Deep-learning-assisted and GPU-accelerated vector Doppler imaging with aliasing-resistant velocity estimation.

Area of Science:

• Medical imaging
• Ultrasound technology
• Computational fluid dynamics

Background:

• Vector flow imaging visualizes complex blood flow dynamics using ultrasound.
• High frame rates (over 1000 fps) are achieved using multi-angle vector Doppler and plane wave sensing.
• Doppler aliasing, caused by low pulse repetition frequency (PRF), leads to flow vector estimation errors, hindering practical application.

Purpose of the Study:

• To develop a fast and robust framework for vector Doppler estimation resilient to aliasing artifacts.
• To leverage deep learning and GPU computing for efficient aliasing correction in vector flow imaging.

Main Methods:

• A convolutional neural network (CNN) was employed to detect aliased regions in vector Doppler images.
• An aliasing correction algorithm was applied selectively to detected regions.
• The CNN was trained on 15,000 in vivo vector Doppler frames from femoral and carotid arteries.

Main Results:

• The framework achieved 90% precision in aliasing segmentation.
• Real-time processing throughputs of 25-100 fps were demonstrated.
• Aliasing-free vector flow maps were rendered, significantly improving visualization quality.

Conclusions:

• The proposed deep learning framework offers a computationally efficient solution for Doppler aliasing in vector flow imaging.
• This approach enhances the real-time visualization of complex blood flow dynamics.
• The method holds potential for improving diagnostic accuracy in ultrasound applications.