A compressed sensing MRI reconstruction method based on cross-frame guidance and dual-view interaction enhancement and a system thereof

Abstract

The application discloses a compressed sensing MRI reconstruction method based on cross-frame guidance and dual-view interactive enhancement and a system thereof. MRI image pairs are acquired, and a network including a sampling module, an initial reconstruction module and a plurality of iterative reconstruction modules is used for image compressed sensing reconstruction. Each iterative reconstruction module includes two branches, a dual-view interactive enhancement module, one branch outputs key frame reconstruction guidance information, and the other branch reconstructs auxiliary frame reconstruction images, and combines the key frame reconstruction guidance information to perform adjacent slice joint reconstruction. The application retains frequency domain information by constructing image pairs with different sampling ratios, adjacent slices have high spatial correlation, a cross-frame guidance module is used to perform feature interaction and information compensation between slices with different sampling ratios, a dual-view interactive enhancement module is used to strengthen structural continuity from different axial directions, global structural consistency is ensured, and details are retained.

Description

Technical Field

[0001] This invention relates to the fields of computer vision and medical image processing technology, specifically to a compressed sensing MRI reconstruction method and system based on cross-frame guidance and dual-view interactive enhancement. Background Technology

[0002] Magnetic Resonance Imaging (MRI) is a non-invasive imaging technique widely used in clinical and research fields. Due to its excellent soft tissue contrast and imaging safety, it plays a crucial role in neurosurgery, oncology, neuroscience, and many other areas. However, the long scan time of MRI increases patient discomfort and limits its accessibility for efficient clinical applications. To improve imaging efficiency, Compressed Sensing (CS) has been introduced into the field of MRI reconstruction. This method, through undersampling in the k-space and combined with reconstruction algorithms, significantly shortens scan time while maintaining image quality as much as possible. In CS-MRI, the rationality of the sampling strategy and the design of the reconstruction algorithm have a significant impact on the final image quality and reconstruction efficiency, thus becoming a core research issue in this field.

[0003] Recently, several Deep Unfolding Networks (DUNs) have been developed to integrate the interpretability of traditional model-based methods with the efficiency of data-driven approaches, resulting in better reconstruction performance. DUNs are deep learning models for compressed sensing that combine traditional optimization iterative methods with deep neural networks. They not only provide explicit interpretability but also fully leverage the advantages of neural networks to achieve fast and accurate signal or image reconstruction. However, existing methods have two main limitations: the first is that they only consider reconstruction under a single sampling ratio condition, ignoring the complementary information between different sampling rates, making it difficult to obtain robust and high-quality reconstruction results under multi-sampling rate or multimodal conditions; the second limitation is that they fail to effectively utilize the structural continuity and contextual information of MRI data in three-dimensional space, neglecting the connections between global structures. Summary of the Invention

[0004] The purpose of this invention is to improve the MRI imaging reconstruction effect based on compressed sensing methods by addressing issues such as high sampling ratio and neglect of global structure. A compressed sensing MRI reconstruction method and system based on cross-frame guidance and dual-view interactive enhancement is proposed.

[0005] In a first aspect, the present invention provides a compressed sensing MRI reconstruction method based on cross-frame guidance and dual-view interactive enhancement, the method comprising:

[0006] Acquire adjacent MRI image pairs and preprocess them to obtain preprocessed image pairs. and The high spatial correlation between adjacent slices is beneficial for the model to mine cross-graph information in subsequent reconstruction.

[0007] Using preprocessed images and Build the dataset;

[0008] The compressed sensing MRI reconstruction network was trained, tested, and validated using a dataset;

[0009] Image compressed sensing reconstruction was performed using a well-trained, tested, and validated compressed sensing MRI reconstruction network.

[0010] The compressed sensing MRI reconstruction network includes a sampling module, an initial reconstruction module, and multiple iterative reconstruction modules.

[0011] The sampling module is used to process the input image pairs. and Perform group sampling, where the image Divided into keyframes, image Divided into auxiliary frames, through the first mask Second mask Two compressed sensing measurements are generated by sampling the keyframe and the auxiliary frame respectively. , Among them, the first mask Second mask The sampling ratios are different;

[0012] The initial reconstruction module is used to process two compressed sensing measurements. , Initial reconstruction was performed separately to obtain the initial reconstructed images of the keyframes. and auxiliary frames initial reconstructed image ;

[0013] Each iterative reconstruction module includes two branches and a dual-view interactive enhancement module. One branch is responsible for extracting keyframe reconstruction guidance information, while the other branch is responsible for reconstructing auxiliary frame reconstruction images. It combines the keyframe reconstruction guidance information to perform joint reconstruction of adjacent slices and extracts the guided auxiliary frame images. The dual-view interactive enhancement module realizes feature interaction and structural modeling between different axes based on the keyframe reconstruction guidance information and the guided auxiliary frame images, and obtains keyframe reconstruction images and auxiliary frame reconstruction images.

