VibNet: Vibration-Boosted Needle Detection in Ultrasound Images

Area of Science:

• Medical Imaging
• Biomedical Engineering
• Machine Learning

Background:

• Ultrasound (US)-guided interventions require precise percutaneous needle detection.
• Speckles, artifacts, and low resolution in US images hinder robust needle detection, particularly when visibility is compromised.

Purpose of the Study:

• To develop a robust and accurate needle detection framework for ultrasound images.
• To leverage external periodic vibration of needles to enhance detection capabilities.

Main Methods:

• Proposed VibNet, a learning-based framework integrating neural Short-Time Fourier Transform and Hough Transform modules.
• Extracted motion features in spatiotemporal space and aggregated frequency features for detection in Hough space.
• Utilized subtle periodic vibrations for robust feature extraction in the frequency domain.

Main Results:

• VibNet demonstrated superior performance in detecting severely reduced visibility needles on ex vivo porcine samples.
• Achieved significantly lower tip error (${1.61}\pm {1.56}$ mm) compared to UNet (${8.15}\pm {9.98}$ mm) and WNet (${6.63}\pm {7.58}$ mm).
• Showcased reduced needle direction error (${1.64}\pm {1.86}^{\circ}$) versus UNet (${9.29}~\pm ~{15.30}^{\circ}$) and WNet (${8.54}~\pm ~{17.92}^{\circ}$).

Conclusions:

• VibNet offers a more effective approach to needle detection in ultrasound imaging compared to intensity-based methods.
• The frequency domain features derived from subtle vibrations provide enhanced robustness.
• The framework shows promise for improving the safety and efficacy of ultrasound-guided interventions.