Compressed Sensing MRI Reconstruction Using Improved U-net based on Deep Generative Adversarial Networks

Abstract

Magnetic Resonance Imaging as non-invasive imaging can produce detailed anatomical images. MRI is a time- consuming imaging technique. Several imaging techniques, like parallel imaging, have been suggested to enhance imaging speed. Compressive Sensing MRI utilizes the sparsity of MR images to reconstruct MR images with under-sampled k-space data. It has already been shown that convolutional neural networks work better than sparsity-based approaches in image quality and reconstruction speed. In this paper, a novel method based on very deep CNN for the reconstruction of MR images is proposed using Generative Adversarial Networks. Generative and discriminative networks are designed with improved ResNet architecture. Using improved architecture has led to deepening generative and discriminative networks, reducing aliasing artifacts, more accurate reconstruction of edges, and better reconstruction of tissues. Compared to DLMRI and DAGAN methods, we demonstrate 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. Reconstruction of brain data with a Cartesian mask of 30% in the proposed method has improved the SSIM criteria up to 0.99. Also, image reconstruction time is approximately 20 ms on GPU, which is suitable for real-time applications.