Catheter segmentation in X-ray fluoroscopy using synthetic data and transfer learning with light U-nets

Area of Science:

• Medical Imaging
• Computer Vision
• Surgical Technology

Background:

• Automated segmentation and tracking of surgical instruments, specifically catheters and guidewires, in X-ray fluoroscopy are crucial for enhancing image guidance in endovascular procedures.
• Current methods often require extensive manual data annotation, posing a significant bottleneck in the development of effective segmentation models.

Purpose of the Study:

• To present a novel method for real-time segmentation of catheters and guidewires in 2D X-ray images using Convolutional Neural Networks (CNNs).
• To develop a lightweight U-Net architecture through a transfer learning approach, utilizing synthetic fluoroscopic images to minimize the need for manually annotated data.

Main Methods:

• A transfer learning strategy was employed, generating high-fidelity synthetic fluoroscopic images from real fluoroscopic backgrounds.
• A two-stage process involving initial end-to-end training and subsequent fine-tuning was implemented. A small set of manually annotated in-vivo images was used to fine-tune the deepest layers of the U-Net model.
• The network processes grayscale images and outputs a binary mask for pixel-wise catheter/guidewire segmentation.

Main Results:

• The developed U-Net models achieved accurate catheter/guidewire segmentation, with average Dice coefficients around 0.55, 0.26, and 0.17 across different evaluations.
• The proposed U-Net model demonstrated comparable segmentation accuracy to state-of-the-art CNN models, with an average Dice coefficient difference within ±5%.
• A significant 84% reduction in testing time was achieved compared to existing methods, enhancing flexibility for real-time applications and adaptability to higher input resolutions.

Conclusions:

• The study presents an effective approach for developing CNN models for surgical catheter segmentation in X-ray fluoroscopy by leveraging synthetic images and transfer learning.
• This methodology substantially reduces the dependency on large volumes of manually annotated data, addressing a key limitation in CNN model development.
• The simplified U-Net model, combined with the proposed training strategy, offers significant advantages over current state-of-the-art solutions for real-time image guidance in endovascular procedures.