FEATURE EXTRACTION IN ULTRASOUND MEDICAL IMAGES

Abstract

Ultrasound imaging has an important place in the domain of medical imaging technologies because of its cost effectiveness, non-invasiveness and easy to use. However, the diagnosis of a disease based on medical images is still restricted to the clinician's visual interpretation of the image and his/her judgment based on his/her experience and expertise. 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 helping the clinician in their diagnostic decision. However, the processing of the ultrasound medical images is a difficult task because of presence of speckle and other artifacts in the ultrasound images. The main objective of the thesis has been to develop feature extraction methods for ultrasound medical images. To achieve this objective, the segmentation of ultrasound images in both textured and non-textured modes has been aimed. Further to improve the performance of segmentation algorithms, few pre-processing algorithms especially suited for ultrasound medical images were planned to be developed either by improving the earlier algorithms or by suggesting new algorithms. More specifically, the aim has been to design and develop effective algorithms for speckle reduction and feature preservation for robust segmentation of ultrasound medical images to help in efficient medical diagnosis. The entire work is divided into three parts, viz. 1) artifacts removal and feature preservation, 2) texture based edge detection, 3) automatic segmentation for texture and non-texture ultrasound medical images. Under the artifacts removal and feature preservation in ultrasound medical images, textural patterns in these images are considered as noise (speckle). The main focus in this part has been to enhance the relevant medical details (features) by suppressing the noise. The noise suppression in resultant facilitates true segmentation and analysis of the image. Unlike CT and MR images, ultrasonic images cannot generally be easily segmented and classified into separate anatomical or structures elements. The main disadvantage of medical ultrasonography is the poor quality of images. The ultrasound B-scan images xv have some special features that must be preserved during the filtering such as, contrast of the organs, delineation contour between organs, and structures with dimension comparable to speckles size. The analysis of these images, in general, is complex due to data composition, which is described in terms of speckle formation dependent on the geometry and sub-wavelength structure of internal tissue. The presence of speckle is undesirable since it degrades image quality and it adversely affects the tasks of interpretation and diagnosis. It also affects the accurate detection of the abnormal tissues in the images during segmentation and classification procedure in high-level image processing stage. The suppression of this noise and artifacts is a relatively difficult task. Several techniques for suppressing speckle noise have been proposed and developed by the researchers. But, these types of speckle-filtering techniques have major limitations in edge preservation. Most of them involve the use of general statistical approach and butdo not consider the special features typical for ultrasound medical images (e.g. structures with dimensions comparable to the speckle size and boundaries between two adjacent regions with slightly different in-homogeneity). For preserving the feature values, the various local statistical based adaptive filters such as adaptive weighted median filter (AWMF), adaptive speckle suppression filter (ASSF), homogeneous region growing mean filter (HRGMF) have been applied. These statistical based adaptive filters are sensitive to the size and shape of the filtering kernel. The refined version of the adaptive filter based on the region growing and local statistical features the aggressive region growing filtering (ARGF) has been proposed and implemented by few researchers. This technique is used to avoid premature blockage of the filtering region growth due to speckle. In the present work, ARGF is further improved by the second order statistical parameters (gray level co-occurrence matrix (GLCM)) based homogeneous region selection and region aggregation. The proposed filter is termed as GLCM based adaptive speckle suppression filter (GASF). The experimental results show that the proposed method is able to preserve the edges and their directions while smoothing the speckles. The performance this method is compared with aggressive region growing filter and better performance is validated by edge preservation test, speckle reduction test, and structural similarity test. xvi Further, the conventional filters such as median filter, Wiener adaptive filter, Lee/Frost/Kuan/Gamma filter do not enhance the edge features, rather they inhibit the smoothing perpendicular to edges i.e., the high contrast speckle/noise is retained in the neighborhood of edge points. In vicinity of an edge, complete smoothing should be prevented surrounding the edge. To accomplish this, smoothing is encouraged in the directions perpendicular to the edge and inhibited in directions