Ultrason görüntülerinin havza sınırlama yöntemi kullanılarak bölütlenmesi

Ultrasound imaging is an important medical application as it is a safe, non-invasive, and interactive procedure for creating detailed visualization of various structures within the body. Computational methods help improving the quality of ultrasound images as well as they aid in the discovery of important structures. In this line of research, this thesis studies the segmentation problem in general, which aims to identify regions of interest of a given image, and in particular, focuses on its application on ultrasound images using the watershed algorithm. The thesis consists of three main parts: (i) preprocessing of the ultrasound images, (ii) the watershed algorithm, and (iii) postprocesssing the segmented images. Despite the technical advances in the clinical field, speckle noise continues to pose a challenge in the interpretation of ultrasound images. The first part of the thesis addresses this issue using preprocessing algorithms, namely; histogram equalization and median-based filters to reduce the speckle noise. This yields an initial improvement in the ultrasound image quality. Next, in the second part, the watershed algorithm is used on the filtered images to discover important regions. The algorithm works based on an analogy from geography where watershed lines separate catchment basins from each other, hence creating different segments. The algorithm is applied on both intensity values and local statistics which are based on mean and variance values. A well-known drawback of the watershed algorithm however is that it leads to oversegmentation. Therefore, in the third part, region merging algorithms are studied to overcome this issue. A novel region merging algorithm that incorporates several criteria is proposed. The criteria include the size (the number of pixels in a region), intensity (color values associated with each pixel) and statistical measures (such as variance and mean ratio). These algorithms include three parameters denoted as Tsize, Tincrement and Tmerge which are image dependent. Extensive tests run on a phantom cyst image and a clinical liver image. Our test results show that the numbers of regions for the cyst and liver images are reduced by about ~40 and ~50 times, respectively. Overall, we achieved a significant reduction in oversegmentation using a mixture of pre- and postprocessing. The results demonstrate the effects of each studied component; median-based filtering, watershed segmentation, and region merging.