Self-supervised learning for interventional image analytics: toward robust device trackers.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Cardiovascular Interventions

Background:

• Accurate device tracking (e.g., guiding catheters) is crucial for endovascular cardiac interventions.
• Challenges include device obscuration, changing acquisition angles, and patient motion, impacting procedural safety and efficacy.

Purpose of the Study:

• To develop a robust and efficient method for tracking medical devices in live X-ray image sequences.
• To overcome limitations of existing tracking methods in challenging interventional scenarios.

Main Methods:

• A self-supervised learning approach using masked image modeling on over 16 million interventional X-ray frames.
• Utilized frame interpolation-based reconstruction to learn inter-frame temporal correspondences.
• Fine-tuned features in a lightweight model for downstream tasks.

Main Results:

• Achieved state-of-the-art performance, notably in robustness, outperforming optimized reference solutions.
• Reduced maximum tracking error by 66.31% and standard deviation of errors by 20%.
• Demonstrated a 97.95% success score at 42 frames-per-second inference speed on GPU.

Conclusions:

• The proposed data-driven approach offers superior robustness and speed for device tracking compared to multi-modular methods.
• Results support its application in interventional image analytics requiring spatio-temporal understanding.