Abstract
Guidewire tracking and segmentation are critical for navigating endovascular procedures, yet they remain challenging due to low contrast, anatomical noise, and real-time constraints. This paper presents a two-stage deep learning framework that achieves state-of-the-art performance. The first stage utilizes a spatiotemporally refined YOLOv5 detector to localize the guidewire, while the second stage employs a novel "HessianNet" for precise segmentation. By incorporating a physics-informed Hessian-based enhancement layer and a dual self-attention module, the system achieves a Dice score of 0.8990 and a processing speed of 35 FPS, enabling real-time surgical assistance and potential robotic integration.
Introduction
During endovascular interventions such as stenting or ablation, clinicians must navigate thin, flexible guidewires through complex vasculature under X-ray fluoroscopy. This task is hindered by several factors: the wires are extremely thin with low contrast, anatomical structures like ribs can mimic their appearance, and the wire occupies only a tiny fraction of the image pixels (extreme class imbalance). Furthermore, any automated system must operate in real-time (at least 25-30 FPS) to be clinically viable.
Methodology
The proposed framework addresses these challenges through a hierarchical approach consisting of localization followed by fine-grained segmentation.
Guidewire Localization and Refinement
The system first employs a YOLOv5s detector to identify a bounding box around the guidewire. To mitigate false positives caused by static artifacts or noise, a spatiotemporal refinement module is introduced. This module analyzes three consecutive frames to ensure that detected motion is physically consistent with a moving guidewire, effectively reducing false positives to zero in experimental trials.
HessianNet Segmentation
Once localized, the guidewire is segmented within the cropped region using HessianNet. This network introduces two primary innovations:
- Hessian-based Enhancement Layer: Traditional Hessian-based filters are reformulated as trainable neural network layers. This allows the model to leverage geometric priors for tube-like structures while remaining adaptive to data.
- Dual Self-Attention Module: To capture long-range spatial dependencies efficiently, the network uses a dimensionality-reduced attention mechanism via max-pooling.
Experimental Results
The framework was evaluated on a clinical dataset of 102 X-ray sequences (5,238 images) from 21 patients. Data augmentation, including elastic deformations and synthetic image generation, was used to enhance robustness.
Performance and Robustness
HessianNet achieved a Dice score of 0.8990, significantly outperforming baseline models such as U-Net and ResUnet. Qualitative results demonstrate that HessianNet maintains high continuity in segmented wires even in regions of low contrast where other models fail.
Real-time Capability
The entire pipeline processes images at approximately 35 FPS on an NVIDIA Quadro RTX 6000. This speed exceeds the standard fluoroscopy frame rate, making it suitable for live intraoperative use.
Discussion and Conclusion
By combining physics-informed layers with modern deep learning architectures, this work provides a robust solution for real-time guidewire tracking. The two-stage approach effectively handles class imbalance and reduces computational overhead. While currently optimized for single-wire scenarios, this framework establishes a foundation for future developments in multi-wire tracking and autonomous robotic surgery.
Summary updated