Assistant Professor Virginia Polytechnic Institute and State University Blacksburg, Virginia, United States
Introduction:: Ultrasound is a widely available, safe, and portable medical imaging modality, but its usefulness in differential diagnosis (e.g., differentiating between benign and malignant tumors) is limited by unspecific contrast for certain disease types. Novel contrast mechanisms can be achieved by imaging biomechanical properties that change with disease processes, with sound speed being a promising candidate. However, sound speed imaging using pulse-echo ultrasound is challenging due to image quality and reconstruction speed issues with physics-based methods and model generalizability issues with end-to-end deep learning methods. This study proposes a real-time, generalizable deep learning method for sound speed imaging using pulse-echo ultrasound and evaluates the method’s feasibility through full-wave simulations.
Materials and Methods:: The proposed method utilizes a deep-learning model that converts ultrasound signals into sound speed images. Unlike end-to-end deep learning approaches that use system-dependent raw channel signals as inputs, we convert the raw signals into a system-independent form through physical modeling to enable generalizability. Full-wave acoustic simulations using the k-Wave toolbox were performed for multiple linear ultrasound array systems under various settings to test the feasibility of our method. Sound speed imaging was performed using the proposed method, an existing physics-based method, and an end-to-end deep learning method, and image quality and model generalizability were compared.
Results, Conclusions, and Discussions:: Our proposed method yielded more accurate sound speed imaging results than the physics-based method, with significantly faster reconstruction times allowing for real-time imaging. Compared with the end-to-end deep learning method, our method was more generalizable across systems and settings. These simulation results demonstrate the feasibility of our approach to achieve real-time, generalizable deep-learning-based sound speed imaging. The method could potentially lead to a practical new imaging modality to significantly enhance the clinical usefulness of diagnostic ultrasound.