Simulation-based framework for guidewire detection and tracking in fluoroscopy: from synthetic sequences to clinical generalization

Le résumé fourni par la source

Real-time guidewire localization and tracking during percutaneous coronary interventions (PCI) is critical for navigation yet remains technically challenging due to patient and guidewire motion as well as limited vessel visibility without contrast agent. Frequent contrast injections are unsuitable for continuous guidance due to potential risk to the patient. Despite this clinical need, lack of publicly available intraoperative image data makes it difficult to develop a robust guidewire segmentation and tracking algorithm. We present a simulation-based framework that generates realistic, annotated fluoroscopic sequences of guidewire navigation under physiologically plausible motion. Starting from a static 3D CTA volume, we simulate anatomical deformation using a cardio-respiratory motion model. Respiratory motion is generated via amplitude-controlled superior-inferior and anterior-posterior displacements, while cardiac motion is characterized using a PCA-based motion model derived from population 4D cardiac MRI data. Guidewire paths are defined from 3D vasculature centerlines and deformed synchronously with the anatomy. Dynamic scenes are rendered via conebeam projection, producing 2D fluoroscopic images with pixel-aligned labels. The framework applies physics-based augmentations including simulated detector noise, blurring, path shape, and variable radiopacity to simulate diverse clinical conditions. To evaluate model performance, we trained a YOLOv8-based detector on simulated data and tested on 30 real fluoroscopy sequences from two publicly available datasets. We further demonstrate the practical utility of the framework by implementing tracking via the ByteTrack algorithm and evaluating the fine-grained segmentation capabilities of SAM 2 under various prompting strategies. Across 1567 test frames, YOLOv8 + ByteTrack achieved detection rates of above 93% and SAM 2 achieved a median tip error of 0.37 mm and a mean clDice of 0.90 ± 0.16. Runtime benchmarking confirms that the proposed framework achieves near real-time inference speeds aligned with standard fluoroscopic acquisition rates, supporting its potential use in intraoperative clinical workflows.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.