Comparison of Deep Learning Architectures for Thoracic Aorta and Calcium Segmentation in Contrast-Enhanced Computed Tomography

Accurate segmentation of the thoracic aorta and aortic calcium in contrast-enhanced CT angiography (CTA) is essential for risk assessment and planning procedures like transcatheter aortic valve implantation (TAVI). This study presents a deep learning model capable of segmenting the aorta and quantifying calcium deposits in five anatomical regions—valve, tubular aorta, superior arch, inferior arch, and descending aorta—using only contrast-enhanced CTA, eliminating the need for non-contrast scans. Several neural network architectures were evaluated, including 2D/3D UNets, SegFormer, and UMamba. Among them, the 2D UNet with an EfficientNet-B0 backbone achieved the best performance, with a mean Dice score of 0.94 for aortic segmentation and 0.888 for calcium segmentation. This method enables region-specific calcium scoring directly from CTA, enhancing procedural planning and risk stratification. By leveraging routine clinical imaging, the model streamlines the diagnostic workflow, reduces additional imaging requirements, and provides detailed calcium assessments that support clinical decision-making in TAVI candidates.