A learning-based method for compensating 3D-2D model mismatch in ring-array ultrasound computed tomography

Ultrasound computed tomography (USCT) is an emerging imaging technique that holds great promise for breast cancer diagnosis and screening. Full-waveform inversion (FWI)-based image reconstruction methods can produce high spatial resolution images that depict the acoustic properties of soft tissues. However, FWI is computationally demanding, especially when wave physics is modeled in three dimensions (3D). A common USCT design employs a circular ring-array comprised of elevation-focused ultrasonic transducers, where the array is translated orthogonally to achieve volumetric (3D) imaging. This design allows efficient two-dimensional (2D) slice-by-slice reconstruction methods to estimate a 3D volume by stacking reconstructed cross-sectional images at each ring-array position. However, these 2D methods do not account for the 3D wave propagation physics and the focusing properties of the transducers and thus can result in out-of-plane scattering-based artifacts and inaccuracies. Previous work has investigated a learning-based method, in which a deep neural network was used to map the 3D ring-array USCT data to idealized 2D USCT measurements, from which acoustic properties can be estimated by use of 2D FWI. The presented work extends that previous study in two ways. First, sophisticated anatomically realistic 3D numerical breast phantoms are used to construct clinically relevant training and testing sets. Second, a high-fidelity 3D wave propagation forward imaging model incorporating elevation focusing effects is used to virtually image the phantoms. The results show promise in improving image accuracy compared to both 3D FWI and conventional 2D FWI, achieving both mitigation of out-of-plane scattering and 3D-2D model mismatches and significant reduction of computational cost compared to 3D FWI. Additionally, the use of multiple ring-array measurements from adjacent elevations are explored as concurrent neural network inputs, showing improved accuracy compared to only using single-ring data.