[0014] Preferably, in each iterative reconstruction module, one of the two branches includes a first alternating direction multiplier method reconstruction module, whose input is the keyframe reconstructed image output by the previous iterative reconstruction module. The input to the first alternating direction multiplier method reconstruction module in the first iterative reconstruction module is initialized with the keyframe initial reconstruction image. The reconstructed output is keyframe reconstruction guidance information. Where n represents the nth iterative reconstruction module;

[0015] The other branch includes a second alternating direction multiplier method reconstruction module and a cross-frame guided reconstruction module. The input of the second alternating direction multiplier method reconstruction module is the auxiliary frame reconstruction image output by the previous iteration reconstruction module. In the first iterative reconstruction module, the input to the second alternating direction multiplier method reconstruction module is initialized with the initial reconstructed image of the auxiliary frame. The reconstructed output is auxiliary frame reconstruction guidance information. The cross-frame guidance reconstruction module is used to reconstruct guidance information based on keyframes. and auxiliary frame reconstruction guidance information Joint reconstruction of adjacent slices is performed to achieve feature guidance and detail compensation of auxiliary frames by keyframes, and the guided auxiliary frame images are extracted. ;

[0016] The dual-view interaction enhancement module is used to realize feature interaction and structural modeling between different axes. Through viewpoint switching and spatial transformation operations, features are mapped and fused between different planes, thereby capturing global continuity across slices; its input is keyframe reconstruction guidance information. and guided auxiliary frame images The output is a keyframe reconstructed image. and auxiliary frame reconstructed image ;

[0017] Finally, the compressed sensing MRI reconstruction network outputs the keyframe reconstructed image processed by the last iteration reconstruction module. and auxiliary frame reconstructed image N represents the number of iterative reconstruction modules.

[0018] Preferably, the cross-frame guided reconstruction module includes a first branch and a second branch;

[0019] The first branch includes a first convolutional block; wherein the first convolutional block is used to reconstruct guidance information from keyframes. Perform convolution to generate key K and value V;

[0020] The second branch includes a second convolutional block, a first fusion module, a second fusion module, a third fusion module, and a stitching layer; wherein, the second convolutional block is used to reconstruct guidance information from auxiliary frames. The first fusion module performs convolution processing to generate a query Q; the second fusion module performs a tensor product on the query Q generated by the first convolution block and the key K generated by the second convolution block to obtain a feature matrix; the third fusion module performs a tensor product on the tensor product result of the first fusion module after processing it with the SoftMax function and the value V of the second convolution block to obtain the fused features. The third fusion module is used to integrate the fusion features obtained by the second fusion module. and auxiliary frame reconstruction guidance information The process involves dimensional stitching, followed by sequential processing through a residual network, convolutional blocks, and activation functions. Finally, the result is combined with the fused features obtained from the second fusion module. Perform dot product; the stitching layer is used to combine the result processed by the third fusion module with the auxiliary frame reconstruction guidance information. The images are stitched together to obtain guided auxiliary frame images. .

[0021] Preferably, the dual-view interaction enhancement module includes a third convolutional block, a fourth convolutional block, a viewpoint transition module, a residual block, an inverse viewpoint transition module, a fourth fusion module, an attention module, a fifth convolutional block, and a sixth convolutional block;

[0022] The third convolutional block is used to extract guided auxiliary frame images. Features;

[0023] The fourth convolutional block is used to extract keyframe reconstruction guidance information. Features;

[0024] The perspective conversion module is used to convert the feature perspective after the result of the third convolution block and the result of the fourth convolution channel splicing process from the initial XY plane to the YZ plane to obtain a new spatial perspective.

[0025] The residual block ResNet extracts deep feature information from the results of the viewpoint transformation module;

[0026] The inverse perspective conversion module re-converts the results of the residual block ResNet from the YZ plane to the XZ plane along the axis.

[0027] The fourth fusion module adds the results of the inverse perspective conversion module, the third convolution block, and the fourth convolution block after channel splicing.

[0028] The attention module is used to dynamically adjust the amplitude weights of the fourth fusion module, thereby more accurately optimizing the feature distribution during the reconstruction process.

