Automatic aorta segmentation is important in medical imaging for accurate diagnosis and treatment planning. To allow for minimal invasive repairs for aortic diseases, a detailed 3D analysis of the aortic and branch vessel anatomy is essential. In this study, we propose an enhanced multi-class segmentation approach that combines region-based, voxel-wise, and topological loss functions within the nnU-Net framework. Our method was evaluated on the official test set of the AortaSeg24 challenge, achieving a validation Dice score of \(0.752 \pm 0.052\) and a Normalized Surface Distance (NSD) of \(0.782 \pm 0.055\) , representing an improvement over the challenge baseline (Dice: \(0.723 \pm 0.058\) , NSD: \(0.746 \pm 0.067\) ). Due to the challenge’s inference time constraints, a trade-off between ensemble size and sliding window step size was required. Experiments showed that using a larger ensemble of networks with an increased sliding window step size yielded better performance than using fewer networks with a smaller step size. The code for inference is available on Github .