A deep ensemble network for compressed sensing MRI

Abstract

Compressed sensing theory has been proven to accelerate magnetic resonance imaging by measuring less K-space data called CS-MRI. Conventional sparse-optimization based CS-MRI methods lack enough capacity to encode rich patterns within the MR images and the iterative optimization for sparse recovery is often time-consuming. Although the deep convolutional neural network (CNN) models have achieved the state-of-the-art performance on CS-MRI reconstruction recently, the fine structure details can be degraded due to the information loss when the network goes deep. In order to better transfer the information in lossless way, we design deep ensemble network (DEN) architecture inspired by the novel interpretation of deep neural network in ensemble respect. The DEN model is formed by cascaded basic blocks for CS-MRI. Within the blocks, information flows forward through different depth. The intermediate outputs reconstructed by each block are merged via 3 × 3 convolution to generate the final reconstruction result. The experimental results show the proposed DEN model outperforms other state-of-the-art nondeep and deep CS-MRI models.