Abstract
Compression sensing magnetic resonance imaging (CS-MRI) provides a theoretical basis for reducing magnetic resonance (MR) data acquisition time and accelerating the imaging process. In recent years, the CS-MRI algorithm based on deep learning has attracted significant attention and developed rapidly in theoretical research. However, the compression sensing based methods do not capture the detail component of an image very well. Inspired by this, we propose an improved learning iterative shrinkagethresholding algorithm for convolutional sparse coding (CSC) and unfold it with neural networks, named an approximate residual convolutional sparse coding network (ARCSCNet). The new network improves learning ability and efficiency to capture the highfrequency details of the image compared to other algorithms. Firstly, unlike the traditional CSC algorithms, we relax the constraints on a single convolutional dictionary and extract higher-level detail features using multiple iterative layers corresponding to different convolutional dictionaries. Secondly, we add residual structures to the network to preserve the low-frequency MR images of under-sampled k-space data. Thirdly, we introduce the data consistency layer to enhance the fidelity of the k-space data, thereby improving the reconstruction performance of the network. Experimental results show that ARCSC-Net is superior to the state-of-the-art CS-MRI methods in terms of runtime and reconstructed image quality.Video
Citation
@inproceedings{Wang_2022_BMVC,
author = {Qian Wang and Pengyu Li and Jinjia Wang},
title = {ARCSC-Net: An Approximate Residual Convolutional Sparse Coding Network For Compressed Sensing MRI},
booktitle = {33rd British Machine Vision Conference 2022, {BMVC} 2022, London, UK, November 21-24, 2022},
publisher = {{BMVA} Press},
year = {2022},
url = {https://bmvc2022.mpi-inf.mpg.de/0120.pdf}
}
Copyright © 2022
The British Machine Vision Conference is organised by