A global-local mamba based magnetic resonance image restoration network
Patent Information
- Application Number
- CN202510513150.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-04-23
- Publication Date
- 2026-10-27
AI Technical Summary
但这些方法未能有效平衡全局感受野和高效计算之间的关系，恢复图像纹理细节的能力有限，因此限制了其在磁共振图像恢复任务中的进一步发展，仍然需要新的解决方案来应对这些问题
本发明提出了一种用于加速磁共振图像恢复的Mamba网络，该网络通过多尺度局部特征与全局特征融合的策略，有效提升了图像恢复的质量和效率；利用全局Mamba块实现全局建模，有效提取磁共振图像中的全局信息；通过多尺度局部Mamba块提取多尺度局部特征，有效恢复了图像中的纹理细节；特征融合模块实现了全局与局部特征的交互融合，进一步增强了图像恢复效果；基于Mamba的恢复网络，克服了卷积运算感受野有限的缺陷，有效地在全局感受野与计算复杂度中取得平衡，在去除图像伪影、恢复更多的纹理细节方面展现出显著优势。
Smart Images
Figure CN122989479A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of image processing, specifically relating to a magnetic resonance image restoration network based on global-local Mamba, used to accelerate magnetic resonance image acquisition. Background Technology
[0002] Magnetic resonance imaging (MRI) is a widely used medical imaging technique with advantages such as being radiation-free and providing excellent soft tissue contrast. However, its lengthy acquisition process can cause patient discomfort and may also result in severe motion artifacts due to patient movement and physiological activities, thus affecting image quality. Therefore, shortening MRI acquisition time can not only improve patient experience and image quality but also provide better support for subsequent diagnostic analysis.
[0003] To accelerate MRI scans, a common approach is to undersample k-space data to obtain low-resolution images, and then use restoration algorithms to recover texture details from these low-resolution images, improving the perceived quality of the images. With the rapid development of artificial intelligence, deep learning (DL), with its powerful learning capabilities and flexible modeling methods, has shown great potential in the field of medical image restoration. Unlike traditional model-based or sparse coding-based restoration methods, deep learning can automatically extract complex features from large-scale data using multi-layered neural networks, effectively modeling the nonlinear relationship between image content and degradation patterns.
[0004] Although Convolutional Neural Networks (CNNs) and Transformers, as two representative models, have demonstrated significant advantages in magnetic resonance image restoration tasks, these methods have failed to effectively balance the relationship between global receptive field and efficient computation, resulting in limited ability to restore image texture details. This restricts their further development in magnetic resonance image restoration, and new solutions are still needed to address these issues. Summary of the Invention
[0005] To address the above problems, this invention proposes a magnetic resonance image restoration network based on global-local Mamba, which can accelerate magnetic resonance imaging and acquire high-quality magnetic resonance images.
[0006] To achieve the above objectives, the technical solution adopted by this invention is: a magnetic resonance image restoration network based on global-local Mamba, comprising the following steps: The collected full-sampled K-space data is undersampled to obtain undersampled K-space data; The undersampled K-space data is subjected to an inverse Fourier transform to obtain an undersampled magnetic resonance image; The undersampled magnetic resonance image is input into the Mamba recovery network for recovery to obtain the recovered magnetic resonance image.
[0007] Optionally, the collected fully sampled K-space data can be undersampled to obtain undersampled K-space data, including: Fully sampled K-space data is acquired using a magnetic resonance imaging (MRI) device. A two-dimensional Cartesian random undersampling template is applied to the fully sampled K-space data, and undersampling is performed according to a set undersampling rate to obtain undersampled K-space data.
[0008] Optionally, the undersampled K-space data is subjected to an inverse Fourier transform to obtain an undersampled magnetic resonance image, including: Undersampled K-space data at various sampling rates are converted from K-space to the image domain through inverse Fourier transform to obtain undersampled magnetic resonance images.
[0009] Optionally, the Mamba recovery network employs dense connections, comprising five cascaded GL_Mamba blocks and one recovery convolutional block: The GL_Mamba block includes one global Mamba block, one data consistency module, one multi-scale local Mamba block, and one feature fusion module.
[0010] The restored convolutional block includes two 1×1 convolutions, three 3×3 convolutions, and three ELU activation functions.
