GPU-Based Acceleration for Real-Time Speed-of-Sound Imaging

Speed-of-sound (SoS) imaging provides critical information for tissue characterization and phase aberration correction. Despite its potential, real-time reconstruction remains challenging for physics-based pulse-echo SoS imaging, which relies on explicit physical modeling rather than data-driven machine learning approaches. To address this challenge, we developed a GPU-accelerated framework in which the computationally intensive steps in physics-based SoS imaging, including phase shift computation and SoS inversion, were reformulated as tensor operations. Specifically, phase shift computation was reformulated as a single-pass, multi-channel tensor operation, and SoS inversion was reformulated as a precomputed linear projection. Both steps were executed on the GPU. Implemented in PyTorch with cuBLAS and cuDNN backends, the method processed each frame within milliseconds on an NVIDIA A40 GPU. Scalability evaluations further demonstrated that the proposed approach maintained approximately constant runtimes when the number of input phase-shift maps or the kernel size was increased.