In this paper, an efficient approximated expectation propagation assisted decentralized signal detection scheme is proposed for uplink multiple-input multiple-output–orthogonal time-frequency space (MIMO-OTFS) systems. Concretely, a decentralized baseband processing (DBP) architecture is first proposed for uplink MIMO-OTFS, which is capable of offloading the signal processing operation of the central processing module to multiple local signal processing units. Next, a decentralized approximated expectation propagation (DeAEP) algorithm is conceived for the DBP architecture, where the means and variances of the local Gaussian probability density function (PDF) are updated at the local signal processing unit and integrated at the central processing module by utilizing maximum ratio combining (MRC). Simulation results show that the BER performance of the proposed DeAEP is comparable to the conventional minimum mean square error (MMSE) detector while its computational complexity can be significantly reduced.