[0011] Optionally, the GL_Mamba block includes one global Mamba block, one data consistency module, one multi-scale local Mamba block, and one feature fusion module, comprising: The global Mamba block includes two 2D-Mamba blocks; The multi-scale local Mamba block consists of a Window Partition, two 2D-Mamba blocks, and a Window Reverse. The feature fusion module includes channel attention, spatial attention, and multi-scale convolution; Optionally, the step of inputting the undersampled magnetic resonance image into a Mamba recovery network for recovery to obtain the recovered magnetic resonance image includes: The undersampled magnetic resonance image is input into a multi-scale local Mamba block, and the Window Partition is used to divide the undersampled magnetic resonance image into multiple image blocks of the same size that do not overlap. The image patch is processed by extracting features from two 2D-Mamba blocks to obtain a local feature map; The local feature map is restored to the same size as the undersampled magnetic resonance image using Window Reverse to obtain local features; The undersampled magnetic resonance image is input into a global Mamba block, and features are extracted using two 2D-Mamba blocks to obtain a global feature map. The global feature map is corrected using a data consistency module to obtain global features; The local features and the global features are processed by a feature fusion module to obtain fused features; The fused features are input into the recovery convolution block to obtain the final recovered magnetic resonance image.
[0012] Optionally, the step of correcting the global feature map through a data consistency module to obtain global features includes: The global feature map is transformed to K space using Fourier transform to obtain the recovered K space data; The recovered K-space data is kept consistent with the data at the corresponding positions of the fully sampled K-space data to complete the correction of the recovered K-space data, thus obtaining the corrected K-space data; The global features are obtained by performing an inverse Fourier transform on the corrected K-space data.
[0013] Optionally, the 2D-Mamba block includes layer normalization, linear layers, Sine activation functions, state-space models, and convolutional layers.
[0014] Compared with existing technologies, the beneficial effects of this solution are as follows: This invention proposes a Mamba network for accelerating magnetic resonance image restoration. This network effectively improves the quality and efficiency of image restoration through a strategy of fusing multi-scale local and global features. It utilizes global Mamba blocks to achieve global modeling, effectively extracting global information from the magnetic resonance image. Multi-scale local features are extracted through multi-scale local Mamba blocks, effectively restoring texture details in the image. The feature fusion module achieves interactive fusion of global and local features, further enhancing the image restoration effect. The Mamba-based restoration network overcomes the limitation of the limited receptive field in convolution operations, effectively balancing the global receptive field and computational complexity, and exhibits significant advantages in removing image artifacts and restoring more texture details. Attached Figure Description
[0015] To more clearly illustrate the technical solutions in the embodiments of the present invention, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. In the drawings: Figure 1 This is a schematic diagram of the magnetic resonance image restoration network based on global-local Mamba according to an embodiment of the present invention; Figure 2 This is a diagram of the magnetic resonance image restoration network structure based on global-local Mamba, as described in an embodiment of the present invention. Figure 3 This invention provides a multi-scale local Mamba block structure diagram in the network for recovery according to an embodiment of the invention. Figure 4 This is a structural diagram of the feature fusion module in the recovery network according to an embodiment of the present invention; Figure 5 This is a diagram illustrating the 2D-Mamba block structure in the network as described in an embodiment of the present invention. Figure 6 This is a visualization of the recovered image under a two-dimensional Cartesian random sampling mask, as shown in an embodiment of the present invention. Detailed Implementation
[0016] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the protection scope of the present invention.
[0017] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.
[0018] like Figure 1 As shown, this embodiment provides a magnetic resonance image restoration method based on global-local Mamba, including: Step S100: Undersample the collected full-sampled K-space data to obtain undersampled K-space data; Step S200: Perform an inverse Fourier transform on the undersampled K-space data to obtain an undersampled magnetic resonance image; Step S300: Input the undersampled magnetic resonance image into the Mamba recovery network for recovery to obtain the recovered magnetic resonance image.
[0019] In step S100, magnetic resonance imaging (MRI) equipment is used to acquire full-sample MRI data to obtain raw K-space data. A two-dimensional Cartesian random mask is then used to undersample the raw data at sampling rates of 12.5% and 25% to obtain undersampled K-space data.
[0020] In step S200, the undersampled K-space data at each sampling rate are converted from K-space to the image domain through inverse Fourier transform to obtain an undersampled magnetic resonance image.