[0029] The fifth convolutional block is used to convolve the results of the attention module to obtain the keyframe reconstruction image. ;

[0030] The sixth convolutional block is used to convolve the results of the attention module to obtain the auxiliary frame reconstruction image. .

[0031] Preferably, the loss function of the compressed sensing MRI reconstruction network during training is the mean squared error.

[0032] In a second aspect, the present invention provides an image compressed sensing reconstruction system, comprising:

[0033] The data acquisition module is used to acquire MRI image pairs after mask sampling;

[0034] The reconstruction module takes the masked MRI image and inputs it into a trained, tested, and validated compressed sensing MRI reconstruction network to gradually reconstruct image details and finally output a high-precision reconstructed image.

[0035] Thirdly, the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed in a computer, causes the computer to perform the method.

[0036] Fourthly, the present invention provides a computing device, including a memory and a processor, wherein the memory stores executable code, and the processor executes the executable code to implement the method.

[0037] The beneficial effects of the present invention include at least the following:

[0038] 1. This invention proposes a compressed sensing MRI reconstruction method based on multi-sampling-ratio joint reconstruction, aiming to fully exploit the complementary information between adjacent slices to improve reconstruction accuracy. The method employs a sampling module to perform k-space sampling on adjacent slices using differentiated sampling rates, generating differentiated compressed sensing measurements. In the iterative reconstruction module, keyframes and auxiliary frames are jointly optimized using an alternating direction multiplier method, fully utilizing the complementary information between adjacent slices. This method effectively enhances the mutual information representation of data under limited measurement conditions, improving not only structural consistency and texture fidelity under different sampling modes but also significantly improving detail reconstruction performance in low-sampling-ratio scenarios, demonstrating excellent sampling efficiency and reconstruction stability.

[0039] 2. This invention designs a cross-frame guided reconstruction module to guide and constrain auxiliary frames from keyframes during the joint reconstruction of adjacent slices. Using keyframes as references, this module achieves high-quality information transfer and sharing through feature alignment and cross-frame attention mechanisms, thereby building efficient structural associations between slices. By introducing an adaptive weighting mechanism, the model can focus on recovering detailed and boundary region information while maintaining global consistency. This module effectively improves the spatiotemporal collaboration capability between adjacent slices, exhibiting stronger robustness and structural fidelity in complex structural regions.

[0040] 3. This invention further proposes a dual-view interactive enhancement module to overcome the limitations of traditional in-plane feature modeling. This module achieves interactive propagation of features between different spatial planes through viewpoint transformation and axis switching operations, and combines a channel-dimensional weighted fusion mechanism to capture the global continuity and local differences of the 3D structure. This design not only strengthens the network's ability to model cross-axis structural information but also improves the fusion quality of multi-view features, resulting in reconstructed images exhibiting superior visualization effects and diagnostic value in terms of texture details, tissue boundaries, and structural consistency. Attached Figure Description

[0041] To more clearly illustrate the technical solution of the present invention, the drawings used in the embodiments 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.

[0042] Figure 1 This is a schematic diagram of the MRI reconstruction network structure of the present invention.

[0043] Figure 2 This is a schematic diagram of the dual-view interactive enhancement module structure in the MRI reconstruction network of the present invention.

[0044] Figure 3 This is a schematic diagram of the cross-frame guided reconstruction module structure in the MRI reconstruction network of this invention.

[0045] Figure 4 This is a schematic diagram of the perspective conversion module in the MRI reconstruction network of this invention.

[0046] Figure 5 This is a comparison chart of the performance of the MRI reconstruction network of this invention with other advanced networks on the Brain dataset.

[0047] Figure 6 This is a comparison chart of the performance of the MRI reconstruction network of this invention with other advanced networks on the FastMRI dataset. Detailed Implementation

[0048] The present invention will be further analyzed below with reference to specific embodiments.

[0049] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0050] The terms "comprising" and "having," and any variations thereof, used in the embodiments of this invention are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or device that includes a series of steps or units is not limited to the steps or units listed, but may optionally include other steps or units not listed, or may optionally include other steps or units inherent to such processes, methods, products, or devices.

[0051] This embodiment provides a compressed sensing MRI reconstruction method based on cross-frame guidance and dual-view interactive enhancement, the method comprising:

[0052] Step S1: Obtain adjacent MRI image slice pairs and perform preprocessing to obtain preprocessed image pairs. and .

[0053] The adjacent MRI image slices are a pair of two-dimensional images that are spatially adjacent and have an index difference of 1 in the three-dimensional MRI data.

