This systematic review analyzes 75 papers (2020-2025) applying Deep Learning (DL) techniques to cryptocurrency trading. It evaluates various DL architectures, including Long Short-Term Memory (LSTM), Gated Recurrent Unit (GRU), Convolutional Neural Network (CNN), and Transformers, and finds that DL methods outperform traditional approaches in managing the high volatility and non-linear patterns of crypto markets. Key findings highlight the promise of hybrid and ensemble models, the benefits of integrating blockchain data, sentiment analysis, and macroeconomic factors for improved predictions, and the potential of Deep Reinforcement Learning (DRL) for developing autonomous trading strategies with risk-adjusted returns. However, challenges such as model interpretability, nonstationary data, and real-world deployment persist. The review emphasizes emerging directions like explainable Artificial Intelligence (AI) for transparent decision-making and high-frequency trading applications, providing a critical synthesis of methodologies, empirical results, and research gaps to inform both academic research and practical trading system development.