Vessel segmentation on CT images using deep learning methods

The aorta is the most important artery in the human body and is responsible for carrying blood from the heart to all other organs. Diseases affecting the aorta, such as atherosclerosis, aortic dissection, and abdominal aortic aneurysm, can be fatal if not diagnosed early. Today, physicians use computed tomography and magnetic resonance imaging techniques to examine the structure of patients' aortas. These imaging techniques are necessary for analyzing the aorta, identifying deformed regions, and planning and monitoring surgery. In medical imaging applications, image processing techniques are used to analyze the structure of the aorta in computed tomography images. Segmentation of the three-dimensional aorta structure is quite challenging with these techniques. In recent years, deep learning methods have been used to overcome segmentation problems encountered with traditional methods. In this thesis, the segmentation of the contrast-enhanced aorta structure was performed using deep learning methods. First, two-dimensional images and masks were resized to 256x256. Window level and normalization operations were applied to the images during preprocessing. U-Net, Attention U-Net, Inception U-Netv2, and LinkNet models were trained using three-dimensional computed tomography images. After training, the two-dimensional mask outputs were combined to obtain a three-dimensional output. The obtained outputs were further processed by removing small objects and transferring the pixel range and center coordinate information of the input image to the mask scanning process, allowing them to be visualized in three dimensions. The highest success rates in the DONGYANG dataset were obtained with the Inception U-Netv2 model, with a 93.5 % dice similarity coefficient, 87.8 % Jaccard, 92.8 % specificity, and 100% sensitivity. The highest success in the KıTS dataset was achieved with the Inception U-Netv2 model, with an 86.58 % dice similarity coefficient, 79.78 % Jaccard, 83.18 % specificity, and 100 % sensitivity.