Image Reconstruction for Ultrasound Computed Tomography by Use of the Regularized Dual Averaging Method

Waveform inversion methods can produce high-resolution reconstructed sound speed images for ultrasound computed tomography; however, they are very computational expensive. Source encoding methods can reduce this computational cost by formulating the image reconstruction problem as a stochastic optimization problem. Here, we solve this optimization problem by the regularized dual averaging method instead of the more commonly used stochastic gradient descent. This new optimization method allows use of non-smooth regularization functions and treats the stochastic data fidelity term in the objective function separately from the deterministic regularization function. This allows noise to be mitigated more effectively. The method further exhibits lower variance in the estimated sound speed distributions across iterations when line search methods are employed.