Lightweight wavelet convolutional network for guidewire segmentation.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Accurate guidewire segmentation is critical for vascular interventions.
• Current methods struggle with complex models and limited dual-guidewire data.
• There's a need for efficient, robust segmentation models for clinical use.

Purpose of the Study:

• To develop a lightweight, efficient, and robust method for segmenting single and dual guidewires in X-ray fluoroscopy.
• To overcome challenges of data scarcity and model complexity in guidewire segmentation.
• To enable real-time clinical deployment of guidewire segmentation technology.

Main Methods:

• Proposed a lightweight wavelet convolutional network (WT-CMUNeXt) integrating wavelet convolution and channel attention.
• Developed a dual guidewire data augmentation algorithm to synthesize data from single guidewire images.
• Evaluated the model on multiple patient X-ray fluoroscopy sequences.

Main Results:

• WT-CMUNeXt achieved state-of-the-art single guidewire segmentation (F1: 0.9048, IoU: 0.8284).
• Demonstrated strong dual guidewire segmentation performance (F1: 0.8668), outperforming most methods.
• The model is lightweight (3.26M parameters) with low computational cost (2.99 GFLOPs).

Conclusions:

• WT-CMUNeXt offers an efficient and accurate solution for both single and dual guidewire segmentation.
• The proposed data augmentation effectively addresses data scarcity issues.
• The model's efficiency and accuracy make it suitable for real-time clinical applications.