Impact of CT Acquisition Parameters on Deep Learning of Aortic Segmentation Performance: Systematic Review

Automatic segmentation of computed tomography (CT) images is fundamental for quantitative anatomical analysis in a wide range of clinical applications. Despite remarkable advances in artificial intelligence (AI), CT acquisition parameters critically affect image quality, structural fidelity, and model performance. This narrative review investigates how variations in CT acquisition influence image quality and accuracy of AI-based aortic segmentation models. Following PRISMA guidelines, a narrative review was conducted to identify studies assessing how CT technical parameters influence image quality and AI-based aortic segmentation accuracy. From 376 initial records, 13 studies met the inclusion criteria. Thinner slices improved segmentation accuracy, increasing the Dice Similarity Coefficient (DSC) from 0.75 to 0.87 and reducing the 95th Percentile Hausdorff Distance from 2.1 to 1.2 mm. Larger slice spacing (3 mm) worsened accuracy by about 12%, while 1 mm slices reduced aortic diameter errors to below 2%. Faster acquisitions decreased image noise from 32.3 ± 6.3 to 22.1 ± 5.9 HU. Among segmentation models, 2D-3D U-Net ensembles achieved the highest accuracy (DSC = 0.928 ± 0.026), with thrombus segmentation (DSC = 0.782 ± 0.170) outperforming vessel segmentation (DSC = 0.481 ± 0.155). Optimal pixel spacing (0.665 × 0.665 mm) improved image fidelity, yielding a PSNR of 16.28 dB and lower MSE (2729.67). CT acquisition parameters critically impact the reliability of AI-driven aortic segmentation. Optimizing these parameters improves model performance, but the opacity of AI systems still limits clinical interpretability, highlighting the need for standardized acquisition protocols and explainable AI approaches.