Phase aberration in ultrasound imaging, often caused by heterogeneous sound speed distribution in tissue, introduces errors in estimating relative time delays during conventional delay-and-sum beamforming, resulting in degraded image quality. Differentiable beamforming in ultrasound autofocusing (DBUA) has recently emerged as a promising method to adaptively estimate spatial sound speed distributions in tissue by minimizing a loss function based on focusing criterion, such as the common midpoint phase error. However, implementing DBUA for large aperture arrays with extended fields of view remains computationally demanding due to increased element numbers and longer acquisition samples, even with graphical processing units (GPUs) for parallel computing. This study investigates refinement techniques for DBUA using a 384-element extended aperture phased array (88 mm aperture, 0.23 mm interelement pitch (0.37λ), 2.5 MHz operating frequency), processed on an Nvidia Titan RTX GPU. Specifically, we optimize control parameters, including the f-number, learning rate, and the size and distribution of the patches (i.e., selective regions for focusing criterion calculations). The results demonstrate over a 30% improvement in computational efficiency while achieving comparable imaging quality, with 1% accuracy in sound speed estimation in phantom studies, advancing DBUA toward practical usage in abdominal imaging. Work supported by NIH R01CA258807, R01EB033967, and R01CA271309.
Song et al. (Tue,) studied this question.