High spatial resolution attenuation imaging on a breast tomography scanner based on deep learning

Attenuation imaging has been used as an independent indictor for breast cancer in ultrasound imaging. Both quantitative ultrasound (QUS) and ultrasound tomography (USCT) can be used to create attenuation images. In QUS, the spectral log difference method is most often applied on the backscattered signal, generating a low spatial resolution image. In USCT, the full wave inversion method is used to reconstruct the imaginary part of the wavenumber, which is attenuation, but suffers from being poorly conditioned, leading to a lower image quality. To overcome these issues, we propose using deep learning (DL) to reconstruct attenuation images of the breast using an ultrasound tomography scanner, i.e., QTI Breast Acoustic CT Scanner. Our U-Net neural network uses both 60-angle RF data as the input and, considering the Kramers–Kronig relations, the sound speed image as an additional input. The attenuation image is the output. The network was trained on simulated breast phantoms, and tested on simulation, physical phantoms, and in vivo breast data. The results demonstrate that our method can generate a high spatial resolution attenuation image with accurate values, and the relationship between generated attenuation and sound speed can serve as a new indicator for breast cancer.