Automatic Explainable Segmentation of Abdominal Aortic Aneurysm from Computed Tomography Angiography

Introduction: : Abdominal aortic aneurysm (AAA) is a dilation of the abdominal aorta where its normal diameter increases by more than 50%. If there is no proper treatment, it gradually grows and leads to an increased risk of rupture. In the last ten years, there has been a notable change in the way aortic aneurysms are treated. Instead of relying primarily on open surgery, clinical experts have increasingly turned to a less invasive method called endovascular aneurysm repair (EVAR), as outlined by Moll et al. AAA screening programs have shown that population-based screening significantly reduces AAA-related mortality, as indicated by several trials and meta-analysis. The Viborg study reported a prevalence of 4.0%, while the Chichester study reported 7.7%, emphasizing variation in different populations (Lindholt and Norman). Screening programs are essential tools that have proven effective in AAA detection and can reduce specific and overall mortality, especially in the male population (Thompson et al., Sweeting et al.). Screening in women has long been a topic of debate despite being currently indicated as an at-risk population and suggesting the diameter of the aneurysm indexed to body size as a predictor of rupture (Spinella et al.). CTA scans, often used to diagnose other conditions, incidentally detect AAA in approximately 1.9% to 8.3% of the cases. These incidental findings, although different from the dedicated screening results, are important for effective monitoring and elective AAA repairs. Effective clinical management of AAAs still depends on early discovery and monitoring, even with improvements in medical imaging and surgical techniques.

Materials and

Methods: : The dataset used in this study consists of 2D high-resolution CTA images collected from two clinical centers: Allegheny General Hospital (Pittsburgh, PA) and Northwestern Memorial Hospital (Chicago, IL). Ethical clearance and Institutional Review Board (IRB) approvals were obtained prior to data collection. The axial CTA slices have resolutions of 512 × 512 pixels, with a slice thickness ranging from 1 mm to 3 mm. Ground truth annotations were generated using a semi-automated tool, AAAVasc (version 0.168), developed at The University of Texas at San Antonio, which detects both lumen and vessel wall boundaries.
To achieve accurate and explainable segmentation of AAA, we propose a framework that integrates multiple specialized U-Net models and adaptive routing mechanisms. The complete architecture is shown in Figure 1. It includes three expert U-Net models trained with diverse objectives, an error learner router and a robust weight router to perform spatially adaptive fusion. Each expert specializes in different structural attributes of AAA. In scenarios where segmentation fails (e.g., in low-contrast or thin-wall regions), misclassified areas are corrected using a NURBS-based interactive tool. This allows clinicians to manually adjust the boundaries through control points, yielding a smoothed and updated segmentation mask, as shown in Figure 2.

Results, Conclusions, and Discussions:: The model accurately segments the aortic lumen and the outer wall, with DS scores of 0.9648 and 0.9615, and IoU scores of 0.9324 and 0.9264, respectively. The 95th percentile Hausdorff distances were 1.3490 mm and 1.3670 mm. The model processed each image in about 17 ± 1 milliseconds, demonstrating that it can work in real time. In complex cases, such as noisy frames or very thin walls, the built-in NURBS correction tool allowed experts to adjust the results within 3 to 20 seconds per frame, which helped improve boundary precision.
The predicted shape and size measurements were similar to the expert-labeled values. For example, the neck lumen area was 3.9220 cm² in the ground truth and 3.9717 cm² in the prediction. The maximum diameter was 5.8105 cm compared to the predicted 5.8586 cm. The differences in the areas of the left and right iliac arteries were small, always within ±0.05 cm², showing the model’s utility for planning surgeries and choosing graft sizes.
Figures 3 show the model’s strong agreement with expert annotations in 22 patients, highlighting its robustness and consistency in segmenting patient-specific anatomical variability. The model includes attention-based mechanisms and error-aware routing to identify and segment the lumen and outer wall using CT imaging. This approach improves segmentation performance while enhancing interpretability through explainable feature representations. Furthermore, the integrated NURBS-based interface enables clinicians to perform interactive refinements in complex cases, facilitating improved boundary delineation and overall segmentation accuracy.
Compared to existing models, the proposed approach shows good performance in terms of DS, IoU and HD95%, indicating improved delineation accuracy and enhanced anatomical structure. In general, the accuracy, speed, and explainable design of the model make it a useful tool for diagnosing and planning AAA treatment. It also handles complex images and allows experts to fine-tune the results, which makes it useful in a clinical setting.

Acknowledgements and/or References (Optional):: Ender Finol, Author, thanks the Zachry Mechanical Engineering Department Chair Endowment and the National Institutes of Health (Grant No. R01HL159300) for providing support for the execution of this work.