MRI (Magnetic Resonance Imaging) reconstruction method and system based on multi-scale soft threshold operation and storage medium

Abstract

The invention discloses an MR I reconstruction method and system based on multi-scale soft threshold operation, and a storage medium. The method comprises the following steps: obtaining full-sampling MR I data, and carrying out simulation undersampling to obtain simulation undersampling data; inputting the analog under-sampled data into a reconstruction network for iterative reconstruction, wherein the reconstruction network is initialized to input the analog under-sampled data; and carrying out iterative reconstruction, which specifically comprises the steps of extracting omnibearing features, carrying out auxiliary feature transmission, carrying out multi-scale feature fusion and carrying out loss calculation and model optimization. And outputting a final reconstructed image after iterative reconstruction of a preset number of iterations. According to the invention, through the technical means of multi-scale feature extraction, local and global feature fusion, feature transmission between iterations, improved soft threshold processing and the like, key problems in the existing MR I reconstruction method are effectively solved.

Description

Technical Field

[0001] The present invention relates to the field of image processing technology, and more specifically, to an MRI reconstruction method, system and storage medium based on multi-scale soft threshold operation. Background Art

[0002] Magnetic resonance imaging (MRI) has become an indispensable tool in clinical diagnosis and medical research due to its excellent characteristics such as no radiation, high resolution, and high contrast. However, since MRI requires multiple radio frequency pulses to fill the k-space, its acquisition time is often longer than other medical imaging methods such as X-rays or computed tomography, and it is susceptible to motion artifacts. In order to improve the performance of MRI reconstruction, a variety of representative methods have been proposed in the past few decades, covering partial Fourier reconstruction, parallel imaging, sparse-driven compressed sensing, and emerging reconstruction methods based on deep learning models.

[0003] Partial Fourier reconstruction techniques acquire partial k-space data and cleverly exploit its symmetry to complete reconstruction, effectively shortening acquisition time and reducing artifacts. However, this technique's reconstruction performance is somewhat limited when used for high-resolution images with complex structures. Parallel imaging techniques utilize multiple receiving coils to simultaneously acquire data and employ specialized algorithms for reconstruction, significantly improving acquisition speed and image quality. However, even at high acceleration factors, noise enhancement and residual artifacts remain unavoidable. Compressed sensing techniques, by precisely exploiting signal sparsity, can accurately reconstruct signals from extremely limited sampled data, achieving remarkable results in significantly reducing acquisition time. However, at high acceleration factors, they are prone to image blurring and loss of detail. In recent years, deep learning technology has emerged as a significant force in the field of medical image reconstruction, achieving a series of impressive results. Deep learning-based reconstruction methods, particularly at high acceleration factors, achieve faster reconstruction speeds and higher-quality images through end-to-end training. For example, the deep unfolding model cleverly maps the steps of traditional iterative reconstruction algorithms into the layered structure of a neural network, enabling efficient image reconstruction.

[0004] Existing deep learning-based MRI reconstruction methods often focus solely on local information within the reconstructed image, failing to fully utilize similar information in the global space. This limits their effectiveness and leads to redundant learning parameters. Furthermore, most current methods only process images at their initial resolution and fail to utilize information at different spatial scales, which can result in missing or blurred image details. Effective technical solutions are urgently needed to address these issues. Summary of the Invention

[0005] The purpose of the present invention is to provide an MRI reconstruction method, system and storage medium based on multi-scale soft threshold operation, which extracts some important local and global information of the image in each reconstruction iterative block as auxiliary features and passes them into the next reconstruction iterative block as a priori. At the same time, in each iterative block, an encoder-decoder architecture is used to extract the multi-scale spatial features of the reconstructed image to further improve the reconstruction effect and achieve high-quality MRI image reconstruction.

[0006] A first aspect of the present invention provides an MRI reconstruction method based on a multi-scale soft threshold operation, comprising the following steps:

[0007] Acquire full-sampled MRI data and perform simulated undersampling to obtain simulated undersampling data;

[0008] Inputting the simulated under-sampled data into a reconstruction network for iterative reconstruction, which includes initializing the reconstruction network to input the simulated under-sampled data; and

[0009] Perform iterative reconstruction, including extraction of omnidirectional features, auxiliary feature transfer and multi-scale feature fusion, as well as loss calculation and model optimization;

[0010] After a preset number of iterative reconstructions, the final reconstructed image is output.

