VeNet: a lightweight neural network for efficient brain vessel segmentation in endovascular robotic surgery.

Telesurgery integrates artificial intelligence (AI), robotic systems, sensing technologies, and wireless communication to enable remote and computer-assisted surgical interventions. Beyond remote operation, AI-driven assistance can improve procedural stability and consistency by supporting motion smoothing, tremor reduction, event detection, and workflow guidance. Achieving these capabilities requires accurate, patient-specific anatomical models derived from limited medical imaging data, together with data-efficient and computationally lightweight algorithms suitable for near-real-time use. Robotic endovascular surgery presents additional challenges due to the complex topology of the cerebral vasculature and the limited availability of high-quality annotated imaging datasets. Current vessel segmentation pipelines often rely on manual annotation or classical vesselness filters, while deep learning approaches typically require large annotated datasets and substantial computational resources, limiting their applicability in resource-constrained or real-time settings. Here, we present VeNet, a lightweight neural network designed for three-dimensional segmentation of tubular structures such as blood vessels. VeNet employs a matrix-based operator and contains approximately 6,000 trainable parameters, enabling efficient training and inference on standard CPU hardware. The model demonstrates robust segmentation performance in low-data regimes, making it suitable for scenarios where extensive manual annotation is impractical. Using VeNet, we generated a large, semi-automatically annotated brain vessel magnetic resonance angiography dataset derived from the IXI cohort, with subsequent expert review. Together, VeNet and the accompanying dataset provide a resource-efficient foundation for vascular segmentation and simulation workflows. The model has been integrated into a robotic endovascular research platform with haptic feedback to support patient-specific vascular reconstruction and interactive simulation.