The type of modulation used by an incoming radio signal is critical for applications such as radio signal surveillance, multiple-input multiple-output communication, cognitive radio, and electronic intelligence. By combining automatic modulation classification with adaptive modulation, communication systems can enhance transmission reliability and data rates by adapting to channel conditions, thus improving spectrum efficiency in cognitive radio systems. This paper evaluates various deep neural network models to develop a robust and accurate classification system capable of operating in environments with significant noise and interference. Utilizing a synthetic dataset of 450,120 data points that include 12 modulation types, both digital and analog, we trained and assessed multiple neural network algorithms. Our results demonstrate that convolutional neural networks achieve the highest performance by leveraging the matrix structure of inputs, reaching an accuracy of more than 99.3% for signal-to-noise ratios equal to or greater than 0 dB. In contrast, other deep and recurrent neural networks were less effective for this classification task.