[0011] In this solution, the acquisition of fully sampled MRI data and the simulated undersampling to obtain the simulated undersampling data specifically includes:

[0012] Acquire fully sampled MRI data, specifically complete k-space data as the basis for training and testing;

[0013] The fully sampled MRI data is simulated under-sampled based on a preset sampling method to generate the simulated under-sampled data, wherein the sampling method includes at least one of random sampling and Cartesian sampling.

[0014] In this solution, the initialization of the reconstruction network to input the simulated under-sampling data specifically includes:

[0015] Initialize the preset step size ρ and regularization parameter λ of the gradient descent update module;

[0016] Initialize convolution weights, soft threshold parameters, and other learnable parameters;

[0017] The simulated under-sampled data is input into the reconstruction network to initialize the reconstructed image.

[0018] In this solution, the omnidirectional features extracted during iterative reconstruction specifically include:

[0019] After initializing the reconstruction network and inputting the simulated under-sampled data, performing local feature extraction based on a preset preliminary feature extraction block, including extracting convolution features; and

[0020] Global feature extraction is performed based on a preset global feature extraction module, which includes downsampling to obtain low-frequency features, and enhancing global features based on variance modulation to perform feature aggregation to obtain the global features.

[0021] In this solution, during iterative reconstruction, auxiliary feature transfer specifically includes:

[0022] Extract auxiliary features before soft thresholding

[0023] Extract auxiliary features after soft thresholding

[0024] Pass auxiliary features z through cascade and convolutional layers k , in, Represents the concatenation of two inputs.

[0025] In this solution, multi-scale feature fusion specifically includes:

[0026] The multi-scale reconstruction block based on the encoder-decoder structure extracts multi-scale features. The formula of the encoding stage is as follows:

[0027]

[0028] Among them, Down represents the downsampling operation, and Conv represents the convolution operation on the downsampled features. Represents input data, and Represent the results of the first downsampling and the second downsampling respectively;

[0029] The extracted multi-scale features are scaled and concatenated, and a sparse representation of the image is obtained through soft thresholding. The image is then downsampled twice as a skip connection to fuse the features of the decoder stage.

[0030] In the decoder stage, the encoder output is skip-connected with the sparse representation and then upsampled twice to obtain the final output fusion features.

[0031] A second aspect of the present invention further provides an MRI reconstruction system based on a multi-scale soft threshold operation, comprising a memory and a processor, wherein the memory includes an MRI reconstruction method program based on a multi-scale soft threshold operation, and when the MRI reconstruction method program based on a multi-scale soft threshold operation is executed by the processor, the following steps are implemented:

[0032] Acquire full-sampled MRI data and perform simulated undersampling to obtain simulated undersampling data;

[0033] Inputting the simulated under-sampled data into a reconstruction network for iterative reconstruction, which includes initializing the reconstruction network to input the simulated under-sampled data; and

[0034] Perform iterative reconstruction, including extraction of omnidirectional features, auxiliary feature transfer and multi-scale feature fusion, as well as loss calculation and model optimization;

[0035] After a preset number of iterative reconstructions, the final reconstructed image is output.

[0036] In this solution, the acquisition of fully sampled MRI data and the simulated undersampling to obtain the simulated undersampling data specifically includes:

[0037] Acquire fully sampled MRI data, specifically complete k-space data as the basis for training and testing;

[0038] The fully sampled MRI data is simulated under-sampled based on a preset sampling method to generate the simulated under-sampled data, wherein the sampling method includes at least one of random sampling and Cartesian sampling.

[0039] In this solution, the initialization of the reconstruction network to input the simulated under-sampling data specifically includes:

[0040] Initialize the preset step size ρ and regularization parameter λ of the gradient descent update module;

[0041] Initialize convolution weights, soft threshold parameters, and other learnable parameters;

[0042] The simulated under-sampled data is input into the reconstruction network to initialize the reconstructed image.

[0043] In this solution, the omnidirectional features extracted during iterative reconstruction specifically include:

[0044] After initializing the reconstruction network and inputting the simulated under-sampled data, performing local feature extraction based on a preset preliminary feature extraction block, including extracting convolution features; and

[0045] Global feature extraction is performed based on a preset global feature extraction module, which includes downsampling to obtain low-frequency features, and enhancing global features based on variance modulation to perform feature aggregation to obtain the global features.

[0046] In this solution, during iterative reconstruction, auxiliary feature transfer specifically includes:

[0047] Extract auxiliary features before soft thresholding

[0048] Extract auxiliary features after soft thresholding

[0049] Pass auxiliary features z through cascade and convolutional layers k , in, Represents the concatenation of two inputs.