[0021] Step S300 involves inputting the undersampled magnetic resonance image into a Mamba recovery network for recovery, thereby obtaining the recovered magnetic resonance image, including: Step S310: Input the undersampled magnetic resonance image into a multi-scale local Mamba block to obtain local features; Step S320: Input the undersampled magnetic resonance image into a global Mamba block to obtain global features; Step S330: The local features and global features are fused using a feature fusion module to obtain fused features; Step S340: Input the fused features into the recovery convolution block to obtain the final recovered magnetic resonance image.
[0022] like Figure 2 The diagram shows the structure of the magnetic resonance image restoration network based on global-local Mamba in an embodiment of the present invention. The proposed network adopts a strategy of fusing multi-scale local features and global features, which significantly improves the quality and efficiency of image restoration.
[0023] In step S310, the undersampled magnetic resonance image is input into a multi-scale local Mamba block to obtain local features, including: Step S311: Input the undersampled magnetic resonance image into a multi-scale local Mamba block, and use WindowPartition to divide the undersampled magnetic resonance image into multiple image blocks of the same size that do not overlap; Step S312: Extract features from the image patch using two 2D-Mamba blocks to obtain a local feature map; Step S313: Restore the local feature map to the same size as the undersampled magnetic resonance image using Window Reverse to obtain local features; Specifically, such as Figure 3As shown, the undersampled magnetic resonance image is first subjected to dimensionality reduction and shallow feature extraction via a convolutional layer with a kernel size of 1. The shallow features are then windowed into multiple non-overlapping image patches. Two 2D-Mamba layers are used for each patch to extract as much local feature information as possible. Subsequently, the extracted features are restored to the size of the undersampled magnetic resonance image using Window Reverse, and a convolutional layer with a kernel size of 3 is used to process the fused feature map to filter out redundant information and retain more useful features. Finally, a residual connection is established between the fused feature map and the shallow features to obtain local features.
[0024] In step S320, the undersampled magnetic resonance image is input into a global Mamba block to obtain global features, including: Step S321: Input the undersampled magnetic resonance image into a global Mamba block, extract features using two 2D-Mamba blocks, and obtain a global feature map; Step S322: Correct the global feature map using the data consistency module to obtain global features; Specifically, the undersampled magnetic resonance image is first subjected to dimensionality reduction and shallow feature extraction using a convolutional layer with a kernel size of 1. The shallow features are then extracted using two 2D-Mamba layers, and the fused feature map is processed using a convolutional layer with a kernel size of 3 to filter out redundant information and retain more useful features. A residual connection is then established between the fused and the shallow features to obtain a global feature map. This global feature map is then transformed to K-space using a Fourier transform to obtain the recovered K-space data. The recovered K-space data is then aligned with the corresponding positions in the fully sampled K-space data to complete the correction of the recovered K-space data, resulting in corrected K-space data. Finally, the corrected K-space data is subjected to an inverse Fourier transform to obtain the global features.
[0025] In step S330, the local features and global features are fused using a feature fusion module to obtain fused features, including: Step S331: Generate channel attention weights by applying average pooling, a convolution with a kernel of 1, and a sigmoid activation to the local and global features. Multiply the channel attention weights by the local and global features to obtain the channel features.
[0026] Step S332: Aggregate the local and global features using spatial attention, which is achieved through average pooling, max pooling, a convolution with a kernel of 7, and a sigmoid activation, to generate spatial attention weights. Multiply the spatial attention weights by the local and global features to obtain the spatial features.
[0027] Step S333: Aggregate the channel features and spatial features through multi-scale convolution consisting of convolutions with kernels of 3, 5, and 7 to obtain fused features; In step S340, the fused features are input into the recovery convolution block to obtain the final recovered magnetic resonance image, including: The fused features are further refined by restoring the convolutional blocks using two 1×1 convolutions, three 3×3 convolutions, and three ELU activation functions to obtain the final restored magnetic resonance image.
[0028] like Figure 4 The diagram shown is a structural diagram of the feature fusion module in the recovery network according to an embodiment of the present invention.
[0029] Step S322 corrects the recovered K-space data using a data consistency layer. The data consistency layer is a commonly used strategy to improve the quality of recovery. By utilizing a portion of the values in the undersampled data that belong to the original data, the recovered K-space data is made consistent with the data at the corresponding positions of the original values, thereby achieving the purpose of correcting the predicted data.
[0030] like Figure 5 The diagram shown is a 2D-Mamba block structure diagram in the recovery network of this invention embodiment; the 2D-Mamba block consists of layer normalization, layer normalization, linear layers, Sine activation function, state space model and convolutional layers.
