A Cascaded Residual Dense Network for Cardiac MR Imaging
Ziwen Ke1,2, Shanshan Wang2, Cheng Li2, Huitao Cheng1,2, Leslie Ying3, Xin Liu2, Hairong Zheng2, and Dong Liang1,2

1Research center for Medical AI, Shenzhen Institutes of Advanced Technology, Shenzhen, China, 2Paul C. Lauterbur Research Center for Biomedical Imaging, Shenzhen Institutes of Advanced Technology, Shenzhen, China, 3Department of Biomedical Engineering and Department of Electrical Engineering, The State University of New York, Buffalo, NY, United States


Cardiac MR imaging plays an important role in clinical diagnosis. But the long scan time limits its wide applications. To accelerate data acquisition, deep learning based methods have been applied to effectively reconstruct the undersampled images. However, current deep convolutional neural network (CNN) based methods do not make full use of the hierarchical features from different convolutional layers, which impedes their performances. In this work, we propose a cascaded residual dense network (C-RDN) for dynamic MR image reconstruction with both local features and global features being fully explored. Our proposed C-RDN achieves the best performance on in vivo datasets compared to the iterative optimization methods and the state-of-the-art CNN method.


Cardiac MR imaging is a non-invasive imaging technique which could provide both spatial and temporal information for the underlying anatomy. Nevertheless, both physiological and hardware constraints have made it suffer from slow imaging speed or long imaging time, which may lead to patients’ discomfort or sometimes cause severe motion artifacts. Therefore, it is of great necessity to accelerate MR imaging. CNN-based methods [1][2] can achieve improved reconstruction results in shorter time compared to classical CS or low rank based methods [3-5]. Despite the successes, available CNN-based methods do not take full advantage of the hierarchical features from different convolutional layers, which may contain a large amount of useful local and global information. Inspired by [6], we propose a cascaded residual dense network, C-RDN, for cardiac MR image reconstruction. Specifically, residual dense block (RDB) is designed to extract abundant local features. These local features are fused to get more informative features via dense connected layers. Then, global feature fusion is performed to integrate all the dense local features. By fully utilizing both local and global features, the proposed C-RDN generated the best reconstruction results from undersampled cardiac MR images compared to both the conventional iterative optimization methods, k-t FOCUSS [3] and L+S [5], and the state-of-the-art CNN method, D5C5 [1].

Theory and method

In this work, we propose a cascaded residual dense network for cardiac MR image reconstruction (Figure 1). C-RDN consists of a cascade of RDNs. Each RDN contains five major components: shallow feature extraction, residual dense blocks (RDBs), global feature fusion, global residual learning and data consistency (DC). Firstly, Undersampled cardiac MR images are fed into the network for shallow feature extraction. Secondly, the shallow features go through D RDBs for local feature fusion. The details of one RDB are shown in Figure 2. RDB includes dense connections, local feature fusion, and local residual connections. Dense connections refer to the direct connections of each convolutional layer to subsequent layers, which can enhance the transmission of local features. All local features are concatenated together and pass through a 1*1 convolutional layer to achieve local feature fusion. Residual connections are introduced in RDB to further improve information propagation. Thirdly, these residual dense features from D RDBs are merged via global feature fusion (concatenation + 1*1 convolution). We believe the combination of local feature fusion and global feature fusion can make full use of features at different levels. Fourthly, a global residual connection combines the shallow features with the global fused features. Finally, a data consistency layer [1] is appended to correct MR images by the accurate k-space samples.


We collected 101 fully sampled cardiac MR data from using a 3T scanner (SIMENS MAGNETOM Trio) with T1-weighted FLASH sequence. Multi-coil data were combined to a single channel and then retrospectively undersampled using 1D random Cartesian masks [3]. After normalization and extraction, we got 17502 cardiac data, where 15000, 2000, and 502 were used for training, validating, and testing, respectively. The models were implemented on a Ubuntu 16.04 LTS (64-bit) operating system equipped with an Intel Xeon E5-2640 Central Processing Unit (CPU) and a Tesla TITAN Xp Graphics Processing Unit (GPU, 12GB memory). The open framework Tensorflow was used.

Results and discussion

To demonstrate the efficacy of the C-RDN, we compare it with k-t FOCUSS, L+S, and the state-of-the-art method D5C5. We adjust the parameters of the CS-MRI methods to their best performance. The reconstructions of these methods are shown in Figure 3. The two CS-based reconstructions contain less structural details than the constructions of CNN-based methods. Compared to D5C5, the proposed method can not only retain the details, but also remove the artifacts better. The evaluation metrics are presented in Table 1, which confirm that the proposed C-RDN achieved the optimal performance with the lowest MSE, the highest PSNR, and the highest SSIM among the compared methods.


This paper proposed a cascaded residual dense network, C-RDN, for dynamic MR imaging. C-RDN can fully explore both local and global features. Comparisons to k-t FOCUSS, L+S and the state-of-the-art CNN method D5C5 on in vivo datasets demonstrate that our C-RDN can effectively reconstruct cardiac MR images with improved performance.


Grant support: China NSFC 81830056, 61771463, 61471350, Science and Technology Planning Project of Guangdong Province(2017B020227012), the Basic Research Program of Shenzhen JCYJ20150831154213680, and Key Laboratory for Magnetic Resonance and Multimodality Imaging of Guangdong Province 2017YFC0108802, and US NIH R21EB020861 for Ying.


[1]. J. Schlemper, J. Caballero, J.V. Hajnal, A. Price, D. Rueckert, “A Deep Cascade of Convolutional Neural Networks for Dynamic MR Image Reconstruction”, IEEE TMI, DOI: 10. 1109/TMI.2017.2760978 (2017)

[2]. Qin, Chen, Joseph V. Hajnal, Daniel Rueckert, Jo Schlemper, Jose Caballero, and Anthony N. Price. "Convolutional recurrent neural networks for dynamic MR image reconstruction." IEEE transactions on medical imaging (2018).

[3]. Jung, Hong, Jong Chul Ye, and Eung Yeop Kim. "Improved k–t BLAST and k–t SENSE using FOCUSS." Physics in Medicine & Biology 52, no. 11 (2007): 3201.

[4]. Lingala, Sajan Goud, Yue Hu, Edward DiBella, and Mathews Jacob. "Accelerated dynamic MRI exploiting sparsity and low-rank structure: kt SLR." IEEE transactions on medical imaging30, no. 5 (2011): 1042-1054.

[5]. Otazo, Ricardo, Emmanuel Candès, and Daniel K. Sodickson. "Low‐rank plus sparse matrix decomposition for accelerated dynamic MRI with separation of background and dynamic components." Magnetic Resonance in Medicine 73, no. 3 (2015): 1125-1136.

[6]. Zhang, Yulun, Yapeng Tian, Yu Kong, Bineng Zhong, and Yun Fu. "Residual dense network for image super-resolution." In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR). 2018.


Figure 1. The proposed cascaded residual dense network for cardiac MR reconstruction.

Figure 2.Residual dense block (RDB) for local feature fusion.

Figure 3.Cardiac MR reconstruction results of different methods (k-t FOCUSS, L+S, D5C5 and the proposed method). (a) ground truth, (b) zero-filling image, (c) mask and its k-t extraction, (d-g) reconstructions of k-t FOCUSS, L+S, D5C5, and the proposed method, respectively; (h-k) their corresponding error maps.

Table 1. Evaluation metrics of the reconstruction results generated by zero-filling, k-t FOCUSS, D5C5 and the proposed method.

Proc. Intl. Soc. Mag. Reson. Med. 27 (2019)