[0050] In this solution, multi-scale feature fusion specifically includes:

[0051] The multi-scale reconstruction block based on the encoder-decoder structure extracts multi-scale features. The formula of the encoding stage is as follows:

[0052]

[0053] Among them, Down represents the downsampling operation, and Conv represents the convolution operation on the downsampled features. Represents input data, and Represent the results of the first downsampling and the second downsampling respectively;

[0054] The extracted multi-scale features are scaled and concatenated, and a sparse representation of the image is obtained through soft thresholding. The image is then downsampled twice as a skip connection to fuse the features of the decoder stage.

[0055] In the decoder stage, the encoder output is skip-connected with the sparse representation and then upsampled twice to obtain the final output fusion features.

[0056] The third aspect of the present invention provides a computer-readable storage medium, which includes a machine-readable MRI reconstruction method program based on multi-scale soft threshold operation. When the MRI reconstruction method program based on multi-scale soft threshold operation is executed by a processor, the steps of the MRI reconstruction method based on multi-scale soft threshold operation as described in any one of the above items are implemented.

[0057] The present invention discloses an MRI reconstruction method, system, and storage medium based on multiscale soft thresholding. By employing multiscale feature extraction, local and global feature fusion, inter-iteration feature transfer, and improved soft thresholding, the method effectively addresses key challenges in existing MRI reconstruction methods. Compared to existing technologies, the method maintains high-quality reconstruction even at high acceleration factors; significantly reduces image blur and detail loss; better suppresses noise and artifacts; improves the clarity and diagnostic value of reconstructed images; and maintains high computational efficiency, making it suitable for clinical applications. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] Figure 1A flowchart of an MRI reconstruction method based on multi-scale soft thresholding operation according to the present invention is shown;

[0059] Figure 2 A block diagram of an MRI reconstruction system based on multi-scale soft threshold operation according to the present invention is shown. DETAILED DESCRIPTION

[0060] In order to more clearly understand the above-mentioned objects, features and advantages of the present invention, the present invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments. It should be noted that, in the absence of conflict, the embodiments of the present application and the features therein can be combined with each other.

[0061] In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention may also be implemented in other ways different from those described herein. Therefore, the scope of protection of the present invention is not limited to the specific embodiments disclosed below.

[0062] Existing MRI reconstruction methods make insufficient use of global information: traditional deep learning methods mainly focus on local features and fail to fully utilize similar information in the global space of the image, which limits the reconstruction effect; and insufficient use of multi-scale features: most existing methods only process images at the initial resolution and fail to effectively utilize information at different spatial scales, resulting in missing or blurred image details; and feature information is lost during the iterative process: during the iterative process of reconstruction, some potentially useful intermediate features are not retained and passed to subsequent iterations, affecting the reconstruction quality; and reconstruction quality degrades under high acceleration factors: under high acceleration factors, existing methods are prone to problems such as noise enhancement, residual artifacts or image blurring. Among them, the defect of some Fourier reconstruction techniques is insufficient reconstruction effect when facing high-resolution and complex structure images; the defect of parallel imaging technology is noise enhancement and residual artifact problems under high acceleration factors; the defect of compressed sensing technology is image blurring and loss of details under high acceleration factors; and the defect of traditional deep learning methods is that they only focus on local information, parameter redundancy, and lack multi-scale processing.

[0063] Figure 1 The flowchart of the MRI reconstruction method based on multi-scale soft threshold operation of the present application is shown.

[0064] like Figure 1 As shown, the present application discloses an MRI reconstruction method based on multi-scale soft threshold operation, comprising the following steps:

[0065] S102, acquiring fully sampled MRI data and performing simulated undersampling to obtain simulated undersampling data;

[0066] S104, inputting the simulated under-sampled data into a reconstruction network for iterative reconstruction, which includes initializing the reconstruction network to input the simulated under-sampled data;

[0067] S106, performing iterative reconstruction, specifically including extracting omnidirectional features, auxiliary feature transfer and multi-scale feature fusion, as well as loss calculation and model optimization;

[0068] S108 , after a preset number of iterative reconstructions, output a final reconstructed image.

