Abstract
Purpose:
Accurate assessment of aortic landing zones (LZs) is essential for planning thoracic endovascular aortic repair (TEVAR). Manual measurements, while the clinical standard, are time-consuming and prone to inter-operator variability. This study evaluates the reliability of an artificial intelligence (AI)-based tool for standardized LZ assessment.
Material and Methods:
A retrospective clinical data set of 35 patients was analysed using manual and an AI-assisted methods for evaluating key geometric parameters relevant to TEVAR planning: diameters at the endpoints of the proximal and distal aneurysm (P0, D0) and at the ends of the LZs (P2, D2), conicity indices (TR
P
, TR
D
), and lengths (L
PLZ
, L
DLZ
) of the proximal and distal LZs. Agreement was assessed using intraclass correlation coefficients (ICCs). The data set was stratified according to the anatomical zones of Ishimaru (Group A = 0, 1, and 2 and Group B = 3) to evaluate the influence of the anatomical location of the PLZ on the evaluation and the concordance of the method.
Results:
Artificial intelligence–assisted and manual measurements showed strong agreement, with ICC values consistently >0.75. No significant differences were observed in TR
D
between groups (2.2% vs 3.5%,
P
= .691). However, TR
P
showed a trend towards higher proximal conicity in Group A vs Group B (−10.6% vs 0.8%,
P
= .086), with reduced agreement in Group B (ICC = 0.62).
Conclusions:
Artificial intelligence–assisted methods showed high concordance with manual measurements for most aortic LZ parameters. However, anatomical complexity, especially near zone 0, can affect agreement, particularly in evaluating aortic neck conicity. These findings highlight the importance of considering anatomical variability when validating AI-assisted tools for clinical planning.
Clinical Impact
This study shows that AI-assisted, centreline-based assessment of aortic landing zones can be integrated into routine TEVAR planning, providing diameters, lengths and conicity indices that agree closely with manual measurements while reducing operator dependency and post-processing time. For clinicians, this translates into faster, standardised and reproducible sizing, less variability between observers and centres, and better-supported decisions on endograft oversizing. The innovation lies in coupling deep-learning segmentation with automated detection of the maximum-area cross-section and of a fixed 25 mm sealing zone along the inner curvature. The reduced agreement observed in anatomically complex arch zones indicates where expert review remains indispensable.