Automated Segmentation of Cerebral Aneurysm Using a Novel Statistical Multiresolution Approach

Abstract

Cerebral Aneurysm (CA) is a vascular disease that threatens the lives of many adults. It a ects almost 1:5 - 5% of the general population. Sub- Arachnoid Hemorrhage (SAH), resulted by a ruptured CA, has high rates of morbidity and mortality. Therefore, radiologists aim to detect it and diagnose it at an early stage, by analyzing the medical images, to prevent or reduce its damages. The analysis process is traditionally done manually. However, with the emerging of the technology, Computer-Aided Diagnosis (CAD) algorithms are adopted in the clinics to overcome the traditional process disadvantages, as the dependency of the radiologist's experience, the inter and intra observation variability, the increase in the probability of error which increases consequently with the growing number of medical images to be analyzed, and the artifacts added by the medical images' acquisition methods (i.e., MRA, CTA, PET, RA, etc.) which impedes the radiologist' s work. Due to the aforementioned reasons, many research works propose di erent segmentation approaches to automate the analysis process of detecting a CA using complementary segmentation techniques; but due to the challenging task of developing a robust reproducible reliable algorithm to detect CA regardless of its shape, size, and location from a variety of the acquisition methods, a diversity of proposed and developed approaches exist which still su er from some limitations. This thesis aims to contribute in this research area by adopting two promising techniques based on the multiresolution and statistical approaches in the Two-Dimensional (2D) domain. The rst technique is the Contourlet Transform (CT), which empowers the segmentation by extracting features not apparent in the normal image scale. While the second technique is the Hidden Markov Random Field model with Expectation Maximization (HMRF-EM), which segments the image based on the relationship of the neighboring pixels in the contourlet domain. The developed algorithm reveals promising results on the four tested Three- Dimensional Rotational Angiography (3D RA) datasets, where an objective and a subjective evaluation are carried out. For the objective evaluation, six performance metrics are adopted which are: accuracy, Dice Similarity Index (DSI), False Positive Ratio (FPR), False Negative Ratio (FNR), speci city, and sensitivity. As for the subjective evaluation, one expert and four observers with some medical background are involved to assess the segmentation visually. Both evaluations compare the segmented volumes against the ground truth data.