Search

Purpose: Recent studies have witnessed that self-attention modules can better solve the vision understanding problems by capturing long-range dependencies. However, there are very few works designing a lightweight self-attention module to improve the quality of MRI reconstruction. Furthermore, it can be observed that several important self-attention modules (e.g., the non-local block) cause high computational complexity and need a huge number of GPU memory when the size of the input feature is large. The purpose of this study is to design a lightweight yet effective spatial orthogonal attention module (SOAM) to capture long-range dependencies, and develop a novel spatial orthogonal attention generative adversarial network, termed as SOGAN, to achieve more accurate MRI reconstruction., Methods: We first develop a lightweight SOAM, which can generate two small attention maps to effectively aggregate the long-range contextual information in vertical and horizontal directions, respectively. Then, we embed the proposed SOAMs into the concatenated convolutional autoencoders to form the generator of the proposed SOGAN., Results: The experimental results demonstrate that the proposed SOAMs improve the quality of the reconstructed MR images effectively by capturing long-range dependencies. Besides, compared with state-of-the-art deep learning-based CS-MRI methods, the proposed SOGAN reconstructs MR images more accurately, but with fewer model parameters., Conclusions: The proposed SOAM is a lightweight yet effective self-attention module to capture long-range dependencies, thus, can improve the quality of MRI reconstruction to a large extent. Besides, with the help of SOAMs, the proposed SOGAN outperforms the state-of-the-art deep learning-based CS-MRI methods., (© 2020 American Association of Physicists in Medicine.)

Low-rank structure is a powerful priori characteristic that is exploited in constrained magnetic resonance imaging (MRI). In this paper, we build two low rank matrices T V and T H from weighted k-space data according to the duality between the sparsity in the difference image and the low-rankness of a reciprocal spectral domain. Then, we propose a two-step constrained MR image reconstruction method. First, the vertical and horizontal difference images are recovered via enforcing low-rankness of matrices T V and T H . Then, the image is reconstructed via the least squares method. In the first step, the nuclear norm of a matrix is replaced by the minimum Frobenius norm of two factorization matrices and the alternating direction method of multipliers (ADMM) algorithm is applied to recover the difference images. This singular value decomposition (SVD) free method leads to fast reconstruction. The experimental results demonstrate that the proposed method outperforms other low rank based methods., (Copyright © 2019 Elsevier Inc. All rights reserved.)

This work aims to develop a novel magnetic resonance (MR) image reconstruction approach motivated by the recently proposed sampling framework with union-of-subspaces model (SUoS). Based on SUoS, we propose a mathematical formalism that effectively integrates a block sparsity constraint and support information which is estimated in an iterative fashion. The resulting optimization problem consists of a data fidelity term and a support detection based block sparsity (SDBS) promoting term penalizing entries within the complement of the estimated support. We provide optional strategies for block assignment, and we also derive unique and robust recovery conditions in terms of the structured restricted isometric property (RIP), namely the block-RIP. The block-RIP constant we derive is lower than that of the previous structured sparse method, which leads to a reduction of the measurements. Simulation results for reconstructing individual and multiple T1/T2-weighted images demonstrate the consistency with our theoretical claims, and show considerable improvement in comparison with methods using only block sparsity or support information., (Copyright © 2015 Elsevier Inc. All rights reserved.)

Compressed sensing (CS) based methods make it possible to reconstruct magnetic resonance (MR) images from undersampled measurements, which is known as CS-MRI. The reference-driven CS-MRI reconstruction schemes can further decrease the sampling ratio by exploiting the sparsity of the difference image between the target and the reference MR images in pixel domain. Unfortunately existing methods do not work well given that contrast changes are incorrectly estimated or motion compensation is inaccurate. In this paper, we propose to reconstruct MR images by utilizing the sparsity of the difference image between the target and the motion-compensated reference images in wavelet transform and gradient domains. The idea is attractive because it requires neither the estimation of the contrast changes nor multiple times motion compensations. In addition, we apply total generalized variation (TGV) regularization to eliminate the staircasing artifacts caused by conventional total variation (TV). Fast composite splitting algorithm (FCSA) is used to solve the proposed reconstruction problem in order to improve computational efficiency. Experimental results demonstrate that the proposed method can not only reduce the computational cost but also decrease sampling ratio or improve the reconstruction quality alternatively.

Purpose: It is challenging for radiologists and gynecologists to identify the type of ovarian lesions by reading magnetic resonance (MR) images. Recently developed convolutional neural networks (CNNs) have made great progress in computer vision, but their architectures still need modification if they are used in processing medical images. This study aims to improve the feature extraction capability of CNNs, thus promoting the diagnostic performance in discriminating between benign and malignant ovarian lesions. Methods: We introduce a feature fusion architecture and insert the attention models in the neural network. The features extracted from different middle layers are integrated with reoptimized spatial and channel weights. We add a loss function to constrain the additional probability vector generated from the integrated features, thus guiding the middle layers to emphasize useful information. We analyzed 159 lesions imaged by dynamic contrast‐enhanced MR imaging (DCE‐MRI), including 73 benign lesions and 86 malignant lesions. Senior radiologists selected and labeled the tumor regions based on the pathology reports. Then, the tumor regions were cropped into 7494 nonoverlapping image patches for training and testing. The type of a single tumor was determined by the average probability scores of the image patches belonging to it. Results: We implemented fivefold cross‐validation to characterize our proposed method, and the distribution of performance matrics was reported. For all the test image patches, the average accuracy of our method is 70.5% with an average area under the curve (AUC) of 0.785, while the baseline is 69.4% and 0.773, and for the diagnosis of single tumors, our model achieved an average accuracy of 82.4% and average AUC of 0.916, which were better than the baseline (81.8% and 0.899). Moreover, we evaluated the performance of our proposed method utilizing different CNN backbones and different attention mechanisms. Conclusions: The texture features extracted from different middle layers are crucial for ovarian lesion diagnosis. Our proposed method can enhance the feature extraction capabilities of different layers of the network, thereby improving diagnostic performance. [ABSTRACT FROM AUTHOR]

