Graph neural network (GNN) has been considered for massive multiple-inputmultiple-output (MIMO) detection recently due to its strong capability in solving graph-based classification problems. However, current GNN detectors still struggle between performance and complexity under high order modulations. In this paper, we propose a novel GNN-based MIMO detector named GNN-BP, which combines GNN and belief propagation (BP) messages. The GNN is constructed based on the Markov random field (MRF) MIMO model, while extra node features are introduced from BP as a supplement to effectively enhance the detection performance with acceptable extra cost. Simulation results show that the proposed GNN-BP detector can outperform state-of-the-art (SOA) GNN-based detectors, and achieves similar performance as GNN with expectation propagation (GEP) with much lower computational cost.