We propose to use symbol decision based on neural networks to alleviate the impacts of quantization noise, laser phase noise, IQ impairments, and their interactions. Our DSP-assisted transceiver design allows the use of low-cost transceiver hardware while satisfying the performance requirements. Extensive computer simulations show the effectiveness of neural networks when higher-order QAM signals are contaminated by the above impairments.