[0069] It should be noted that, in this embodiment, a fully sampled MRI image is first obtained and simulated undersampling preprocessing is performed to obtain a simulated undersampling MRI image; then, all-round feature extraction is performed on the image through local and global feature extraction modules, and the extracted features are input into a network with an encoder and decoder structure for multi-scale feature extraction and fusion. The soft threshold processing can improve the sparsity of the reconstructed image in the transformation domain. In addition, in order to retain and utilize some potentially useful features in each reconstruction iteration block, the intermediate features are obtained as auxiliary features before and after the soft threshold processing and transmitted to the next reconstruction iteration block. Subsequently, the MRI reconstruction model loss function of the multi-scale soft threshold operation is constructed, and the model parameters are fine-tuned to make the model accurately adapt to the reconstruction requirements. Finally, the undersampled MRI image is input into the reconstruction model to obtain a high-quality and fine reconstructed image.

[0070] Specifically, in this embodiment, reconstruction is based on an optimized compressed sensing algorithm: Given a linear measurement y, traditional image CS (Compressed Sensing) methods typically reconstruct the original image x by solving the following (usually convex) optimization problem: Where Φ is the measurement matrix, x is the image to be reconstructed, y is the undersampled k-space data, and Ψ X Represents the transformation coefficient of x about a certain transformation Ψ, vector Ψ X The sparsity of norm guarantees, λ is a (usually predefined) regularization parameter, further, where the iterative shrinkage-thresholding algorithm (ISTA) is a popular iterative algorithm for solving sparse optimization problems, which is very suitable for solving many large-scale linear inverse problems. Specifically, ISTA solves the optimization problem by iterating between the following update steps to reconstruct the CS reconstruction problem in the original image x formula, Formula 1: Formula 2: Among them, r (k)is the gradient output, Φ is the measurement matrix, x is the image to be reconstructed, y is the undersampled k-space data, k is the number of ISTA iterations, ρ is the step size, and Formula 2 is actually a special case of the so-called proximal mapping, that is, when Φ(x)=||Ψx||1, prox λΦ (r (k) ), formally, the proximal mapping prox of the regularizer Φ λφ (r) is defined as:

[0071] Furthermore, in this embodiment, in each stage of the reconstruction module, a gradient descent update module (GDU) method is used as the beginning of the network. This method involves the calculation of gradients related to the data fidelity term. Then, a convolution-guided multiscale module (CMM) is designed as a proximal mapping network to map the output of the GDU. Mathematically, the GDU and CMM processes of the kth stage are iteratively solved by the above formula to obtain the final MRI reconstructed image. The entire CMM is centered on soft thresholding and illustrates the nonlinear solution of proximal mapping for sparse correlation problems with complex changes. In the kth stage, the CMM transforms the output of the GDU r (k) As input, it first passes through the preliminary feature extraction block (PFEB, Primary Feature Extraction Block) for local feature extraction, then passes through the global feature extraction module (EASA, Enhanced Attention-based Spatial Aggregation Module) for global feature extraction, and then inputs the extracted features into the multi-scale soft threshold module (SUnet) for multi-scale feature decomposition. The obtained multi-scale features are fused and then soft thresholded to obtain a sparse representation of the image. Finally, this sparse representation is reconstructed through a network symmetrical to the one before soft thresholding to obtain the final MRI reconstructed image x k In addition, in order to enhance the effective use of the features of the previous iterative stage, an auxiliary iterative reconstruction block (AIRB) is designed in each iterative stage of CMM to receive the output z from the previous iterative stage. k-1 As input, it assists the image reconstruction of the current stage and generates z k It is transferred to the next iteration stage as output.

[0072] According to an embodiment of the present invention, acquiring fully sampled MRI data and performing simulated undersampling to obtain simulated undersampling data specifically includes:

[0073] Acquire fully sampled MRI data, specifically complete k-space data as the basis for training and testing;

[0074] The fully sampled MRI data is simulated under-sampled based on a preset sampling method to generate the simulated under-sampled data, wherein the sampling method includes at least one of random sampling and Cartesian sampling.

[0075] It should be noted that, in this embodiment, an MRI device is used to collect complete k-space data y as the basic data for training and testing, and undersampling simulation is performed. Specifically, the fully sampled MRI data is simulated undersampled to generate the simulated undersampled data to simulate the actual accelerated acquisition scenario, wherein the sampling method includes at least one of random sampling and Cartesian sampling.

[0076] According to an embodiment of the present invention, initializing the reconstruction network to input the simulated under-sampled data specifically includes:

[0077] Initialize the preset step size ρ and regularization parameter λ of the gradient descent update module;

[0078] Initialize convolution weights, soft threshold parameters, and other learnable parameters;

[0079] The simulated under-sampled data is input into the reconstruction network to initialize the reconstructed image.

