GRU-based Modeling for Predicting Guidewire Trajectories in Interventional Robotics With Morphological Feature Fusion

We present a causality-preserving sequential estimator for guidewire trajectory reconstruction during interventional navigation. Unlike generic recurrent baselines, the proposed model time-broadcasts sequence-level constants (including guidewire stiffness, insertion angle, and an effective friction descriptor) and concatenates them with dynamic geometric tokens (centerline coordinates and local diameter) before a two-stage feature encoder and a unidirectional gated recurrent unit decoder that emits 2D positions stepwise. To cope with variable sequence lengths, we adopt a time-step length classification training strategy with mask-based loss function, which limits padding-induced invalid gradients and improves training and inference efficiency without altering the network architecture. On a phantom platform covering multiple guidewire types and insertion angles, the method achieves a 0.40 ~ 0.54 mm position-error range (mean 0.46 mm) while preserving strict causality; relative to a baseline without the time-step classification strategy, it reduces epochs-to-convergence by 42%, training time by 52%, and per-inference latency by 51%. These results indicate a deployable, real-time basis for guidewire trajectory estimation and intraoperative navigation.