Segmentation of the Aorta in CTA Images Using Deep Learning Methods

Abstract

Abstract Doctors manually segmenting patient CT images is both time-consuming and labor-intensive. Additionally, classic image processing techniques are insufficient in non-contrast CT images because the pixel values of tissues are very close. Automatic segmentation of the aorta in human anatomy can be a useful clinical application that can help reduce the workload of healthcare workers in preoperative planning. In this study, the segmentation of the thoracic aorta, abdominal aorta, and iliac arteries in contrast and non-contrast CT images was performed using U-Net, U-Net attention, and Inception U-Netv2 segmentation models. First, 2D axial images were extracted from all datasets. Preprocessing such as resizing, gray level normalization and histogram equalization were applied to the resulting axial images. The edge structures of the aortic structure were determined using the Contrast limited adaptive histogram equalization (Clahe) method. Then, 5-Fold Cross Validation was applied to the segmentation models to perform training and test operations. The resulting 2D sections from the test were merged to make a 3D structure and the spatial coordinate information of the original image was transferred to the predicted mask. The 3D image was improved by removing small objects incorrectly defined as negative around the 3D aortic segmentation obtained. In this study, the test results obtained from the Dongyang and KITS dataset, a U-Net model gave a 89.5% Dice, 81.0% IoU, 86.9% sensitivity, and 99% specificity score, a U-Net attention model gave a 89.7% Dice, 81.3% IoU, 87.3% sensitivity and 99% specificity score and Inception U-Netv2 model gave a 90.4% Dice, 82.7% IoU, 89.1% sensitivity and 99% specificity score. The Inception U-Netv2 model gave the highest predictive segmentation results.