[0054] This embodiment uses the publicly available MRI datasets Brain and FastMRI for training, containing 100 brain MRI images and 4501 knee MRI images, respectively. The test set consists of 50 brain MRI images and 657 knee MRI images. The validation set comprises 50 images randomly selected from the FastMRI test set. The Brain dataset contains brain MRI images, and the FastMRI dataset contains knee MRI images, covering MRI data from different locations and at different resolutions.

[0055] The network input data is undersampled K-space data with a uniform size of 256×256. Since K-space data is inherently complex, the model treats its real and imaginary parts as two independent channels. To simulate undersampling, a two-dimensional Gaussian mask is used. Undersampling rates of 10% and 5% are applied to the two-dimensional Gaussian mask.

[0056] To ensure the model's robustness and generalization ability, all input images are real-valued MRI images during both training and testing, and all selected image samples are independent and without duplication. This approach allows the model to be effectively trained and tested under different undersampling rates and sampling modes, ensuring its reconstruction performance at low sampling rates.

[0057] As an example, the preprocessing process described above mainly includes intensity normalization of MRI images to eliminate brightness differences under different scanning conditions, removal of irrelevant background, and standardization of input size, thereby obtaining standardized image data suitable for subsequent processing.

[0058] Step S2: Use the preprocessed image to... and Build the dataset.

[0059] Step S3: Use the dataset to train, test and validate the compressed sensing MRI reconstruction network based on cross-frame guidance and dual-view interactive enhancement.

[0060] See Appendix Figure 1 The compressed sensing MRI reconstruction network (CFG-MVINet) based on cross-frame guidance and dual-view interactive enhancement includes a sampling module, an initial reconstruction module (ARM), and multiple cascaded iterative reconstruction modules.

[0061] The sampling module is used to process the input image pairs. and Perform group sampling, where the image Divided into keyframes, image Divided into auxiliary frames, using a first mask with different sampling ratios Second mask The first compressed sensing measurement value is generated by sampling the key frame and the auxiliary frame respectively. Second compressed sensing measurement value Among them, the first mask Second mask The sampling ratios are different. Specifically, the group sampling process of the sampling module is as follows:

[0062] , a = 0 or 1 (1)

[0063] in, It is the Fourier transform function.

[0064] The initial reconstruction module is used to process the first compressed sensing measurement value. Second compressed sensing measurement value Initial reconstruction is performed separately, specifically by using inverse Fourier transform to reconstruct the first compressed sensing measurements. Initial reconstruction yields keyframes and initial reconstructed images. The second compressed sensing measurement value Initial reconstruction yields auxiliary frames and initial reconstructed images. This is for use by the subsequent reconstruction module; specifically:

[0065] , a=0 or 1 (2)

[0066] in, It is the inverse Fourier transform function.

[0067] Each iterative reconstruction module includes two branches and a dual-view interactive enhancement module (DVI). One branch includes a first alternating direction multiplier method reconstruction module, and the other branch includes a second alternating direction multiplier method reconstruction module and a cross-frame guided reconstruction module (CFG).

[0068] The input to the first alternating direction multiplier method reconstruction module is the keyframe reconstruction image output by the previous iterative reconstruction module. The input to the first alternating direction multiplier method reconstruction module in the first iterative reconstruction module is initialized with the keyframe initial reconstruction image. The reconstructed output is keyframe reconstruction guidance information. Where n represents the nth iterative reconstruction module. The input of the second alternating direction multiplier method reconstruction module is the auxiliary frame reconstruction image output by the previous iterative reconstruction module. In the first iterative reconstruction module, the input to the second alternating direction multiplier method reconstruction module is initialized with the initial reconstructed image of the auxiliary frame. The reconstructed output is auxiliary frame reconstruction guidance information. In one implementation, the first alternating direction multiplier reconstruction module and the second alternating direction multiplier reconstruction module have the same structure, and the network unfolding alternating direction multiplier (ADMM) algorithm is used. The implementation process is as follows:

[0069] (3)

[0070] Where i∈{0,1} represents the keyframe and the auxiliary frame, respectively. For the sampling mask matrix, For undersampled k-space observation data, and The learnable penalty parameters for the current stage. It is the identity matrix. It is a Lagrange multiplier. This represents an auxiliary variable.

