Computationally efficient deep learning-based image reconstruction for sound speed estimation in ultrasound computed tomography

Abstract

Ultrasound computed tomography (USCT) has great potential for breast cancer imaging due to its ability to quantify acoustic tissue properties such as the speed of sound (SOS). Full-waveform inversion (FWI)-based image reconstruction methods can produce high spatial resolution SOS images, but these methods are computationally demanding, especially when formulated in the time-domain. Deep learning (DL)-based methods can be computationally efficient and provide high-quality object estimates as studied in various medical image reconstruction problems. However, there remains a further need to develop and assess the DL-based image reconstruction methods for USCT for use in medical imaging applications. Here, we present a supervised DL approach for USCT image reconstruction that is formulated as an image-to-image translation problem. Specifically, a low-resolution SOS image obtained via traveltime tomography and a high-resolution reflection tomography image are used as inputs to a U-Net, which is trained to estimate a high-resolution SOS map. These input images are complementary, providing both quantitative information about SOS values and information about tissue boundaries. Numerical studies that employ realistic numerical breast phantoms demonstrated that the proposed method produced high-quality SOS estimates based on image quality metrics such as NMSE, SSIM, and PSNR.