[0080] It should be noted that, in this embodiment, the step size ρ and regularization parameter λ of the gradient descent update module (GDU), as well as the convolution weights, soft threshold parameters and other learnable parameters are initialized, and the simulated under-sampling data is input into the reconstruction network to initialize the reconstructed image, wherein the specific process of the gradient descent update module (GDU) has been described in the above embodiment.

[0081] According to an embodiment of the present invention, in iterative reconstruction, extracting omnidirectional features specifically includes:

[0082] After initializing the reconstruction network and inputting the simulated under-sampled data, performing local feature extraction based on a preset preliminary feature extraction block, including extracting convolution features; and

[0083] Global feature extraction is performed based on a preset global feature extraction module, which includes downsampling to obtain low-frequency features, and enhancing global features based on variance modulation to perform feature aggregation to obtain the global features.

[0084] It should be noted that in this embodiment, the CMM starts from the preliminary feature extraction block (PFEB), which helps to deepen the extraction of local features. First, a convolution operation Conv(·) is performed to convert the initial single-channel image input into a multi-channel format containing 32 channels. Then, the leaky ReLU activation function LReLU(·) is applied, and then a channel attention mechanism C(·) is integrated to finely calibrate the importance of each dimension. This calibration process has a dual purpose of improving feature representation and reducing redundant information.

[0085] Specifically, the above process can be mathematically formalized as follows: in represents the PFEB operation, and Represents the output of the module, r k The output of the gradient descent update module GDU is used to capture long-distance dependencies. The fully local feature extraction block (EASA) is introduced to obtain low-frequency components through downsampling operations and feed them into a 3×3 depth convolution to generate global structural information. in, Represents adaptive max pooling with a scaling factor of 8, DWConv 3×3 (·) is a 3×3 depth convolution layer to embed the modulated global representation A global description of The variance of is used as the statistics of spatial information and is combined with the global representation through 1×1 convolution merge: μ) 2 ; in, yes The variance of N is the total number of pixels, x i represents the value of each pixel, μ is the mean of all pixel values, Represents the modulation feature. This variance modulation mechanism helps to better explore the global information. Finally, the modulation feature is used to aggregate the input features. To extract representative structural information

[0086] According to an embodiment of the present invention, in iterative reconstruction, auxiliary feature transfer specifically includes:

[0087] Extract auxiliary features before soft thresholding

[0088] Extract auxiliary features z after soft threshold k2 ;

[0089] Pass auxiliary features z through cascade and convolutional layers k , in, Represents the concatenation of two inputs.

[0090] It should be noted that in this embodiment, two auxiliary iterative reconstruction blocks (AIRB) are customized, one before and one after the soft threshold operation, to retain a wider range of multi-channel information and enhance the network's modeling ability for complex data. and z k-1 As input, AIRB produces two outputs, namely and z k1 represents the output of AIRB before soft threshold operation, and z k2 represents the output of AIRB after the soft threshold operation, namely: in, Represents the concatenation of two inputs.

[0091] According to an embodiment of the present invention, multi-scale feature fusion specifically includes:

[0092] The multi-scale reconstruction block based on the encoder-decoder structure extracts multi-scale features. The formula of the encoding stage is as follows:

[0093]

[0094] Among them, Down represents the downsampling operation, and Conv represents the convolution operation on the downsampled features. Represents input data, and Represent the results of the first downsampling and the second downsampling respectively;

[0095] The extracted multi-scale features are scaled and concatenated, and a sparse representation of the image is obtained through soft thresholding. The image is then downsampled twice as a skip connection to fuse the features of the decoder stage.

[0096] In the decoder stage, the encoder output is skip-connected with the sparse representation and then upsampled twice to obtain the final output fusion features.

[0097] It should be noted that in this embodiment, unlike the traditional single-scale soft threshold operation, this embodiment uses a multi-scale reconstruction block (MTRB) based on an encoder-decoder structure to extract multi-scale features, and then applies soft threshold processing after fusing these multi-scale features to improve sparsity.

