Fast inter-frame motion correction in contrast-free ultrasound quantitative microvasculature imaging using deep...

Area of Science:

• Medical Imaging
• Ultrasound Technology
• Artificial Intelligence in Medicine

Background:

• Contrast-free ultrasound quantitative microvasculature imaging is valuable for assessing lesions.
• Physiological motion, particularly from carotid artery pulsation, significantly degrades thyroid microvasculature images.
• This degradation leads to diagnostic errors in high-frame-rate ultrasound imaging.

Purpose of the Study:

• To reduce inter-frame motion artifacts in high-frame-rate ultrasound imaging.
• To achieve more accurate visualization of tumor microvessel features in the thyroid.
• To address image degradation caused by carotid artery pulsation.

Main Methods:

• Proposed a low-complexity deep learning network using depth-wise separable convolutional layers.
• Incorporated hybrid adaptive and squeeze-and-excite attention mechanisms for motion correction.
• Validated the network using phantom data with simulated motion and in-vivo thyroid data with physiological motion.

Main Results:

• Achieved average improvements of 35% in Pearson correlation coefficients (PCCs) for simulated motion data.
• Demonstrated PCC improvements from 31% to 35% in microvasculature image reconstruction.
• Showcased average PCC improvement of 20% and mean inter-frame correlation improvement of 40% for in-vivo thyroid data.
• Achieved 5000 times reduction in motion-corrected frame prediction latency compared to conventional methods.

Conclusions:

• The proposed deep learning network effectively corrects inter-frame motion in high-frame-rate ultrasound imaging.
• The method significantly enhances the accuracy and diagnostic quality of microvasculature imaging, particularly in motion-prone organs like the thyroid.
• The network offers a computationally efficient solution suitable for real-time applications.