Abstract: In order to shorten the scanning time of magnetic resonance imaging (MRI), a combination of sub-nyquist sampling data and nonlinear reconstruction algorithms is adopted in the compressed sensing method to realize the real-time or quasi real-time imaging requirement. Considering the sparseness of MRI images in the transform domain and the gradient domain, a conjugate gradient algorithm is proposed to reconstruct the magnetic resonance image using the prediction line search method. The conventional conjugate gradient algorithm suffers a long computation time because of too many attempts in the line search process. We address this problem by optimizing the searching step size with a prediction approach, which significantly reduces the line search cost and accelerates convergence. Simulation results, with under-sampling rate at 10%, 20% and 30%, show that the prediction line search method achieves the best image reconstruction resolution in comparison to the zero filling method and the FR conjugate gradient method, while its time cost is less than the backtracking line search method, which illustrate the effectiveness of the proposed algorithm.