Total variation (TV) regularization is a technique commonly utilized to promote sparsity of image in gradient domain. In this article, we address the problem of MR brain image reconstruction from highly undersampled Fourier measurements. We define the Moreau enhanced function of  L1 norm, and introduce the minmax‐concave TV (MCTV) penalty as a regularization term for MR brain image reconstruction. MCTV strongly induces the sparsity in gradient domain, and fits the frame of fast algorithms (eg, ADMM) for solving optimization problems. Although MCTV is non‐convex, the cost function in each iteration step can maintain convexity by specifying the relative nonconvexity parameter properly. Experimental results demonstrate the superior performance of the proposed method in comparison with standard TV as well as non‐local TV minimization method, which suggests that MCTV may have promising applications in the field of neuroscience in the future. [ABSTRACT FROM AUTHOR]

Medical image fusion increases accuracy of clinical diagnosis and analysis through integrating complementary information of multi-modality medical images. A novel multi-modality medical image fusion algorithm exploiting a moving frame based decomposition framework (MFDF) and the nonsubsampled shearlet transform (NSST) is proposed. The MFDF is applied to decompose source images into texture components and approximation components. Maximum selection fusion rule is employed to fuse texture components aimed at transferring salient gradient information to the fused image. The approximate components are merged using NSST. Finally, a components synthesis process is adopted to produce the fused image. Experimental results verify that the proposed method achieves better performance than other compared state-of-art methods in both visual effects and objective criteria. [ABSTRACT FROM AUTHOR]

Multimodal medical image fusion technique contributes to the reduction of information uncertainty and improves the clinical diagnosis accuracy, the aim of which is to preserve salient image features and detail information of multiple source images to produce a visual enhanced single fused image. In this paper, by taking full advantages of structure tensor and nonsubsampled shearlet transform (NSST) to effectively extract geometric features, a novel unified optimization model is proposed for fusing computed tomography and magnetic resonance imaging images. The proposed model includes two terms, which constrain the gradient and NSST coefficients of the final fused image close to the pre-fused gradient and NSST coefficients. The pre-fused gradient is obtained from weighted structure tensor and the pre-fused NSST coefficients are generated by fusing NSST coefficients of source images with proposed fusion rules. The final fused image can be obtained by solving the constructed optimization problem via conjugate gradient method. Experimental results demonstrate that the proposed approach outperforms the compared multi-resolution and gradient based methods in both visual effects and quantitative assessments. [ABSTRACT FROM AUTHOR]

We propose a novel multimodality medical image fusion algorithm which involves L 0 gradient minimization smoothing filter (GMSF) and pulse coupled neural network (PCNN). Firstly, an excellent multi-scale edge-preserving decomposition framework based on GMSF is proposed to decompose each source image into one base image and a series of detail images. For extracting and preserving more salient features and detail information, different fusion rules are designed to fuse the separated subimages. The base images are fused using the regional weighted sum of pixel energy and gradient energy, and a biologically inspired feedback neural network is used to fuse the detail images. The final fused image is obtained by synthesizing the fused base image and detail images. Experimental results on several datasets of CT and MRI images show that the proposed algorithm outperforms other compared methods in terms of both subjective and objective assessment. [ABSTRACT FROM AUTHOR]

ABSTRACT Compressed sensing (CS) enables to reconstruct MR images from highly undersampled k-space data by exploiting the sparsity which is implicit in the images. In this article, an MR image [ABSTRACT FROM AUTHOR]

ABSTRACT This article addresses the problem of reconstructing a magnetic resonance image from highly undersampled data, which frequently arises in accelerated magnetic resonance imaging. We propose to impose sparsity of first and second order difference sparse coefficients within the complement of the known support. Second order variation is involved to overcome blocky effects and support information is used to reduce the sampling rate further. The resulting optimization problem consists of a data fidelity term and first-second order variation terms penalizing entries within the complement of the known support. The efficient split Bregman algorithm is used to solve the problem. Reconstruction results from magnetic resonance imaging data corresponding to different sampling rates are shown to illustrate the performance of the proposed method. Then, we also assess the tolerance of the new method to noise briefly. © 2015 Wiley Periodicals, Inc. Int J Imaging Syst Technol, 25, 277-284, 2015 [ABSTRACT FROM AUTHOR]

Compressed sensing based magnetic resonance imaging (CS-MRI) methods greatly shorten the scanning time while ensuring the quality of image reconstruction in an efficient way. Recently deep learning has been introduced into MRI reconstruction to further improve the image quality and shorten reconstruction time. In this paper, we propose an efficient structurally-strengthened Generative Adversarial Network, termed as ESSGAN, for reconstructing MR images from highly under-sampled k -space data. ESSGAN consists of a structurally-strengthened generator (termed as SG) and a discriminator. In SG, we introduce strengthened connections to enhance feature propagation and reuse between the concatenated strengthened convolutional autoencoders (termed as SCAEs), where each SCAE is a variant of a typical convolutional autoencoder. In addition, we creatively introduce the residual in residual blocks (termed as RIRBs) to SG. The RIRBs can effectively enhance the expression ability of the proposed SG. To further preserve more image details, we introduce an enhanced structural loss to SG. The experimental results demonstrate that ESSGAN can provide higher image quality with fewer model parameters than the state-of-the-art deep learning-based methods at different undersampling rates under different types of undersampling patterns. [ABSTRACT FROM AUTHOR]