In recent years, the continued advances of the deep learning as a part of machine learning produces an accuracy which resembles the people’s performance in processing various challenges of the real world. U-Net, as convolutional neural network (CNN), is one of the deep learning architectures that have been utilized to perform segmentation in several applications. The flexible design of the U-Net, utilizing the data augmentation approach, has been contributed in the achievement of successful predictive results for different image sizes particularly with training few datasets implementing efficient computations. However, the accuracy of one application may need adding additional improvement on the basic U-Net, due to the encoding and decoding processes, which causes some information loss. Another challenge is that the training and testing of a large amount of labeled data is a very computation-intensive process which needs to be minimized. Therefore, this review aims to describe the basic building blocks of 2D U-Net architecture, addressing its challenges and then it explains the most important cooperation issue between software and hardware. Finally it introduces important conclusions with considerable remarks that may help in selecting a suitable model.