[0071] The cross-frame guidance reconstruction module is used to reconstruct guidance information based on keyframes. and auxiliary frame reconstruction guidance information Joint reconstruction of adjacent slices is performed to achieve feature guidance and detail compensation of auxiliary frames by keyframes, and the guided auxiliary frame images are extracted. This module uses keyframes as references and achieves high-quality information transfer and sharing through feature alignment and cross-frame attention mechanisms, thereby building efficient structural associations between slices. By introducing an adaptive weighting mechanism, the model can focus on recovering details and boundary region information while maintaining global consistency. This module effectively improves the spatiotemporal collaboration ability between adjacent slices and exhibits stronger robustness and structural fidelity in complex structural regions. One implementation method is described in the appendix. Figure 3 The cross-frame guided reconstruction module (CFG) includes a first branch and a second branch:

[0072] The first branch includes a first convolutional block; wherein, the second convolutional block is used to reconstruct guiding information from keyframes. Perform convolution to generate key K and value V;

[0073] The second branch includes a second convolutional block, a first fusion module, a second fusion module, a third fusion module, and a stitching layer;

[0074] The second convolutional block is used to reconstruct guidance information from the auxiliary frames. Perform convolution processing to generate query Q;

[0075] The first fusion module performs a tensor product on the query Q generated by the first convolutional block and the key K generated by the second convolutional block to obtain the feature matrix;

[0076] The second fusion module processes the tensor product result from the first fusion module using the SoftMax function, and then performs a tensor product with the value V of the second convolutional block to obtain the features. ;

[0077] The third fusion module is used to integrate the features obtained by the second fusion module. and auxiliary frame reconstruction guidance information The data is then processed through dimensional stitching, followed by sequential processing via a residual network, convolutional blocks, and a Sigmad activation function. Finally, the result is combined with the features obtained from the second fusion module. Perform dot product;

[0078] The residual network is the result of the dimensional stitching process of the third fusion module and the auxiliary frame reconstruction guidance information. Perform concat processing to extract deeper feature information;

[0079] The stitching layer is used to combine the results processed by the third fusion module with the auxiliary frame reconstruction guidance information. The images are stitched together to obtain guided auxiliary frame images. .

[0080] The dual-view interaction enhancement module is used to realize feature interaction and structural modeling between different axes. Through viewpoint switching and spatial transformation operations, features are mapped and fused between different planes, thereby capturing the global continuity across slices and overcoming the limitations of traditional in-plane feature modeling. Its input is keyframe reconstruction guidance information. and guided auxiliary frame images The output is a keyframe reconstructed image. and auxiliary frame reconstructed image This module enables the interactive propagation of features across different spatial planes through viewpoint transformation and axis switching operations. Combined with a channel-dimensional weighted fusion mechanism, it captures the global continuity and local differences of the 3D structure. This design not only enhances the network's ability to model cross-axis structural information but also improves the fusion quality of multi-view features, resulting in reconstructed images exhibiting superior visualization effects and diagnostic value in terms of texture details, tissue boundaries, and structural consistency. One implementation method is described in the appendix. Figure 2 The dual-view interaction enhancement module includes a third convolutional block, a fourth convolutional block, a view transition module, a residual block ResNet, an inverse view transition module, a fourth fusion module, an attention module CABM, a fifth convolutional block, and a sixth convolutional block.

[0081] The third convolutional block is used to extract guided auxiliary frame images. Features;

[0082] The fourth convolutional block is used to extract keyframe reconstruction guidance information. Features;

[0083] The perspective transformation module is used to transform the feature perspective of the concatenated channels of the third and fourth convolutional blocks from the initial XY plane to the YZ plane to obtain a new spatial perspective. See Appendix. Figure 4 ;

[0084] The residual block ResNet extracts deep feature information from the results of the viewpoint transformation module;

[0085] The inverse perspective transformation module re-transforms the results of the residual block ResNet from the YZ plane to the XZ plane. (See Appendix) Figure 4 ;

[0086] The fourth fusion module adds the results of the inverse perspective conversion module, the third convolution block, and the fourth convolution block after channel splicing.

[0087] The attention module CBAM is used to dynamically adjust the amplitude weights of the fourth fusion module, thereby more accurately optimizing the feature distribution during the reconstruction process.

[0088] The fifth convolutional block is used to convolve the results of the attention module CBAM to obtain the keyframe reconstruction image. ;

[0089] The sixth convolutional block is used to convolve the results of the attention module CBAM to obtain the auxiliary frame reconstruction image. .

[0090] Finally, the compressed sensing MRI reconstruction network, enhanced with cross-frame guidance and dual-view interaction, outputs the keyframe reconstructed image processed by the last iteration reconstruction module. and auxiliary frame reconstructed image N represents the number of iterative reconstruction modules.

