Accelerated Super-resolution MR Image Reconstruction via a 3D Densely Connected Deep Convolutional Neural Network

Magnetic resonance (MR) images are often acquired with low resolution (LR) due to hardware limitations, practical limitations on acquisition time, and patient comfort. In this study, we propose a novel method to reconstruct high resolution (HR) MR images through efficiently learning the LR to HR nonlinear mapping by a densely connected 3D deep convolutional neural network (CNN) that uses learnable deconvolution on multi-level features for upsampling. Different from the current CNN-based MR image super-resolution (SR) methods that take interpolated patches as input, the proposed method directly takes LR MR images (or LR patches) as input to reduce computational complexity and accelerate SR reconstruction. Improved dense blocks in this architecture are adopted to extract multi-level features from the LR image, and carry the information forward with dense connections. The final deconvolution layer automatically learns a filter to fuse and upscale all feature maps to generate HR MR images. The experimental results on three benchmark datasets demonstrate that the proposed method achieves state-of-the-art MR image SR reconstruction performance with less computational load and memory usage.