Deep Learning For Image Reconstruction Via Dense Connected Cascade Network

Image reconstruction is a classical problem in the field of computer vision,which aims to recover a high-resolution image from a degraded low-resolution image.Based on traditional image reconstruction algorithms and different convolutional neural network(CNN)models,this paper explores and studies the application of deep learning network models in the field of magnetic resonance image(MRI)reconstruction:(1)In the CNN sub-network part,we apply some CNN-based image reconstruction algorithms to MRI,and compare the performances of them.We found that the ResNet-DenseNet hybrid network can extract feature information more effectively.(2)In the cascade structure part,as the neural network gradually deepens,the extracted feature information will be distorted.To solve this problem,we add a fidelity term after CNN to ensure that the extracted feature information is used.(3)In the process of improving the traditional cascade network,benefiting from the inspiration of densely connected networks,we densely connect the output of each cascade unit of the cascade network to improve the use efficiency of potential feature information.To sum up,this paper takes improving the information flowability of the network structure as the breakthrough,applies the classical CNN-based algorithms to the MRI reconstruction,and proposes the corresponding dense connected cascade network.Which achieves a superior performance than other methods in our research.