[0091] To effectively constrain the reconstruction results of the model, Mean Squared Error (MSE) is used as the optimization objective. MSE loss directly measures the pixel-level difference between the original and reconstructed images, encouraging the model to generate results consistent with the real image in overall intensity distribution. One implementation method involves using a loss function based on a compressed sensing MRI reconstruction network enhanced with cross-frame guidance and dual-viewpoint interaction during training. for:

[0092] (4)

[0093] in and This represents a batch of trainable MRI image pairs. and This represents a pair of reconstructed MRI images.

[0094] In the training of the CFG-MVINet model, undersampled K-space data is input into a compressed sensing MRI reconstruction network (CFG-MVINet) based on cross-frame guidance and dual-view interactive enhancement for training. The model generates high-quality reconstructed images. The model is tested using two widely used benchmark datasets, Brain and FastMRI. The CFG-MVINet network can reconstruct arbitrarily undersampled MRI data, outputting high-fidelity MRI images. The quality of the reconstructed images is evaluated using peak signal-to-noise ratio (PSNR) and structural similarity index (SSIM) to ensure that the reconstruction results achieve high levels of both quantitative metrics and visual quality.

[0095] Step S4: Perform compressed sensing reconstruction of the image using a trained, tested, and validated compressed sensing MRI reconstruction network based on cross-frame guidance and dual-view interactive enhancement.

[0096] A single MRI image, sampled using a mask, is input into the pre-trained compressed sensing MRI reconstruction network CFG-MVINet, which is based on cross-frame guidance and dual-view interactive enhancement. The model performs feature extraction and reconstruction through multiple cascaded iterative reconstruction modules, ultimately outputting a high-fidelity reconstructed image corresponding to the MRI data. The reconstructed image can be used for clinical diagnosis or other medical image analysis tasks, significantly improving the efficiency and quality of MRI imaging.

[0097] Comparison of model experimental results:

[0098] As shown in Table 1, experimental results on the fastMRI and Brain datasets validate the effectiveness of the proposed method. Table 1 also shows the results on the sampling ratio... Under the condition of ∈{0.05, 0.10}, the method of this invention consistently outperforms other compressed sensing magnetic resonance imaging (CS-MRI) methods in terms of quantitative metrics. Particularly at a sampling ratio of 0.05, compared to PUERT, CDDN, and ISTA-Net+, CFG-MVINet improves the PSNR by 0.83 dB, 0.74 dB, and 1.60 dB on the fastMRI dataset, respectively, and the SSIM metrics are improved by 0.0753, 0.0550, and 0.0667, respectively, fully demonstrating the robustness and advantages of the proposed method at extremely low sampling ratios.

[0099] also, Figure 5 and Figure 6 Demonstrates the effectiveness of CFG-MVINet compared to other methods on the fastMRI and Brain datasets under different sampling ratios. = 5% and Visual comparison of reconstructed image details at a sampling ratio of 10%. It can be observed that CFG-MVINet, due to its adjacent slice reconstruction mechanism, requires two different masks for a given sampling ratio: the first is the keyframe, and the second is the auxiliary frame. The keyframe images generated by CFG-MVINet exhibit higher fidelity in detail and edge structure, and the reconstruction results of the auxiliary frames also retain relatively clear structural information. Compared with the results of CFG-MVINet, the images reconstructed by LTwIST and PUERT show some blurring in detail and insufficient edge sharpness, resulting in a poorer overall visual effect. This indicates that the CFG-MVINet proposed in this invention can achieve more accurate structure restoration and detail preservation under different sampling ratios and different datasets, outperforming existing methods not only in quantitative metrics but also demonstrating a stronger advantage in subjective visual perception.

[0100] Table 2 illustrates the impact of the number of iteration stages on the reconstructed image quality. We set 5, 7, 9 (default), and 11 reconstruction stages in the network and compared the experimental results. With increasing iteration count, the overall image quality gradually improves: from 5 to 7 iterations, PSNR improves by approximately 1.2 dB, and SSIM improves by approximately 0.01. When the number of iterations further increases to 9, the reconstructed image quality reaches a stable equilibrium point, with PSNR still showing an improvement of approximately 0.01 dB, better preserving detail while avoiding the computational overhead of excessive iteration. In contrast, the performance improvement from 9 to 11 iterations is relatively small and insufficient to offset the additional computational cost. Therefore, 9 iteration stages achieve a better balance between image quality and computational efficiency and have been selected as the default configuration for this method.

[0101] Table 3 explores the functional roles of each module by configuring different components in CFG-MVINet. We first validated the proposed modules, and the results are shown in the table. Both the Cross-Frame Guided Reconstruction (CFG) module and the Dual-View Interactive Enhancement (DVI) module played crucial roles. The table presents experimental results for the benchmark method and its different variants. It can be seen that the Cross-Frame Guided Reconstruction module exhibits efficient feature recovery capabilities in magnetic resonance image reconstruction, significantly improving overall performance. These results demonstrate that the two modules can work synergistically during image reconstruction, effectively recovering more detailed features and thus enhancing the reconstruction quality of magnetic resonance images.

