Needle tracking in low-resolution ultrasound volumes using deep learning

Area of Science:

• Medical Imaging
• Robotics
• Artificial Intelligence

Background:

• 2D ultrasound-guided needle insertion faces challenges in precise needle and probe alignment, leading to out-of-plane movement.
• 3D ultrasound and deep learning show promise for needle tip detection but require high-resolution images, limiting real-time application.
• Current methods struggle with the trade-off between image resolution, acquisition time, and real-time capability.

Purpose of the Study:

• To develop a deep learning approach for direct 3D needle tip position extraction from sparsely sampled, low-resolution ultrasound (US) volumes.
• To maximize the US volume rate by accepting lower image resolution for improved real-time performance.
• To enhance real-time needle navigation and 3D motion analysis in clinical settings.

Main Methods:

• An experimental setup was designed using a robot for controlled needle insertion into water and chicken liver tissue.
• Needle tip position was assessed using known robot pose, bypassing manual annotation.
• A large dataset of low-resolution volumes was acquired using a 16x16 matrix transducer at a 4 Hz volume rate.
• The deep learning approach was compared against conventional needle segmentation techniques.

Main Results:

• Deep learning significantly outperformed conventional methods in both water and liver tissue experiments.
• Sub-millimeter accuracy was achieved, with mean position errors of 0.54 mm in water and 1.54 mm in liver.
• The proposed method demonstrates effective 3D needle tip localization from low-resolution ultrasound data.

Conclusions:

• Deep learning effectively predicts 3D needle positions from low-resolution ultrasound volumes, offering a viable solution for real-time navigation.
• This approach simplifies needle and ultrasound probe alignment, crucial for reducing out-of-plane errors.
• The study represents a significant advancement for real-time needle navigation and enables comprehensive 3D motion analysis.