Artificial intelligence–based fully automated multi-class segmentation of aortic dissection in computed tomography images

Background Computed tomography angiography (CTA) is the first-line imaging modality for aortic dissection, but manual segmentation remains time-consuming and operator-dependent, limiting standardized morphological assessment in routine practice. Purpose We aimed to develop and validate a fully automated deep-learning framework for multi-class segmentation of aortic dissection pathology and aortic anatomical zones on CTA. Methods We developed a deep learning architecture based on the nnU-Zoo framework, comprising two parallel modules for dissection pathology and anatomical zone segmentation. The model was developed using public datasets (ImageTAAD and AortaSeg24) and enriched with expert manual annotations of clinically relevant structures, including refined true/false lumen labels, the aortic root, coronary arteries, and intraluminal thrombus (ILT). For large-scale multicentre validation, model inference was performed on 1,000 CTA cases from internal and external cohorts, including Type A (TAAD), Type B (TBAD), and non-A/non-B dissections. Segmentation performance was benchmarked against ground-truth annotations. Results The model demonstrated robust quantitative performance, achieving a mean Dice similarity coefficient (DSC) of 0.79 on the AortaSeg24 validation set and 0.68 on ImageTAAD. It showed high anatomical robustness, with reliable segmentation of major aortic zones (0-11) and key branch vessels, including the renal, celiac and superior mesenteric arteries. Large anatomical structures, including the true lumen, false lumen and aortic root, achieved excellent accuracy (median DSC >0.80). In contrast, segmentation of smaller or morphologically irregular pathological features, particularly ILT and entry tears, yielded lower performance and remained more challenging. Conclusion This study presents a fully automated deep-learning framework for multi-class CTA segmentation of aortic dissection, validated across multiple centres and dissection subtypes. The model delivers robust luminal delineation and zonal anatomical mapping, offering potential to facilitate standardised imaging assessment within clinical workflows. Further refinement is warranted to improve segmentation of small, complex pathological structures, particularly entry tears and ILT.