[0102] Table 1. Comparison of PSNR (dB) and SSIM of CFG-MVINet and several advanced methods on FASTMRI and Brain at different sampling ratios.

[0103]

[0104] Table 2. Performance comparison of CFG-MVINet at different iteration stages on different datasets.

[0105]

[0106] Table 3 shows the performance of CFG-MVINet using different modules on brain and fast magnetic resonance imaging datasets.

[0107]

[0108] This invention also provides an image compressed sensing reconstruction system, comprising:

[0109] The data acquisition module is used to acquire MRI image pairs after mask sampling;

[0110] The reconstruction module is used to input the masked MRI image into a trained, tested, and validated compressed sensing MRI reconstruction network based on cross-frame guidance and dual-view interactive enhancement, gradually reconstructing image details and finally outputting a high-precision reconstructed image.

[0111] This invention also provides an electronic device, specifically, the electronic device includes a memory and a processor, the memory stores executable code, and when the processor executes the executable code, it implements the method described in any of the embodiments.

[0112] The memory may include high-speed random access memory (RAM) or non-volatile memory, such as at least one disk drive. Communication between this system network element and at least one other network element is achieved through at least one communication interface (which can be wired or wireless), such as the Internet, wide area network, local area network, or metropolitan area network.

[0113] The bus can be an ISA bus, PCI bus, or EISA bus, etc. The bus can be divided into address bus, data bus, control bus, etc.

[0114] The memory is used to store programs. After receiving an execution instruction, the processor executes the program. The method executed by the device for defining the flow process disclosed in any of the foregoing embodiments of the present invention can be applied to the processor or implemented by the processor.

[0115] The processor may be an integrated circuit chip with signal processing capabilities. In implementation, the steps of the above methods can be completed by integrated logic circuits in the processor's hardware or by software instructions. The processor can be a general-purpose processor, including a Central Processing Unit (CPU), a Network Processor (NP), etc.; it can also be a Digital Signal Processor (DSP), an Application Specific Integrated Circuit (ASIC), a Field-Programmable Gate Array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. It can implement or execute the methods, steps, and logic block diagrams disclosed in the embodiments of this invention. The general-purpose processor can be a microprocessor or any conventional processor. The steps of the methods disclosed in the embodiments of this invention can be directly embodied in the execution of a hardware decoding processor, or executed by a combination of hardware and software modules in the decoding processor. The software modules can reside in random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, registers, or other mature storage media in the art. The storage medium is located in the memory, and the processor reads the information in the memory and, in conjunction with its hardware, completes the steps of the above method.

[0116] The computer program product of the readable storage medium provided in the embodiments of the present invention includes a computer-readable storage medium storing program code. The instructions included in the program code can be used to execute the methods described in the foregoing method embodiments. For specific implementation, please refer to the foregoing method embodiments, which will not be repeated here.

[0117] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, essentially, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0118] Finally, it should be noted that the above-described embodiments are merely specific implementations of the present invention, used to illustrate the technical solutions of the present invention, and not to limit it. The scope of protection of the present invention is not limited thereto. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that any person skilled in the art can still modify or easily conceive of changes to the technical solutions described in the foregoing embodiments within the technical scope disclosed in the present invention, or make equivalent substitutions for some of the technical features; and these modifications, changes, or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

[0119] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications are also considered to be within the scope of protection of the present invention.

Claims

1. A compression sensing MRI reconstruction method based on cross-frame guidance and dual-view interaction enhancement, comprising acquiring a pair of adjacent MRI images and preprocessing, constructing a data set; training, testing and verifying the MRI reconstruction network using the data set; using the trained, tested and verified MRI reconstruction network for image compression sensing reconstruction; characterized in that: The MRI reconstruction network comprises a sampling module, an initial reconstruction module and a plurality of iterative reconstruction modules; ​ The sampling module performs grouped sampling on input image pairs respectively as key frames and auxiliary frames, to generate two compressed sensing measurement values; The initial reconstruction module performs initial reconstruction on the two compressed sensing measurement values respectively, to obtain key frame initial reconstruction images and auxiliary frame initial reconstruction images; Each iterative reconstruction module comprises two branches and a dual-view interactive enhancement module; one branch of the two branches is responsible for extracting key frame reconstruction guide information, and the other branch is responsible for reconstructing auxiliary frame reconstruction images, combining the key frame reconstruction guide information to perform adjacent slice joint reconstruction, and extracting guided auxiliary frame images; the dual-view interactive enhancement module obtains key frame reconstruction images and auxiliary frame reconstruction images according to the key frame reconstruction guide information and the guided auxiliary frame images.

