AI Algorithm Suggests Optimal Stent Size for Iliac Arteries

Published in CVIR Endovascular in August 2026, this proof-of-concept study demonstrates that a modular software algorithm can identify vascular stenoses, segment the affected vessel segments, and predict optimal stent diameter and length based on digital subtraction angiography images. The study addresses a significant gap in vascular medicine, where standardization and international guidelines for stent size selection are currently lacking, and the choice of stent remains a highly individualized decision relying heavily on operator experience.

The development of OSCAR was driven by the clinical necessity of precise stent sizing in peripheral artery disease. Stenosis morphology and anatomical location further influence stent choice; for instance, calcified lesions often require balloon-expandable stents due to their higher radial force. Despite the availability of numerous stent models from various manufacturers, there has been no widely accepted algorithm to suggest the optimal stent size based on angiography images for peripheral artery disease until now.

The study was a retrospective multicenter analysis that included 149 patients who underwent stent implantation for symptomatic stenoses of the common and external iliac arteries between August 2017 and July 2024. The cohort consisted of 100 male and 49 female patients, with a median age of 56.9 years. The patients received a total of 149 stents, distributed across the right and left common iliac arteries and the right and left external iliac arteries. The researchers focused on iliac artery interventions because these procedures are common, typically treated with primary stenting, and usually represent isolated stenoses, making them suitable for the initial development and validation of the algorithm.

OSCAR is built on a modular software architecture comprising three independent but interoperable components: stenosis detection, stenosis segmentation, and stent parameter prediction. The detection module uses a YOLOv8-nano-based architecture to autonomously identify vascular narrowing in angiography images. To enhance the model’s ability to generalize across different anatomies, the network was first pre-trained on a large publicly available coronary stenosis dataset containing 8,325 annotated angiographic X-ray images. It was then fine-tuned on the study-specific iliac artery dataset using a leave-one-out cross-validation strategy at the patient level, ensuring that each patient’s data remained unseen during the training process for that specific iteration.

The segmentation module isolates pathological vascular alterations with high spatial precision using a combination of classical image processing techniques and iterative refinement. After the detection module identifies a bounding box around a potential stenosis, the image is Gaussian-smoothed to suppress background noise and cropped to increase the lesion-to-background pixel ratio. An initial segmentation is generated via intensity thresholding, followed by morphological closing to reduce local discontinuities. This mask serves as the initialization for the GrabCut algorithm, which yields a cleaner vascular contour. The structure is then aligned vertically using principal component analysis, and the segmentation process is repeated on this aligned representation to improve consistency in diameter and length extraction.

For stent parameter prediction, the algorithm quantifies the geometric properties of the stenosis to recommend an appropriate stent diameter and length. A graph-based search identifies the vessel centerline from the proximal to distal end, automatically excluding side branches. The trajectory is smoothed using a moving-average filter to ensure a stable geometric reference. From this refined centerline, left and right vessel contours are extracted, averaged, and smoothed to obtain a representative stenosis profile. The lesion length is determined from the distribution of contour values using percentile-based thresholds, while the lesion diameter is derived from the 95th percentile of the contour values, providing a stable estimate that minimizes the influence of outliers.

In terms of performance, OSCAR detected stenoses in 84.6% of the 149 cases. The model achieved a high recall of 0.89, indicating that most expert-annotated stenoses were correctly identified, while a moderate precision of 0.65 indicated the presence of some false-positive detections. In two-thirds of the patients, the algorithm achieved a recall of at least 0.8, and in 64.4% of patients, it achieved perfect recall. The segmentation accuracy was good, with a mean Dice Similarity Coefficient of 0.77, which is comparable to the inter-observer variability among the four board-certified interventional radiologists who provided the ground truth annotations. This similarity suggests that the algorithm’s performance is approaching the limit of human consistency in this specific task.

The stent diameter and length predictions demonstrated mean absolute percentage errors of 0.13 and 0.33, respectively. These errors are comparable to the variability observed among expert annotators. For comparison, the average pairwise mean absolute percentage error between the ground truth segmentation-based stent parameters was 0.07 for diameter and 0.40 for length. The stent parameters determined from the predicted segmentation fell within the minimum and maximum range of the ground truth-based parameters in 31% of cases for diameter and 60% of cases for length.

The study acknowledges several limitations that must be considered when interpreting the results. The primary limitation is the relatively small number of patient datasets included in the work, which may affect the generalizability of the findings. The study focused exclusively on iliac lesions, and the transfer of OSCAR to other vessel locations, such as the coronary arteries, must be evaluated in future work. Additionally, the inter-observer variability in the ground truth annotations, which was mitigated through rigorous observer training, could introduce a certain bias to the segmentations used for training the algorithm. The current pipeline is also limited to the estimation of geometric sizing parameters and does not account for clinically relevant factors influencing stent type selection, such as calcification, vessel tortuosity, or lesion complexity.

Despite these limitations, the study demonstrates the feasibility of AI-supported stent choice during vascular interventions. The modular architecture of OSCAR allows for specific training of individual tasks, providing understandable and explainable outcomes at every step. The high recall of the detection module supports its potential as a reliable screening component in automated stent planning workflows, although the moderate precision highlights the risk of false positives that may necessitate human oversight. The researchers suggest that future work should focus on enhancing stenosis segmentation through data-driven methods, such as deep learning-based approaches, and on extending parameter prediction to such frameworks to improve accuracy and robustness.

From a clinical and economic perspective, the integration of OSCAR into established software platforms used in angio-suites could offer significant benefits. The algorithm has the potential to be integrated into hospital procurement systems, working as a closed-loop solution where implanted stents can be automatically reordered. Statistical analysis could estimate the probability of specific stent types being required, allowing for more accurate inventory planning and reducing the number of stents that expire before implantation. This approach may improve economic efficiency and sustainability in stent implantation procedures, addressing both clinical and logistical challenges in vascular medicine.

The findings of this study contribute to the growing body of evidence supporting the use of artificial intelligence in cardiovascular interventions. While many existing tools focus on coronary arteries, OSCAR represents a significant step forward in applying these technologies to peripheral artery disease. The ability to automatically suggest the optimal stent size based on angiography images could standardize a currently subjective process, potentially leading to improved patient outcomes by reducing the risks associated with undersizing or oversizing stents. As the field continues to evolve, further validation in larger multicenter cohorts and the incorporation of additional clinical and imaging data will be essential to confirm the robustness and generalizability of OSCAR for widespread clinical use.

Subject of Research: Validation of an AI-based algorithm for stent size selection in iliac artery interventions

Article Title: Introduction and validation of OSCAR—optimal stent choice algorithm

Article References: Wegner, F., Buhné, M.-J., Erben, N., Wulff, D., Stahlberg, E., Storch, A., Dünschede, F., Mogadas, S., Ströder, J., Jacob, F., Sieren, M. M., Barkhausen, J., & Kloeckner, R. (2026). Introduction and validation of OSCAR—optimal stent choice algorithm. CVIR Endovascular, 9(1), Article 108. https://doi.org/10.1186/s42155-026-00760-1

Image Credits: AI Generated

DOI: 10.1186/s42155-026-00760-1

Keywords: Artificial Intelligence, Vascular Medicine, Stent Sizing, Iliac Artery, Digital Subtraction Angiography, Peripheral Artery Disease, Medical Imaging, Interventional Radiology, Introduction, validation, OSCAR, optimal