Objective Photoacoustic computed tomography (PACT) is a fast-evolving noninvasive biomedical imaging technique that shows great potential for basic life sciences and clinical practice. To generate high-quality photoacoustic (PA) images, imaging systems need to employ a dense array of ultrasonic detectors. However, due to economic constraints, fabrication complexity, and real-time data processing requirements, ultrasonic detectors are usually arranged sparsely. Such sparsity cannot satisfy the essential conditions of stable image reconstruction and results in significant artifacts in reconstructed images. To address this issue, we develop an innovative PACT image reconstruction algorithm based on a dual-domain neural network. Methods The proposed network ( Fig. 1), which we refer to as DI-Net, consists of a data-domain network (D-Net) , a back projection layer and an image-domain network (I-Net) . Both D-Net and I-Net are designed based on U-Net, a convolutional neural network that is developed for biomedical image segmentation. Based on U-Net, an instance normalization, a skip connection, and a leaky rectified linear unit are used to enhance the performance of the DI-Net. The back projection layer is a sparse matrix with fixed parameters that allows for gradient propagation from I-Net to D-Net. First, the D-Net maps sparse-view PA data into dense-view PA data in the data domain. Then, the back projection layer transforms the dense-view PA data into a PA image. Finally, the reconstructed image is further enhanced in the image domain by the I-Net. The performance of DI-Net is evaluated through numerical simulations and in vivo experimental data that contains 128-views and 256-views undersampled data. In addition, to further demonstrate the effectiveness of the network, two popular algorithms, i.e. , filtered back projection (FBP) and Post-Unet, are compared with the proposed DI-Net. Results and Discussions We first numerically test the performance of DI-Net using a synthetic vascular phantom dataset. The reconstruction results of 128 views show that the image reconstructed by FBP is significantly contaminated by the streak-type artifacts due to sparse-view sampling. Although both the Post-Unet and the DI-Net can reconstruct artifact-free images, some details are lost in the image reconstructed by the Post-Unet while the DINet completely recovers these details. Quantitative evaluation results demonstrate that the DI-Net provides image quality with the lowest mean square error (MSE), and the highest peak signal-to-noise ratio (PSNR) and structural similarity index measure (SSIM) (Fig. 3, Fig. 5, and Table 1) . When the number of views is 256, streak-type artifacts caused by sparse-view sampling are reduced but are still clearly visible in the images reconstructed by FBP, while Post-Unet and DI-Net effectively suppress these artifacts and reconstruct high-quality images. Due to the increasing number of views, Post-Unet completely reconstructs the details; however the quantitative accuracy of the reconstructed image is lower than that of the DI-Net ( Fig. 4, Fig. 5, and Table 1). We further experimentally evaluated the performance of the DI-Net using an in vivo mouse slice dataset. Similar to the numerical simulations, experimental results also demonstrate the effectiveness of the proposed algorithm. Specifically, in the case of 128 views, the image reconstructed by FBP contains significant artifacts that occlude real PA structures, resulting in the loss of image details. Compared with FBP, Post-Unet demonstrate better performance; however, not all PA structures are recovered. DI-Net can achieve accurate reconstruction and the reconstructed image is consistent with the reference image ( Fig. 6, Fig. 8, and Table 2 ). In the case of 256 views, although the three algorithms reconstruct the PA structures, the image generated by DI-Net more closely resembles the reference image and has the lowest MSE and the highest PSNR and SSIM values (Fig. 7, Fig. 8, and Table 2). Conclusions In this paper, we describe an innovative PACT image reconstruction algorithm based on DI-Net, a dual-domain neural network. Both numerical simulations and in vivo experiments are used to evaluate the performance of the proposed DI-Net. The imaging results reveal that DI-Net can effectively suppress streak-type artifacts caused by undersampling and the reconstructed images are comparable with the reference image. The imaging results also demonstrate that the proposed DI-Net provides better image quality compared with the widelyused FBP algorithm and the popular Post-Unet algorithm.
