TotalSegmentator-AI for aortic segmentation in CT: Reliable performance in normal anatomy but limited utility in pathological aortic disease.
Authors
Affiliations (1)
Affiliations (1)
- Department of Diagnostic and Interventional Radiology, Lausanne University Hospital and University of Lausanne, Rue du Bugnon 46, Lausanne 1011, Switzerland.
Abstract
Accurate aortic segmentation on computed tomography angiography (CTA) is essential for diagnosing aortic disease, cardiovascular risk assessment, and surgical planning. Deep learning algorithms, such as TotalSegmentator-AI, offer fully automated multi-organ segmentation, yet their performance in pathological aortic conditions remains uncertain. This study performs a clinical stress-test of TotalSegmentator-AI, mapping its boundaries and structural failure modes across a spectrum of normal and pathological cases. In this monocentric, retrospective study, 60 CTA scans from 2014 to 2024 were categorized into six groups: young, elderly, aneurysm, dissection, venous phase, and non-contrast phase. TotalSegmentator-AI was applied without manual correction. Two radiologists independently rated six aortic segments per scan using a five-point qualitative scale. Quantitative segmentation errors were correlated with qualitative scores using Spearman's correlation, and inter-reader agreement was assessed with weighted Cohen's kappa. All scans were successfully processed, yielding 360 aortic segments. Median segmentation quality was 4 [IQR 4-5], with 77% rated good or excellent. Performance was consistent across segments (p = 0.16) but varied by category (p < 0.001): best in young patients (5 [IQR 4-5]) and adequate in non-contrast and venous-phase scans (4 [IQR 4-5]), poorest in dissections (3 [IQR 3-4]) and aneurysms (4 [IQR 3-4]). A strong negative correlation was observed between qualitative scores and quantitative errors (ρ = -1, p = 0.017). Inter-reader agreement was substantial (κ = 0.72). TotalSegmentator-AI achieves accurate aortic segmentation in normal anatomy but is inadequate for unsupervised clinical use in complex pathologies like aneurysms and dissections. Comprehensive human-in-the-loop quality control or dedicated pathology-inclusive models are mandatory before AI-based segmentation can be safely integrated into vascular clinical workflows.