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

Ultrasound, a widely used and safe imaging modality, utilizes radio frequency (RF) signals obtained from pulse-echo imaging to generate B-mode images, offering a visual representation of internal body organs and tissue microstructures. The processing details and the specific parameters applied for converting RF data to B-mode images in ultrasound devices are typically not made available by manufacturers. In this study, we investigated a convolutional network architecture for conversion of RF data into two different types of B-mode images generated by the Ultrasonix scanners. Additionally, we assessed the network's efficacy in translating subsampled RF data with fewer scan lines into the B-mode images, aiming to expedite the data acquisition and transmission process. Our results on an unseen test set demonstrate the feasibility of reconstructing B-mode images from full and subsampled RF frames with a quality similar to the scanner-generated B-mode images, even when these images undergo multiple nonlinear processing steps. The proposed approach for B-mode image reconstruction from subsampled RF frames permits scanning wider lateral field of view with the same frame rate, offering increased imaging efficiency while maintaining comparable image quality. In addition, it provides a practical solution for reconstructing high-quality B-mode images from engineered RF data in research settings, where scanner processing details are unavailable.