Fully Automated AI-Driven Quantification of Thoracic Aorta Dimensions from Contrast-Enhanced MRA

Background: Aortic valve disease (AVD) affects 2-3% of U.S. adults and can lead to severe complications (e.g., aneurysm, dissection)^11,12. Early detection is crucial. Guidelines¹⁰ recommend imaging (CT or MRA) to assess aortic diameters and growth rates at defined landmarks, as primary criteria for intervention. Extracting diameters from 3D angiography requires time-consuming manual analysis and is limited by observer variability^13,14. There is an unmet need for automated diameter quantification, to reduce analysis time and increase measurement reliability. In this study, we report on the feasibility and performance of an AI-driven, fully automated pipeline for 3D aorta segmentation and diameter analysis, for contrast-enhanced MRA (CE-MRA) scans.

Methods: Fig. 1 illustrates the analysis pipeline, which includes AI models for 3D thoracic aorta segmentation. This was based on 234 bicuspid aortic valve patient CE-MRA scans (50.7±13.6 years, 75 female). The thoracic aorta was manually segmented to generate global and regional (ascending aorta [AAo], arch, descending aorta [DAo]) ground-truth masks. 3D Dense U-Net^1,2,5,7 models were trained (n=174) and tested (n=60). Geometry quality control (gap detection, aspect ratio and convex hull analysis) rejected defective AI segmentations. A centerline was calculated and perpendicular 2D analysis planes were fitted at region boundaries (distal ascending [distAAo], proximal descending [proxDAo] aorta). A mid-ascending aorta (midAAo) plane was placed at the centerline midpoint between distAAo and the sinotubular junction (initialized 1 cm from centerline start); a mid-descending (midDAo) plane at the corresponding superior-inferior axis location in the descending aorta. For each analysis plane, lumen boundaries were fit to CE-MRA signal intensity (Chan-Vese⁴, Canny edge detection³). To capture plane placement variability, diameters were computed stochastically, placing ten planes at 0.5 mm intervals proximally and distally to plane locations, and taking the mean. Diameters were extracted as the major axis length of fitted ellipses. AI quantification was compared with available clinical ground truth (n=26-27).

Results: Automated analysis was completed for 92% of cases. Segmentation results are presented in Fig. 2, showing excellent performance (median Dice coefficient of 0.93 and 0.94 for global and arch segmentation). Diameter results are presented in Fig. 3. Good to excellent agreement with manual measurements was achieved (ICC(3,1): 0.74-0.96), with >±1.0 mm bias at distAAo (-1.5 mm) and midDAo (-2.0 mm).

Conclusion: Our study shows promising performance for fully automated assessment of aortic dimensions from CE-MRA data. Discrepancies mainly arose from minor segmentation errors (contrast, artifacts), centerline fitting constraints, and variability in multi-observer ground truth. Future directions include evaluating advanced segmentation approaches (nnU-Net^8,9, transformer¹⁵-/Mamba⁶-based U-Nets, ensembling) and end-to-end optimization.

Figure 1: Overview of the automated diameter analysis pipeline. Input CE-MRA scans are preprocessed and input into global and regional segmentation AI models. Binary masks are automatically rejected if diameter analysis is infeasible (e.g., due to large gaps). Segmentation into regions follows, followed by centerline calculation and diameter analysis plane placement. CE-MRA intensity is used to refine contours at diameter analysis planes, where an ellipse is fit. Diameters are calculated by measuring the length of the major axis, and averaging over adjacent planes (stochastic measurement).
Figure 2: Segmentation results (ranked by Dice similarity coefficient [DSC]), with indicative analysis plane placement and diameter values. Analysis was completed for the first (best-DSC) and second (median-DSC) cases. The worst-DSC case failed quality control. Regional segmentation results are presented in terms of the arch region.
Figure 3: Validation of automated aortic diameter measurements against manual ground truth at four key anatomical locations (n=26-27 paired measurements). Mid-ascending aorta (midAAo) shows excellent reliability (ICC(3,1)=0.96, 95% CI: 0.92-0.98) with minimal bias (0.8 mm) and mean absolute error of 2.0 mm. Proximal descending aorta (proxDAo) shows good agreement (ICC(3,1)=0.84, 95% CI: 0.73-0.90) with negligible bias (0.1 mm) and excellent precision (93% of measurements within ±2 mm). Underestimation is observed at the distal ascending aorta (distAAo; bias=-1.5 mm, ICC(3,1)=0.86) and mid-descending aorta (midDAo; bias=-2.0 mm, ICC(3,1)=0.74). Limits of agreement ranged from ±3.0 mm (proxDAo) to ±5.4 mm (distAAo).