Two-phase flow is widely encountered in industrial production, where Ultrasonic Tomography (UT) has the advantages of non-invasion, non-radiation and low-cost. The traditional ultrasonic image reconstruction algorithm produces serious artifacts, which hinders its applications. A pre-imaging Convolutional Neural Network (CNN) base on U-net method is proposed to improve the accuracy of the image reconstruction of UT. The raw measurement data pre-processed by SART iteration as the input of the network to extract the features of the reconstructed image more effectively. In order to adapt to the low resolution characteristic of UT reconstructed image, the network structure includes 1 down-sampled layer and 5 up-sampled layers. A simulation data set containing 23,355 samples of ultrasonic measurements was established for the model training. After 433 training epochs, the Image Correlation Coefficient (ICC) on validation set was 0.86. When different levels of white Gaussian noise were added to validation set, the ICC remains the same, which proved the anti-noise capability of the network. The results show that compared with Local Binary Pattern(LBP), Simultaneous Algebraic Reconstruction Technique (SART), the proposed CNN method has a better reconstruction effect.