[0098] Specifically, at the encoder stage, the input Perform two downsampling to obtain multi-scale features. The formula is as follows:

[0099]

[0100] Among them, Down represents the downsampling operation, and Conv represents the convolution operation on the downsampled features. and They represent the results of the first downsampling and the second downsampling respectively. Furthermore, before performing soft threshold processing, the features of the three scales are first resized. The formula is as follows: Interpolate means upsampling by interpolation operation, then fusing the features after unification of the three scales, and performing soft thresholding on the fused multi-scale features to obtain a sparse representation of the image. The formula is as follows: soft represents the soft threshold operation, θ is the shrinkage threshold, which is also a learnable parameter in this module, and Conv_fu represents the fused convolution, which is used to transform the spliced ​​multi-scale features into a shape that conforms to the model input. After the soft threshold processing, the feature is downsampled twice to obtain features of two scales, which are fused with the features of the decoder stage as a skip connection. The formula is as follows: Interpolate means downsampling by interpolation operation. In the decoder stage, the output of the encoder is and After performing the jump connection, upsampling is performed and the result of the first upsampling is compared with After performing the jump connection, upsampling is performed. The formula is as follows:

[0101]

[0102] Among them, Up represents the upsampling operation, and Conv represents the convolution operation on the upsampled features. and Represent the results of the first upsampling and the second upsampling respectively. Finally, and After performing the jump connection, convolution processing is performed to obtain the final output. The formula is as follows: in, Represents the final output of MTRB, and Conv represents the convolution operation.

[0103] It is worth mentioning that during execution, this process completely reverses the processing taken before the soft threshold operation to ensure the symmetry of the network structure before and after the soft threshold processing. The formula is as follows: in, represents the network before soft thresholding, Represents the network after soft threshold processing, and finally, with the input r (k) Establish a residual connection to produce the output x of the kth stage (k) .

[0104] It is worth mentioning that the mean square error (MSE) is used as the loss function, the original image x is used as the true image label, and the reconstructed image is obtained from the measurement y corresponding to x. As the output of the network, the loss function is formulated as: in, represents the aggregate count of training images, x (i) represents the i-th trainable image.

[0105] Figure 2 A block diagram of an MRI reconstruction system based on multi-scale soft threshold operation according to the present invention is shown.

[0106] like Figure 2 As shown, the present invention discloses an MRI reconstruction system based on multi-scale soft threshold operation, including a memory and a processor. The memory includes an MRI reconstruction method program based on multi-scale soft threshold operation. When the MRI reconstruction method program based on multi-scale soft threshold operation is executed by the processor, the following steps are implemented:

[0107] Acquire full-sampled MRI data and perform simulated undersampling to obtain simulated undersampling data;

[0108] Inputting the simulated under-sampled data into a reconstruction network for iterative reconstruction, which includes initializing the reconstruction network to input the simulated under-sampled data; and

[0109] Perform iterative reconstruction, including extraction of omnidirectional features, auxiliary feature transfer and multi-scale feature fusion, as well as loss calculation and model optimization;

[0110] After a preset number of iterative reconstructions, the final reconstructed image is output.

[0111] It should be noted that, in this embodiment, a fully sampled MRI image is first obtained and simulated undersampling preprocessing is performed to obtain a simulated undersampling MRI image; then, all-round feature extraction is performed on the image through local and global feature extraction modules, and the extracted features are input into a network with an encoder and decoder structure for multi-scale feature extraction and fusion. The soft threshold processing can improve the sparsity of the reconstructed image in the transformation domain. In addition, in order to retain and utilize some potentially useful features in each reconstruction iteration block, the intermediate features are obtained as auxiliary features before and after the soft threshold processing and transmitted to the next reconstruction iteration block. Subsequently, the MRI reconstruction model loss function of the multi-scale soft threshold operation is constructed, and the model parameters are fine-tuned to make the model accurately adapt to the reconstruction requirements. Finally, the undersampled MRI image is input into the reconstruction model to obtain a high-quality and fine reconstructed image.

[0112] Specifically, in this embodiment, reconstruction is based on an optimized compressed sensing algorithm: Given a linear measurement y, traditional image CS (Compressed Sensing) methods typically reconstruct the original image x by solving the following (usually convex) optimization problem: Where Φ is the measurement matrix, x is the image to be reconstructed, y is the undersampled k-space data, and Ψ X Represents the transformation coefficient of x about a certain transformation Ψ, vector Ψ X The sparsity of norm guarantees, λ is a (usually predefined) regularization parameter, further, where the iterative shrinkage-thresholding algorithm (ISTA) is a popular iterative algorithm for solving sparse optimization problems, which is very suitable for solving many large-scale linear inverse problems. Specifically, ISTA solves the optimization problem by iterating between the following update steps to reconstruct the CS reconstruction problem in the original image x formula, Formula 1: Formula 2: Among them, r (k) is the gradient output, Φ is the measurement matrix, x is the image to be reconstructed, y is the undersampled k-space data, k is the number of ISTA iterations, ρ is the step size, and Formula 2 is actually a special case of the so-called proximal mapping, that is, when Φ(x)=||Ψx||1, prox λΦ (r (k) ), formally, the proximal mapping prox of the regularizer Φ λφ (r) is defined as:

