In order to improve the reconstruction accuracy in fluorescence molecular tomography (FMT), a common approach is to increase the number of fluorescence data or projections. However, this approach consumes too much memory space and computational time. In this Letter, a data compression strategy that involves the removal of the redundant information from both intra- and inter-projections is proposed to reduce the dimension of the FMT inverse problem. The performance of this strategy is tested with phantom and in vivo mouse experiments. The results demonstrate that the proposed data compression strategy can accelerate the FMT reconstruction nearly tenfold and almost without any quality degradation.