The Transformer architecture has shown its superior ability than recurrent neural networks on many different natural language processing applications. Therefore, this study applies a modified Transformer on the speech enhancement task. Specifically, the positional encoding may not be necessary and hence is replaced by convolutional layers. To further improve PESQ scores of enhanced speech, the L_1 pre-trained Transformer is fine-tuned by MetricGAN framework. The proposed MetricGAN can be treated as a general post-processing module to further boost interested objective scores. The experiments are conducted using the data sets provided by the organizer of the Deep Noise Suppression (DNS) challenge. Experimental results demonstrate that the proposed system outperforms the challenge baseline in both subjective and objective evaluation with a large margin.