Synthesizing Ultrasound B-mode Images from Subsampled RF Data: A Data-Driven Deep Learning Approach

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Signal Processing

Background:

• Ultrasound utilizes radio frequency (RF) signals for B-mode imaging of tissues.
• Manufacturers typically do not disclose RF data processing parameters for B-mode image generation.
• Reconstructing B-mode images from RF data is crucial for research when proprietary details are unavailable.

Purpose of the Study:

• To investigate a convolutional network for converting ultrasound RF data into B-mode images.
• To assess the network's ability to reconstruct images from subsampled RF data, enhancing acquisition speed.
• To evaluate image quality compared to scanner-generated B-mode images.

Main Methods:

• A convolutional network architecture was employed for RF data to B-mode image conversion.
• The network was trained and tested on RF data from Ultrasonix scanners.
• Subsampled RF data (fewer scan lines) were used to evaluate accelerated acquisition potential.

Main Results:

• The convolutional network successfully reconstructed B-mode images from both full and subsampled RF data.
• Reconstructed image quality was comparable to images generated directly by the ultrasound scanner.
• The approach maintained image quality even after multiple nonlinear processing steps.

Conclusions:

• A convolutional network can accurately reconstruct high-quality B-mode ultrasound images from RF data.
• Reconstruction from subsampled RF data improves imaging efficiency by enabling wider fields of view at the same frame rate.
• This method offers a practical solution for B-mode image reconstruction in research settings lacking proprietary scanner data.