Stent-RD-UNet: A Deep Learning Network for Real-Time Pose Estimation and Tracking of Aortic Stents in Image-Assisted Interventions
Summary
No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.
Summarise with VIRAT Free to try with your Science Catalogs account.Abstract
During endovascular aortic repair (EVAR/TEVAR), X-ray fluoroscopy / digital subtraction angiography (DSA) is used to confirm whether the stent graft and delivery system (DS) reach the lesion site, deploy correctly, and are retrieved safely. Device visualization is challenged by low contrast, sparse metallic lattice appearance, background clutter (spine edges, bowel gas, and overlapping instruments), and projection-angle variations, which collectively degrade segmentation-only solutions and limit real-time clinical usability. Methods: Intraoperative device understanding was formulated as a pose estimation and tracking problem for three categories: bifurcation stents (BS), straight stents (SS), and delivery systems (DS). DSA/fluoroscopy videos from 60 patients (36,418 frames) were collected, with 14,279 manually annotated images. Device-specific keypoint skeletons were defined (BS: 34 keypoints / 50 links; SS: 17 keypoints / 25 links; DS: 13 keypoints / 12 links), and supervision was constructed as keypoint heatmaps plus a skeleton probability map. We propose the Stent-RD-UNet, a U-shaped encoder–decoder network strengthened by Residual Dense Feature Fusion (M1), fusion refinement after skip concatenation (M2), and transposed-convolution upsampling (M3). Performance was evaluated using Center–Edge Dice (CE Dice), Object Keypoint Similarity (OKS), detection rate (DR), mean inference speed (mFPS), and HD95 computed on edge maps and skeletons. A structured ablation under both segmentation and pose settings quantified the contribution of M1 and M2 + M3. Results: Stent-RD-UNet achieved CE Dice of 0.94/0.95/0.94 and OKS of 0.96 ± 0.17/0.90 ± 0.05/0.91 ± 0.14 for BS/SS/DS, respectively, while maintaining real-time throughput (22.79/20.93/22.76 mFPS). The model reduced HD95(edge) to 4.2/3.9/3.7 px and HD95(skeleton) to 3.4/3.3/3.1 px for BS/SS/DS, respectively. The ablation confirms that M1 and M2 + M3 consistently improve CE Dice, DR, and HD95 across all device categories. Conclusion: Stent-RD-UNet enables accurate, real-time pose estimation and tracking of aortic stents and delivery systems under fluoroscopy/DSA. Pose estimation provides stable edge and endpoint cues for hollow, deformable lattices, and boundary-sensitive HD95 metrics complement CE Dice to reflect clinically relevant deviations.
Fields of Study & Subjects
Medical Image Segmentation Techniques
Advanced Neural Network Applications
Medical Imaging and Analysis
Publication details
- Volume
- 1
- Issue
- 1
- Pages
- 109–125
- Publication date
- 24 March 2026
- Language
- EN