Objective. The image reconstruction of ultrasound computed tomography is computationally expensive with conventional iterative methods. The fully learned direct deep learning reconstruction is promising to speed up image reconstruction significantly. However, for direct reconstruction from measurement data, due to the lack of real labeled data, the neural network is usually trained on a simulation dataset and shows poor performance on real data because of the simulation-to-real gap.Approach. To improve the simulation-to-real generalization of neural networks, a series of strategies are developed including a Fourier-transform-integrated neural network, measurement-domain data augmentation methods, and a self-supervised-learning-based patch-wise preprocessing neural network. Our strategies are evaluated on both the simulation dataset and real measurement datasets from two different prototype machines.Main results. The experimental results show that our deep learning methods help to improve the neural networks' robustness against noise and the generalizability to real measurement data.Significance. Our methods prove that it is possible for neural networks to achieve superior performance to traditional iterative reconstruction algorithms in imaging quality and allow for real-time 2D-image reconstruction. This study helps pave the path for the application of deep learning methods to practical ultrasound tomography image reconstruction based on simulation datasets.@en

Deep neural networks have proven to excel classical medical image reconstruction techniques. Some networks are based on fully connected (FC) layers to achieve domain transformation such as from the data acquisition domain to the image domain. However, FC layers result in huge numbers of parameters which take a lot of GPU memory. Hence, they do not scale well, and the overall performance is limited. For ultrasound computed tomography (USCT) application, we propose a memory-efficient convolutional network that reconstructs images from the frequency domain to image domain with much less parameters compared with multilayer perceptron, by using data-driven learning. Extensive experiments demonstrate that our method achieves high reconstruction quality. It improves the structural similarity measure (SSIM) from 0.73 to 0.99 when compared with state-of-the-art reconstruction methods in this field while reduces 2/3 parameters when compared with deep neural network with FC layers to reconstruct images from frequency domain to image domain. @en