parallel to the edge. For this purposes, anisotropic diffusion based speckle reduction filter is implemented with modified diffusion coefficients. The modified coefficients, in speckle reduction anisotropic diffusion filter perform smoothing and preserve edges better than the conventional speckle reduction anisotropic diffusion filter. The modified version of the speckle reducing algorithm using anisotropic diffusion is formally known as modified speckle reducing anisotropic diffusion (MSRAD) filter. Under the textured-based segmentation of ultrasound medical images, speckles, textural pattern of the soft tissues and artifacts are considered as the texture. In the process of the texture-based segmentation, the first step is to calculate the edges between the two adjacent textural patterns. The non-textural approaches are not used to detect the texture edges with weak mean gray level difference between the two textural patterns. Since the non-textural approaches inherently consider the speckles and tissue-related textures as noises, and discard them, the edge detection performance may be seriously deteriorated by these approaches by generating false edges of the texcels. The nontextural approaches mayalso lead to a shiftof the true edgeposition, or smear the desired edges, if a substantial denoising or speckle reduction operation is applied. In present research work, various textural approaches have been studied and applied to detect the texture edges. Anew method is proposed for texture edge detection based on early vision model. This model is based on the Gabor filter based feature extraction and used to detect the texture edge with improvement in the conventional edge flow vector. The comparison is based on the boundary detection between texture patterns using conventional edge flow propagation and proposed edge flow vector. The results show the better boundary detection using proposed method. The false and undesired edges into the ultrasound medical images are highly suppressed using proposed edge flow vector. xvn The edges obtained from the proposed textural-based method provide significant clues for segmentation. These edge profiles are further used for high-level processing with the hybrid technique ofregion and edge based segmentation. The final stage ofthe texturalbased approach aims at segmenting the two different textural patterns with weak mean gray level difference. The automatic partitioning and boundary evolution aims at high-level description of ultrasound medical image from the low-level description. In ultrasound medical image analysis, partitioning is an important task for delineation of the contour of anatomical structures within an image and for qualitative and quantitative ultrasound image analysis. It includes visualization and quantification of tissue volume, diagnosis, localization of pathology, and study of anatomical structure. The accurate segmentation provides the more meaningful extraction of the information from the ultrasound images for clinical purpose also. In qualitative study, image segmentation may provide the better visualization of surface of the object of interest. For ultrasound medical images, many methods were recently employed for image segmentation, i.e. texture-based analysis, deformable surface models, region based segmentation, statistical shape models, adaptive threshold techniques, edge-guided boundary delineation approach and hybrid techniques of edge-based and region-based approach. Most of these segmentation algorithms are semi-automatic and require human intervention for initial boundary specification, in order to specify the lesion center and then to identify the region of interest containing the lesion in the image. In the present work, anautomated multi-stage watershed algorithm is applied on the filtered image for accurate delineation of the anatomical structures from the image background. The region boundary obtained from this method is used in boundary evolution algorithm. Thereafter, the region based post processing is applied for final contour identification. The final contour identification using proposed approach is compared with the marker based watershed segmentation technique. The comparison is based on the over segmentation reduction and reducing extraneous regions during the segmentation procedure. In the comparison, the results obtained by the proposed technique show the better contour identification than marker based approach. The over segmentation problem in this approach is sufficiently overcome as compared to the conventional marker based approach. xvin For implementing and evaluating the performance of above discussed proposed methods, the ultrasound medical images were collected from the IIT Roorkee Hospital and from the database of GE, Philips, Siemens and Toshiba. In summary, the present work contributes in the area of ultrasound medical image analysis ranging its horizons in speckle noise suppression and automatic segmentation of textural and non-textural ultrasound medical images. The results obtained by the suggested methods may be useful for the clinicians, radiologist, orexperts, who may use it for clinical diagnosis and treatment planning.