[0113] Furthermore, in this embodiment, in each stage of the reconstruction module, a gradient descent update module (GDU) method is used as the beginning of the network. This method involves the calculation of gradients related to the data fidelity term. Then, a convolution-guided multiscale module (CMM) is designed as a proximal mapping network to map the output of the GDU. Mathematically, the GDU and CMM processes of the kth stage are iteratively solved by the above formula to obtain the final MRI reconstructed image. The entire CMM is centered on soft thresholding and illustrates the nonlinear solution of proximal mapping for sparse correlation problems with complex changes. In the kth stage, the CMM transforms the output of the GDU r (k) As input, it first passes through the preliminary feature extraction block (PFEB, Primary Feature Extraction Block) for local feature extraction, then passes through the global feature extraction module (EASA, Enhanced Attention-based Spatial Aggregation Module) for global feature extraction, and then inputs the extracted features into the multi-scale soft threshold module (SUnet) for multi-scale feature decomposition. The obtained multi-scale features are fused and then soft thresholded to obtain a sparse representation of the image. Finally, this sparse representation is reconstructed through a network symmetrical to the one before soft thresholding to obtain the final MRI reconstructed image x k In addition, in order to enhance the effective use of the features of the previous iterative stage, an auxiliary iterative reconstruction block (AIRB) is designed in each iterative stage of CMM to receive the output z from the previous iterative stage. k-1 As input, it assists the image reconstruction of the current stage and generates z k It is transferred to the next iteration stage as output.

[0114] According to an embodiment of the present invention, acquiring fully sampled MRI data and performing simulated undersampling to obtain simulated undersampling data specifically includes:

[0115] Acquire fully sampled MRI data, specifically complete k-space data as the basis for training and testing;

[0116] The fully sampled MRI data is simulated under-sampled based on a preset sampling method to generate the simulated under-sampled data, wherein the sampling method includes at least one of random sampling and Cartesian sampling.

[0117] It should be noted that, in this embodiment, an MRI device is used to collect complete k-space data y as the basic data for training and testing, and undersampling simulation is performed. Specifically, the fully sampled MRI data is simulated undersampled to generate the simulated undersampled data to simulate the actual accelerated acquisition scenario, wherein the sampling method includes at least one of random sampling and Cartesian sampling.

[0118] According to an embodiment of the present invention, initializing the reconstruction network to input the simulated under-sampled data specifically includes:

[0119] Initialize the preset step size ρ and regularization parameter λ of the gradient descent update module;

[0120] Initialize convolution weights, soft threshold parameters, and other learnable parameters;

[0121] The simulated under-sampled data is input into the reconstruction network to initialize the reconstructed image.

[0122] It should be noted that, in this embodiment, the step size ρ and regularization parameter λ of the gradient descent update module (GDU), as well as the convolution weights, soft threshold parameters and other learnable parameters are initialized, and the simulated under-sampling data is input into the reconstruction network to initialize the reconstructed image, wherein the specific process of the gradient descent update module (GDU) has been described in the above embodiment.

[0123] According to an embodiment of the present invention, in iterative reconstruction, extracting omnidirectional features specifically includes:

[0124] After initializing the reconstruction network and inputting the simulated under-sampled data, performing local feature extraction based on a preset preliminary feature extraction block, including extracting convolution features; and

[0125] Global feature extraction is performed based on a preset global feature extraction module, which includes downsampling to obtain low-frequency features, and enhancing global features based on variance modulation to perform feature aggregation to obtain the global features.

[0126] It should be noted that in this embodiment, the CMM starts from the preliminary feature extraction block (PFEB), which helps to deepen the extraction of local features. First, a convolution operation Conv(·) is performed to convert the initial single-channel image input into a multi-channel format containing 32 channels. Then, the leaky ReLU activation function LReLU(·) is applied, and then a channel attention mechanism C(·) is integrated to finely calibrate the importance of each dimension. This calibration process has a dual purpose of improving feature representation and reducing redundant information.

