A block-based compressed sensing scheme for magnetic resonance (MR) i mage with variable sampling rate is proposed.In view of containing the rich details and c omplex texture of MR image,the Tetrolet transform,which can represent the high dimensional singul arity structure of image sparsely,is introduced.Meanwhile,the adjacent slices of the MR image are bound to constitute a group of pictures (GOP) considering the spatio-temporal correlatio n between contiguous slices,and the disparities of the reference slice with its immediate previous and following slices are calculated to form the difference images.Then,the referenc e slice and the difference images are partitioned into the no-overlapped blocks to assign sampl ing rate according to the changes of contents between sequential image blocks.Finally,GOP can be reconstructed from measurements by using the projected Landweber algorithm under the compresse d sensing framework.The experimental results show that the Tetrolet transform is a suita ble sparse representation tool for MR image.Compared with the methods using discrete cosin e transform and dual-tree wavelet transform,the PSNR of the reconstruct ed images is increased by 0.92 dB and 2.06dB,averagely.Moreover,for diverse GOPs,the qualities of the reconstructed i mages with variable sampling rate are all better than the results obtained from the fixed sampling rate method.This paper provides a feasible scheme for compressed sensing of MR images .