[0031] To comprehensively evaluate the effectiveness of the proposed network, magnetic resonance image restoration results under two-dimensional Cartesian random masks with sampling rates of 12.5% and 25% were compared. Three quantitative evaluation metrics, including Peak Signal-to-Noise Ratio (PSNR), Normalized Root Mean Square Error (NRMSE), and Structure Similarity Index Measure (SSIM), were used to quantitatively evaluate the quality of the restored images. Figure 6 The visualization results show that the network proposed in this invention can effectively remove artifacts caused by undersampling, restore more texture details, and provide a clearer visual effect.
[0032] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be within the scope of protection of the present invention.
Claims
1. A magnetic resonance image restoration method based on global-local Mamba, characterized in that, Includes the following steps: The collected fully sampled K-space data is undersampled to obtain undersampled K-space data; The undersampled K-space data is subjected to an inverse Fourier transform to obtain an undersampled magnetic resonance image; The undersampled magnetic resonance image is input into the Mamba recovery network for recovery to obtain the recovered magnetic resonance image.
2. The magnetic resonance image restoration method based on global-local Mamba according to claim 1, characterized in that, The collected fully sampled K-space data is undersampled to obtain undersampled K-space data, including: Fully sampled K-space data is acquired using a magnetic resonance imaging (MRI) device. A two-dimensional Cartesian random undersampling template is applied to the fully sampled K-space data, and undersampling is performed according to a set undersampling rate to obtain undersampled K-space data.
3. The magnetic resonance image restoration method based on global-local Mamba according to claim 1, characterized in that, The undersampled K-space data is subjected to an inverse Fourier transform to obtain an undersampled magnetic resonance image, including: Undersampled K-space data at various sampling rates are converted from K-space to the image domain through inverse Fourier transform to obtain undersampled magnetic resonance images.
4. The magnetic resonance image restoration method based on global-local Mamba according to claim 1, characterized in that, The Mamba recovery method employs dense connections, consisting of five cascaded GL_Mamba blocks and one recovery convolutional block: The GL_Mamba block includes one global Mamba block, one data consistency module, one multi-scale local Mamba block, and one feature fusion module. The restored convolutional block includes two 1×1 convolutions, three 3×3 convolutions, and three ELU activation functions.
5. The magnetic resonance image restoration method based on global-local Mamba according to claim 1, characterized in that, The GL_Mamba block includes one global Mamba block, one data consistency module, one multi-scale local Mamba block, and one feature fusion module, comprising: The global Mamba block includes two 2D-Mamba blocks; The multi-scale local Mamba block consists of a Window Partition, two 2D-Mamba blocks, and a Window Reverse. The feature fusion module includes channel attention, spatial attention, and multi-scale convolution.
6. The magnetic resonance image restoration method based on global-local Mamba according to claim 1, characterized in that, The undersampled magnetic resonance image is input into the Mamba recovery network for recovery, and the recovered magnetic resonance image is obtained, including: The undersampled magnetic resonance image is input into a multi-scale local Mamba block, and the Window Partition is used to divide the undersampled magnetic resonance image into multiple image blocks of the same size that do not overlap. The image patch is processed by extracting features from two 2D-Mamba blocks to obtain a local feature map; The local feature map is restored to the same size as the undersampled magnetic resonance image using Window Reverse to obtain local features; The undersampled magnetic resonance image is input into a global Mamba block, and features are extracted using two 2D-Mamba blocks to obtain a global feature map. The global feature map is corrected using a data consistency module to obtain global features; The local features and the global features are processed by a feature fusion module to obtain fused features; The fused features are input into the recovery convolution block to obtain the final recovered magnetic resonance image.
7. The method according to claim 6, which involves inputting the undersampled magnetic resonance image into a Mamba recovery network for recovery to obtain a recovered magnetic resonance image, characterized in that... The step of correcting the global feature map through a data consistency module to obtain global features includes: The global feature map is transformed to K-space using Fourier transform to obtain the recovered K-space data; The recovered K-space data is kept consistent with the data at the corresponding positions of the fully sampled K-space data to complete the correction of the recovered K-space data, thus obtaining the corrected K-space data; The global features are obtained by performing an inverse Fourier transform on the corrected K-space data.
8. The global Mamba block according to claim 5 comprises two 2D-Mamba blocks, characterized in that, include: The 2D-Mamba block includes layer normalization, linear layers, Sine activation function, state-space model, and convolutional layers.