to remove the need of manual intervention and decrease the processing time. vi In the second proposed segmentation approach, the features of the RBACM driven by the RSF energy and the two different variants of the Gaussian kernel fuzzy c-mean clustering with spatial information such as the GKFCM_S1 and GKFCM_S2 are utilized. In addition to the previously described methods in the earlier paragraph, two more hybrid segmentation methods using the RBACM approach have been proposed, which starts with the GKFCM_S1 and GKFCM_S2 approach, individually. The proposed approaches utilize the GKFCM_S1 and GKFCM_S2 clustering, individually, not only to initialize the contour, but also to estimate the several contour propagation controlling parameters. These proposed approaches are named as the RBACM-GKFCM_S1 and RBACM-GKFCM_S2. In these segmentation approaches, the intensity information in the local regions as against the global regions in conventional RBACM approach, are utilized to drive the motion of contour toward the desired object boundaries. The RSF formulation is also responsible for attracting the contour toward the object boundaries, thus increasing the speed of the contour propagation. The effectiveness of the proposed approaches is illustrated through several experiments performed on the similar US images, synthetic test images and also compared with the results of the above segmentation methods, including the RBACM using spatial fuzzy clustering (RBACM-SFCM) approach. From the experimental results, it is observed that the segmented images obtained by the proposed approaches are approximately similar to that of the manually delineated region compared to others. It is further observed that the proposed approaches achieve higher segmentation accuracy than the GAC, ACMRSF, RBACM-SFCM and EBACM, which itself signifies improvement in the results of the proposed approaches. These proposed methods also take comparatively less time to segment the image. For the purpose of implementing and evaluating the performance of the above discussed proposed methods, the US images were acquired from the image database available at (http://rad.usuhs.edu/medpix/parent.php3?mode=home_page), (http://ultrasonic s.bioenggineering.illinois.edu), (http://www.ultrasoundcases.info/), (http://radiologyinfo.org/en /photocat/), (http://thelivercarefoundation.org) and from the database of GE, Phillips and Siemens.