Magnetic resonance imaging (MRI) as a non-invasive imaging is able to produce three-dimensional detailed anatomical images without the use of damaging radiation and with excellent visualization of the anatomical structure. MRI is a time-consuming imaging technique. Image quality may be reduced due to spontaneous or non-spontaneous movements of the patient. Several imaging techniques like parallel imaging have been suggested to enhance imaging speed. Compressive Sensing MRI (CS-MRI) violates the Nyquist-Shannon sampling rate and utilizes the sparsity of MR images to reconstruct MR images with under-sampled k-space data. Prior studies in CS-MRI have employed orthogonal transforms such as wavelets and recent methods have employed dictionary learning for adaptive transforms. CS-MRI is an efficient way to decrease the scan time of MR imaging. However, the computational costs are usually expensive and the CS reconstruction process is time-consuming. In addition, CS-MRI methods are based on constant transform bases or shallow dictionaries, which limits modeling capacity. Deep Learning is a novel orientation in Machine Learning and Artificial Intelligence investigation. It has already been shown that C work better than sparsity-based approaches in terms of both image quality and reconstruction speed. In this thesis, a novel method based on very deep convolutional neural networks (C) for reconstruction of MR images is proposed using Generative Adversarial Networks (GANs). In this model, a Generative and Discriminator networks designed with improved ResNet architecture. The Generative network is based on U-net and Discriminator is a classifier network, which improved blocks is used in both of them. Using improved architecture has led to deepening of Generative and discriminator networks, reduction in aliasing artifacts, more accurate reconstruction of edges and better reconstruction of tissues. To achieve better reconstruction adversarial loss, pixel-wise cost and perceptual loss (pre-trained deep VGG network) are combined. Two comparisons have been made with the latest studies in this field. The first comparison has been made with DLMRI which is a well-known conventional method of CS-MRI to reconstruct magnetic resonance images. The second comparison has been made with deep learning method named DAGAN. Compared to DLMRI and DAGAN methods, it has been demonstrated that the proposed method outperforms the conventional methods and deep learning based approaches. Assessment is made on several datasets such as the brain, heart, and prostate and the proposed method leads to a better reconstruction in details of the images. Reconstruction of brain data with a radial mask of %30 in the proposed method has been improved on basis of the SSIM criteria up to 0.98. Also, image reconstruction time is about 20 ms on GPU that is much smaller than the state-of-the-art CS-MRI methods. Keywords Deep Learning, Generative Adversarial Networks (GANs), MRI, Compressive Sensing