Tensorflow MNIST示例中的最后一层LSTM

I'm playing around with the Tensorflow LSTM example on the MNIST dataset. What I don't understand is why in the final layer logistic regression is used. Isn't the last output of the LSTM network a better estimator than using the outputs of the previous 'timesteps'? How could I just use the last output of the LSTM network for classification?I'm playing around with the Tensorflow LSTM exa