Accurate measurement of the transverse phase space distribution of electron beams is of great importance in the design and optimization of accelerators. The computerized tomography theoretically provides the true transverse phase space distribution. However, to understand the details of the distribution more accurately, it is necessary to solve the problems of limited range of rotation angle and insufficient number of projections. In this paper, a neural network model is proposed to address these two problems in the hybrid domains, which combines interpolation and artifact removal neural networks in the sinogram and tomogram domains, respectively. Even with a simple diagnostic beamline and a small number of projections (7), the network model can reconstruct the transverse phase space distribution of beams with high quality. Moreover, since the selection of angles is independent of the normalized phase space, Twiss parameters do not need to be measured. Using the proposed method to measure the transverse phase space improves reconstruction quality to a certain extent and simplifies the measurement process.