Learning needle tip localization from digital subtraction in 2D ultrasound

Area of Science:

• Medical Imaging
• Image Processing
• Machine Learning

Background:

• Ultrasound-guided interventions require accurate needle localization.
• Low intensity of needle shafts and tips in ultrasound images presents a significant challenge.
• Existing methods struggle with subtle intensity changes and motion artifacts.

Purpose of the Study:

• To develop an accurate and efficient method for localizing needles (both in-plane and out-of-plane) in challenging ultrasound-guided interventions.
• To enhance the visibility of low-intensity needle tips and shafts using a novel digital subtraction scheme.
• To integrate a deep learning approach for robust needle tip detection.

Main Methods:

• A digital subtraction scheme was employed to enhance subtle spatiotemporal intensity variations caused by needle tip movement.
• Spatial total variation regularization with the split Bregman method was used to augment the needle tip.
• A deep learning-based end-to-end method was utilized to filter irrelevant motion and detect the needle tip.

Main Results:

• The framework achieved a tip localization error of 0.72 ± 0.04 mm on an extensive ex vivo dataset.
• The processing time was approximately 0.094 seconds per frame, enabling real-time performance (around 10 frames per second).
• The model was trained and validated on diverse phantom data using 17G and 22G needles.

Conclusions:

• The proposed method offers superior speed and accuracy compared to current state-of-the-art techniques.
• The approach demonstrates resilience to spatiotemporal redundancies, making it robust in complex scenarios.
• This technique holds significant potential for improving accuracy in challenging ultrasound-guided interventions.