SplineFormer: An Explainable Transformer-Based Approach for Autonomous Endovascular Navigation

TL;DR

This work proposes SplineFormer, a new transformer-based architecture, designed specifically to predict the continuous, smooth shape of the guidewire in an explainable way, and integrates it into an end-to-end robot navigation system by leveraging the condensed information.

Abstract

Endovascular navigation is a crucial aspect of minimally invasive procedures, where precise control of curvilinear instruments like guidewires is critical for successful interventions. A key challenge in this task is accurately predicting the evolving shape of the guidewire as it navigates through the vasculature, which presents complex deformations due to interactions with the vessel walls. Traditional segmentation methods often fail to provide accurate real-time shape predictions, limiting their effectiveness in highly dynamic environments. To address this, we propose SplineFormer, a new transformer-based architecture, designed specifically to predict the continuous, smooth shape of the guidewire in an explainable way. By leveraging the transformer's ability, our network effectively captures the intricate bending and twisting of the guidewire, representing it as a spline for greater accuracy and smoothness. We integrate our SplineFormer into an end-to-end robot navigation system by leveraging the condensed information. The experimental results demonstrate that our SplineFormer is able to perform endovascular navigation autonomously and achieves a 50% success rate when cannulating the brachiocephalic artery on the real robot.

Figures (8)

• Figure 1: System Overview: The experimental setup includes three main components: i) an anatomically accurate half-body vascular phantom model from Elastrat Sarl Ltd.; ii) a joystick controller (robotic leader); and iii) a robotic follower. Data is collected through teleoperation, where the robotic follower is controlled by the joystick controller.
• Figure 2: Segmentation Failure Cases. Due to challenges in capturing thin and elongated guidewire, segmentation models often produce discontinuous segmented maps.
• Figure 3: Robot Setup in X-ray Room: The data acquisition and teleoperation process is highlighted, showcasing the user navigating the guidewire towards the designated target within the vascular phantom. The setup employs a Bi-planar X-ray. To enhance data variability, two different guidewires are used—the Radifocus™ Guide Wire M Stiff Type with an angled tip and the Nitrex Guidewire straight tip.
• Figure 4: Network Architecture: The input fluoroscopic image $X$ is processed by a visual transformer encoder that divides the image into patches, embeds them, and generates visual feature representations $X'$. Simultaneously, a positional encoder processes embeddings of the target sequence. These encoded features are fed into a transformer decoder composed of $N_D$ layers. Using masked self-attention and cross-attention mechanisms, the decoder sequentially generates the B-spline control points $\mathbf{P}_i$ and knots $t_i$ that define the guidewire's geometry. The decoder predicts these parameters by projecting its outputs onto the B-spline dimensionality, yielding pairs $\{\mathbf{P}_0, t_0\}, \{\mathbf{P}_1, t_1\}, \dots, \{\mathbf{P}_n, t_n\}$. An independent tip predictor module initializes the generation by predicting the starting point $\{\mathbf{P}_0,t_0\}$.
• Figure 5: B-spline Representation of a Guidewire: The figure illustrates how a guidewire, as seen in fluoroscopic imaging, can be represented as a continuous B-spline curve.