We develop a sparse tensor constrained reconstruction (STCR) algorithm which utilizes the nonlocal similarity prior information and divides the computed tomography (CT) image into a series of patch groups. The multidimensional low-rank decomposition method for tensors is used, and the prior information is introduced in the low dose computed tomography (LDCT) reconstruction to establish an object function. The object function is optimized by alternating iteration of the CT reconstruction image update step and the patch group sparse coding step in the iterative process. The performance of the STCR algorithm is verified through experiments based on simulation data and clinical data. Preliminary experimental results show that, compared to the classical reconstruction methods, the proposed method can produce better images in terms of structure preservation and noise suppression.