A communication transmission system with channel coding and deep neural network (DNN)-based decoding is considered. A DNN-based decoding scheme is proposed for reliable transmission. The decoding scheme is accomplished by efficient local decoding at all the neurons and interactions in the input, hidden and output layer. Specifically, firstly, the nonlinear operations at each neuron and the linear operations of the weights and biases at each edge are performed by the local decoding. Secondly, the weights and biases are updated by gradient descent (GD) algorithm, based on the estimated loss value. This process above is performed iteratively until the message sequence has been recovered. Simulation results show that our proposed decoding scheme performs well. Moreover, our decoding scheme performs significantly better than the conventional hard decision.