Because of the sampling scope and quantity limitation in the computed tomography (CT), the completeness of sparse projection data is very low, which leads to a huge search space for the reconstruction algorithm. The iterative algorithm based on convex optimization can not converge to the global minima in finite time due to the fixed search path. Particle swarm optimization has global search capability, but costs tremendous computation and memory. To improve the quality of reconstruction from incomplete projection data, a new stochastic sparse reconstruction algorithm based on particle swarm optimization is proposed. Firstly, the initial solutions with diversity are generated by the stochastic strategy to ensure the search capability. Secondly, the proposed algorithm stochastically chooses either gradient descent direction or random direction based on the local best known solution and the global best known solution in the iteration, to ensure the efficiency of this algorithm and the diversity of search directions. Finally, to avoid trapping in local optimum, the random initial populations are generated based on the fitness evaluation, which represents the current situation. The contrast reconstruction experiments are conducted on both noise-free and noisy limited-angle sparse projection data. The experimental results demonstrate that the proposed algorithm is efficient and evidently superior in reconstruction quality and robustness compared to common iterative algorithms based on convex optimization or particle swarm optimization.