Accurate segmentation of the carotid artery in cross-sectional ultrasound images is essential for vascular assessment, intima-media thickness measurement, and plaque evaluation. Although deep learning approaches have shown promising performance in vascular imaging, segmentation remains challenging in anatomically complex regions due to increased structural variability and irregular artery morphology. This study evaluates fully automated, segmentation-based deep learning models for carotid artery segmentation in cross-sectional ultrasound images from the VIPVIZA cohort. The dataset includes anatomically diverse regions such as the common carotid artery (CCA), carotid bulb, and split segments, which introduce morphological variability, non-circular artery shapes, and complex surrounding tissue structures that increase segmentation difficulty. Five deep learning models were investigated: YOLOv8, YOLOv11, U-Net, Attention U-Net, and DeepLabv3+. Images were preprocessed using intensity normalization and contrast enhancement. Performance was evaluated using the Dice similarity coefficient, Intersection-over-Union (IoU), and pixel-wise accuracy. The results demonstrate that while all models achieved stable performance in the relatively homogeneous CCA region (Dice up to 0.924 ± 0.053), segmentation performance decreased in anatomically complex areas such as the bulb and split regions. The split region posed the greatest challenge for all models, with lower Dice scores overall. DeepLabV3 consistently achieved the highest performance across all regions, including the bulb (Dice 0.914 ± 0.055) and split (Dice 0.852 ± 0.112), indicating accurate boundary delineation in morphologically variable areas. YOLO-based models demonstrated competitive performance, though with slightly lower overlap metrics. These findings highlight the importance of evaluating segmentation models across anatomically diverse subregions to ensure robust and reliable performance.