Deep-Learning Based Adaptive Ultrasound Imaging From Sub-Nyquist Channel Data.

Area of Science:

• Medical Imaging
• Signal Processing
• Machine Learning

Background:

• Traditional ultrasound imaging requires high sampling rates, leading to significant data storage and processing challenges.
• These challenges impact ultrasound hardware development, algorithm performance, and overall imaging efficiency.
• Deep learning shows promise in overcoming limitations in medical image reconstruction.

Purpose of the Study:

• To develop a deep learning-based method for reconstructing high-quality B-mode ultrasound images from sub-sampled channel data.
• To address hardware and software limitations by reducing data acquisition requirements.
• To improve image resolution and contrast-to-noise ratio compared to existing methods.

Main Methods:

• Proposed a deep-learning reconstruction approach using temporally and spatially sub-sampled data.
• Trained an encoder-decoder convolutional neural network (CNN) on partial data.
• Used minimum-variance (MV) beamformed signals from fully-sampled data as training targets.

Main Results:

• Achieved high-quality B-mode image reconstruction from sub-sampled ultrasound data.
• Demonstrated up to two times higher resolution compared to NESTA and delay-and-sum (DAS) beamforming.
• Obtained contrast-to-noise ratio (CNR) comparable to MV beamforming and up to 2 dB higher than DAS and NESTA.

Conclusions:

• Deep learning enables efficient and high-quality ultrasound image reconstruction from reduced data.
• The proposed method offers superior resolution and CNR, surpassing current clinical practices.
• This approach has the potential to advance ultrasound technology, enabling better and more efficient medical imaging.