Abstract: ‍ ‍Maximum likelihood detection （ML） is the optimal algorithm for traditional Multi-input Multi-output （MIMO） signal detection. However， it is constrained by the number of antennas， the ratio of transmitting and receiving antennas， and the modulation scheme， making it only appropriate for scenarios where the number of antennas is relatively small， the ratio of transmitting and receiving antennas is low， and the signal is of low modulation order. As a novel solution， deep learning （DL）-based signal detection algorithms have received much attention but also suffer from deterioration in detection performance when the transmitting and receiving antenna sizes are mutually similar. This paper combines the orthogonal approximate message passing （OAMP） algorithm with a sparsely connected neural network （ScNet） into a trainable network structure and proposes a new signal detection algorithm for the uplink of MIMO systems， called ScNet-OAMP. The algorithm provides an accurate initial solution of the signal transmission parameters through a neural network， which improves the linear and non-linear estimation of the OAMP process， therefore enhancing its noise reduction capability and achieving an improved detection accuracy， with the best detection performance compared to ScNet and OAMP with the same experimental parameters. The experimental results show that the algorithm is suitable for different modulation signals such as QPSK， 4QAM， and 16QAM and can handle system configurations with different ratios and numbers of transceiver antennas， especially when the number of transceiver antennas is nearly the same， with a performance gain of at least 0.5 dB， and even 2.2 dB or more， in the 10^-3 bit error rate.