Convolutional Neural Network for Sparse Reconstruction of MR Images Interposed with Gaussian Noise

Compressive Sensing (CS) reconstructs high-quality images from very few measurements, which are far below Nyquist rate. CS proves to be very useful for acquiring high dimensional image sets like Magnetic Resonance Imaging (MRI). However, the efficiency of MR image reconstruction is affected due to slow acquisition of voluminous k-space data. To improve the quality of reconstructed image and increase the speed of the reconstruction, a novel algorithm namely Adaptive Sparse Reconstruction using Convolution Neural Network AsrCNN has been, proposed for MR Images. AsrCNN employs a CNN, which consists of four convolutional layers and one fully connected layer. The proposed algorithm reconstructs MR images with immense quality, as it is trained over a large dataset with adaptive gradient optimization. The training set consists of 32 x 32 image patches, which is used to create the dictionary by adaptively updating the weights. Subsequently, the dictionary is employed for recovery of sparse MR images corrupted with Gaussian noise. Patch-based approach in AsrCNN enables MR images of varied sizes to be processed without resizing. Experimental results for AsrCNN show an improvement of 1-5 dB in PSNR over previous state-of-art algorithms. Training has been done on GPU using Convolutional Architecture for Fast Feature Embedding (CAFFE) framework as it reduces significant amount of time in reconstructing images.