Recently, unsupervised deep learning methods have shown great potential in image processing. Compared with a large-amount demand for paired training data of supervised methods with a specific task, unsupervised methods can learn a universal and explicit prior information on data distribution and integrate it into the reconstruction process. Therefore, it can be used in various image reconstruction environments without showing degraded performance. The importance of unsupervised learning in MRI reconstruction appears to be growing. Nevertheless, the establishment of prior formulation in unsupervised deep learning varies a lot depending on mathematical approximation and network architectures. In this work, we summarized basic concepts of unsupervised deep learning comprehensively and compared performances of several state-of-the-art unsupervised learning methods for MRI reconstruction.

With the development of deep-learning techniques, the application of deep learning in MR imaging processing seems to be growing. Accordingly, deep learning has also been introduced in motion correction and seemed to work as well as do conventional motion-compensation methods. In this article, we review the motion-correction methods based on deep learning, focusing especially on the motion-simulation methods adopted. We then propose a new motion-simulation tool, which we call view2Dmotion.

Purpose: To understand the effects of datasets with various parameters on pre-trained network performance, the generalization capacity of the artificial neural network for myelin water imaging (ANN-MWI) is explored by testing datasets with various scan protocols (i.e., resolution and refocusing RF pulse shape) and types of disorders (i.e., neuromyelitis optica and edema). Materials and Methods: ANN-MWI was trained to generate a T₂ distribution, from which the myelin water fraction value was measured. The training and test datasets were acquired from healthy controls and multiple sclerosis patients using a multi-echo gradient and spin-echo sequence with the same scan protocols. To test the generalization capacity of ANN-MWI, datasets with different settings were utilized. The datasets were acquired or generated with different resolutions, refocusing pulse shape, and types of disorders. For all datasets, the evaluation was performed in a white matter mask by calculating the normalized root-mean-squared error (NRMSE) between the results from the conventional method and ANN-MWI. Additionally, for the patient datasets, the NRMSE was calculated in each lesion mask. Results: The results of ANN-MWI showed high reliability in generating myelin water fraction maps from the datasets with different resolutions. However, the increased errors were reported for the datasets with different refocusing pulse shapes and disorder types. Specifically, the region of lesions in edema patients reported high NRMSEs. These increased errors indicate the dependency of ANN-MWI on refocusing pulse flip angles and T₂ characteristics. Conclusion: This study proposes information about the generalization accuracy of a trained network when applying deep learning to processing myelin water imaging.

Dynamic magnetic resonance (MR) imaging has generated great research interest, because it can provide both spatial and temporal information for clinical diagnosis. However, slow imaging speed or long scanning time is still a challenge for dynamic MR imaging. Most existing methods reconstruct dynamic MR images from incomplete k-space data under the guidance of compressed sensing (CS) or low-rank theory, which suffer from long iterative reconstruction time. Recently, deep learning has shown great potential in accelerating dynamic MR. Our previous work proposed a dynamic MR imaging method with both k-space and spatial prior knowledge integrated via multi-supervised network training. Nevertheless, there was still some smoothing needed in the reconstructed images at high acceleration. In this work, we propose cascaded residual dense networks for dynamic MR imaging with edge-enhanced loss constraint, dubbed cascaded residual dense networks (CRDN). Specifically, the cascaded residual dense networks fully exploit the hierarchical features from all the convolutional layers with both local and global feature fusion. We further use the higher-degree total variation loss function, which has the edge enhancement properties, for training the networks.

Purpose: Image registration is a fundamental task in various medical imaging studies and clinical image analyses, such as comparison of patient data with anatomical structures. In order to solve the problems of conventional image registration approaches, such as long computational time, recent deep-learning supervised and unsupervised methods have been extensively studied because of their excellent performance and fast computational time. In this study, we propose a deep-learning-based network for deformable medical image registration using unsupervised learning. Materials and Methods: In this paper, we solve the image-registration optimization problem by modelling a function using a convolutional neural network with polyphase decomposition to learn the spatial transformable parameters based on the input images and to generate the registration field. A spatial transformer is used to reconstruct the output warped image while imposing smoothness constraints on the registration field. With polyphase decomposition, our proposed method learns more features based on the input image pairs without the need for any ground-truth registration field. Results: Experimental results using 3D T1 brain MRI volume scans and compared with state-of-the-art image-registration methods demonstrated that our method provides better 3D-image registration. Conclusion: Our proposed method uses less computational time in registering unseen pairs of input images during inference and can be applied for other unimodal image registration tasks, and the hyper-parameters can be adjusted for the specific task.

Purpose: To generate the under-sampling pattern using a self-supervised learning framework based on a graph convolutional network. Materials and Methods: We first decoded the k-space data into the graph and put it into the network. After the processing of graph convolution layers and graph pooling layers, the network generated the under-sampling pattern for MR reconstruction. We trained the network on the simulated brain dataset enabled by the self-supervised learning strategy. We did simulation along with the in vivo brain and liver experiments under different noise levels and accelerating factors to compare the performance between the proposed method and traditional methods using the PSNR and SSIM index. Results: The simulation experiments showed that the proposed method can achieve the best performance with low accelerating factors (2 and 3) at all noise levels and in high accelerating factors (4 and 5) at high noise levels (50 and 70 dB). In in vivo experiments, the proposed method attained the highest PSNR and SSIM in the brain dataset as well as in the liver dataset after fine tuning on a small liver dataset. Conclusion: The self-supervised learning framework based on a graph convolutional network was able to design the under-sampling mask for MR reconstruction. The superior performance in the simulation and in vivo experiments demonstrated the feasibility and flexibility of the proposed method and its potential in clinical use.

Purpose: To develop qMTNet⁺, an improved version of a recently proposed neural network called qMTNet, to accelerate quantitative magnetization transfer (qMT) imaging acquisition and processing. Materials and Methods: Conventional and inter-slice qMT data were acquired with two flip angles at six offset frequencies from seven subjects for developing networks and from four young and four older subjects for testing its generalizability. qMTNet⁺ was designed to incorporate residual and multi-task learning to improve its performance. qMTNet⁺ is composed of multiple fully-connected layers. It can simultaneously generate missing MT-weighted images and qMT parameters from undersampled MT images. The network was trained and validated with 7-fold cross-validation. Additional testing with unseen data was performed to assess the generalizability of the network. Performance of qMTNet⁺ was compared with that of qMTNet-1 and qMTNet-acq for fitting and MT images generation, respectively. Results: qMTNet⁺ achieved quantitative results that were better than qMTNet across all metrics (peak signal-to-noise ratio, structural similarity index, normalized mean squared error) for both conventional and inter-slice MT data. Produced offset images were better quantitatively than those produced by qMTNet-acq. Conclusion: qMTNet⁺ improves qMTNet, generating qMT parameters from undersampled MT data with higher agreement with ground truth values. Additionally, qMTNet⁺ can produce both qMT parameters and unsampled MT images with a single network in an end-to-end manner, obviating the need for separate networks required for qMTNet. qMTNet⁺ has the potential to accelerate qMT imaging for diagnostic and research purposes.