Massive multiple-input-multiple-output (mMIMO) enables a significant increase in capacity in fifth-generation (5G) communications systems, both in beamforming and spatial multiplexing scenarios, demanding highly accurate channel estimates. We present two models based on convolutional neural networks (CNNs) for 5G mMIMO channel estimation that differ in complexity and flexibility. The results achieved with both models are competitive compared to traditional methods, such as least squares (LS) which presents a poor estimate in the low signal-to-noise ratio (SNR) region, or minimum mean square error (MMSE) which requires prior statistical knowledge of the channel and noise estimation. Furthermore, the proposed CNN models outperform estimation structures based on conventional deep neural networks (DNNs). Our approach achieves results close to the MMSE estimates, improving them in the low SNR regime, and enabling them to a wide range of channel conditions, i.e., variability in time, frequency, and SNR, not requiring any prior channel statistics information. Furthermore, we present a deep analysis of the computational and cost complexity, demonstrating the suitability of the proposed models for real hardware structure implementation.