2. The method of claim 1, wherein, In each iterative reconstruction module, one branch of the two branches comprises a first alternating direction multiplier method reconstruction module, the input of which is the key frame reconstruction image output by the previous iterative reconstruction module, and the output after reconstruction is the key frame reconstruction guide information; the other branch comprises a second alternating direction multiplier method reconstruction module and a cross-frame guided reconstruction module, the input of the second alternating direction multiplier method reconstruction module is the auxiliary frame reconstruction image output by the previous iterative reconstruction module, and the output after reconstruction is auxiliary frame reconstruction guide information; the cross-frame guided reconstruction module is used to perform adjacent slice joint reconstruction according to the key frame reconstruction guide information and the auxiliary frame reconstruction guide information, and extract guided auxiliary frame images.

3. The method of claim 1, wherein, In each iterative reconstruction module, the dual-view interactive enhancement module is used to realize feature interaction and structure modeling between different axial directions, map and fuse the features between different planes through view switching and spatial conversion operations, so as to capture the global continuity across slices.

4. The method of claim 1, wherein, The sampling module samples the key frames and the auxiliary frames through two masks with different sampling ratios respectively, to generate two compressed sensing measurement values.

5. The method of claim 2, wherein, The cross-frame guided reconstruction module comprises a first branch and a second branch; The first branch comprises a first convolution block, which is used for convolution processing on the key frame reconstruction guide information to generate a key K and a value V; The second branch comprises a second convolution block, a first fusion module, a second fusion module, a third fusion module and a splicing layer; wherein the second convolution block is used for convolution processing on the auxiliary frame reconstruction guide information to generate a query Q; the first fusion module performs tensor product on the query Q generated by the first convolution block and the key K generated by the second convolution block, to obtain a feature matrix; the second fusion module performs tensor product on the tensor product result processed by the first fusion module after SoftMax function processing and the value V of the second convolution block, to obtain a fusion feature; the third fusion module is used for latitude splicing processing on the fusion feature obtained by the second fusion module and the auxiliary frame reconstruction guide information, and then sequentially processes through a residual network, a convolution block and an activation function, and finally performs point multiplication on the result and the fusion feature obtained by the second fusion module; the splicing layer is used for splicing the result processed by the third fusion module and the auxiliary frame reconstruction guide information, to obtain the guided auxiliary frame image.

6. The method of claim 1, wherein, The dual-view interactive enhancement module comprises a third convolutional block, a fourth convolutional block, a view conversion module, a residual block, an inverse view conversion module, a fourth fusion module, an attention module, a fifth convolutional block, and a sixth convolutional block; the third convolutional block is configured to extract features of a guided auxiliary frame image; the fourth convolutional block is configured to extract features of key frame reconstruction guide information; the view conversion module is configured to convert the features of the third convolutional block and the fourth convolutional block from an XY plane to a YZ plane to obtain a new spatial view; the residual block is configured to extract deep feature information from the result of the view conversion module; the inverse view conversion module is configured to convert the result of the residual block from the YZ plane to an XZ plane; the fourth fusion module is configured to add the result of the inverse view conversion module and the result of the third convolutional block and the fourth convolutional block after channel splicing processing; the attention module is configured to dynamically adjust the amplitude weight of the fourth fusion module; the fifth convolutional block is configured to convolve the result of the attention module to obtain a key frame reconstruction image; and the sixth convolutional block is configured to convolve the result of the attention module to obtain an auxiliary frame reconstruction image.

7. The method of claim 1 wherein, In the training process, the loss function of the MRI reconstruction network adopts a mean square error.

8. An image compressive sensing reconstruction system implementing the method of any one of claims 1-7, characterized in that The method comprises the following steps: The data acquisition module is configured to acquire a pair of mask-sampled MRI images; The reconstruction module is configured to input the pair of mask-sampled MRI images into a trained, tested, and verified MRI reconstruction network, gradually reconstruct image details, and finally output a reconstructed image.

9. A computer readable storage medium having stored thereon a computer program, characterized in that, When the computer program is executed in the computer, the computer is caused to execute the method of any one of claims 1-7.

10. A computing device comprising a memory and a processor, wherein, The memory stores executable code, and the processor executes the executable code to implement the method of any one of claims 1-7.

Images