In recent years, deep learning has played a pivotal role in the decoding of brain-computer interfaces for motor imagery. In particular, convolutional neural network-based models have become the most popular deep network structures for decoding motor imagery EEG. However, convolutional neural networks focus more on extracting local features and tend to ignore the overall features of EEG signals. To overcome this challenge, we propose a model that fuses the transformer block and multi-dimensional convolution layers for the decoding of motor imagery EEG. In this way, multi-dimensional feature extraction is realized, and the comprehensive decoding of bias and global features by the model is taken into account. In addition, to evaluate the decoding capability of the proposed method, validation tests were carried out on the 2a dataset of the BCI IV competition and compared with the benchmark models. The results show that our proposed method achieves an accuracy of 82.56%, demonstrating a significant improvement in decoding performance compared to baseline models.