Three-dimensional particle reconstruction with limited two-dimensional projects is an underdetermined inverse problem that the exact solution is often difficulty to be obtained. In general, approximate solutions can be obtained by optimization methods. In the current work, a practical particle reconstruction method based on convolutional neural network (CNN) is proposed. The proposed technique can refine the particle reconstruction from a very coarse initial guess of particle distribution from any traditional algebraic reconstruction technique (ART) based methods. Compared with available ART-based algorithms, the novel technique makes significant improvements in terms of reconstruction quality and at least an order of magnitude faster with dense particle concentration.