[0127] Specifically, the above process can be mathematically formalized as follows: in represents the PFEB operation, and Represents the output of the module, r k The output of the gradient descent update module GDU is used to capture long-distance dependencies. The fully local feature extraction block (EASA) is introduced to obtain low-frequency components through downsampling operations and feed them into a 3×3 depth convolution to generate global structural information. in, Represents adaptive max pooling with a scaling factor of 8, DWConv 3×3 (·) is a 3×3 depth convolution layer to embed the modulated global representation A global description of The variance of is used as the statistics of spatial information and is combined with the global representation through 1×1 convolution merge: in, yes The variance of N is the total number of pixels, x i represents the value of each pixel, μ is the mean of all pixel values, Represents the modulation feature. This variance modulation mechanism helps to better explore the global information. Finally, the modulation feature is used to aggregate the input features. To extract representative structural information

[0128] According to an embodiment of the present invention, in iterative reconstruction, auxiliary feature transfer specifically includes:

[0129] Extract auxiliary features before soft thresholding

[0130] Extract auxiliary features after soft thresholding

[0131] Pass auxiliary features z through cascade and convolutional layers k , in, Represents the concatenation of two inputs.

[0132] It should be noted that in this embodiment, two auxiliary iterative reconstruction blocks (AIRB) are customized, one before and one after the soft threshold operation, to retain a wider range of multi-channel information and enhance the network's modeling ability for complex data. and z k-1 As input, AIRB produces two outputs, namely and z k1 represents the output of AIRB before soft threshold operation, and zk2 represents the output of AIRB after the soft threshold operation, namely: in, Represents the concatenation of two inputs.

[0133] According to an embodiment of the present invention, multi-scale feature fusion specifically includes:

[0134] The multi-scale reconstruction block based on the encoder-decoder structure extracts multi-scale features. The formula of the encoding stage is as follows:

[0135]

[0136] Among them, Down represents the downsampling operation, and Conv represents the convolution operation on the downsampled features. Represents input data, and Represent the results of the first downsampling and the second downsampling respectively;

[0137] The extracted multi-scale features are scaled and concatenated, and a sparse representation of the image is obtained through soft thresholding. The image is then downsampled twice as a skip connection to fuse the features of the decoder stage.

[0138] In the decoder stage, the encoder output is skip-connected with the sparse representation and then upsampled twice to obtain the final output fusion features.

[0139] It should be noted that in this embodiment, unlike the traditional single-scale soft threshold operation, this embodiment uses a multi-scale reconstruction block (MTRB) based on an encoder-decoder structure to extract multi-scale features, and then applies soft threshold processing after fusing these multi-scale features to improve sparsity.

[0140] Specifically, at the encoder stage, the input Perform two downsampling to obtain multi-scale features. The formula is as follows:

[0141]

[0142] Among them, Down represents the downsampling operation, and Conv represents the convolution operation on the downsampled features. and They represent the results of the first downsampling and the second downsampling respectively. Furthermore, before performing soft threshold processing, the features of the three scales are first resized. The formula is as follows: Interpolate means upsampling by interpolation operation, then fusing the features after unification of the three scales, and performing soft thresholding on the fused multi-scale features to obtain a sparse representation of the image. The formula is as follows: soft represents the soft threshold operation, θ is the shrinkage threshold, which is also a learnable parameter in this module, and Conv_fu represents the fused convolution, which is used to transform the spliced ​​multi-scale features into a shape that conforms to the model input. After the soft threshold processing, the feature is downsampled twice to obtain features of two scales, which are fused with the features of the decoder stage as a skip connection. The formula is as follows: Interpolate means downsampling by interpolation operation. In the decoder stage, the output of the encoder is and After performing the jump connection, upsampling is performed and the result of the first upsampling is compared with After performing the jump connection, upsampling is performed. The formula is as follows:

[0143]

[0144] Among them, Up represents the upsampling operation, and Conv represents the convolution operation on the upsampled features. and Represent the results of the first upsampling and the second upsampling respectively. Finally, and After performing the jump connection, convolution processing is performed to obtain the final output. The formula is as follows: in, Represents the final output of MTRB, and Conv represents the convolution operation.

[0145] It is worth mentioning that during execution, this process completely reverses the processing taken before the soft threshold operation to ensure the symmetry of the network structure before and after the soft threshold processing. The formula is as follows: in, represents the network before soft thresholding, Represents the network after soft threshold processing, and finally, with the input r (k) Establish a residual connection to produce the output x of the kth stage (k) .

[0146] It is worth mentioning that the mean square error (MSE) is used as the loss function, the original image x is used as the true image label, and the reconstructed image is obtained from the measurement y corresponding to x. As the output of the network, the loss function is formulated as: in, represents the aggregate count of training images, x (i) represents the i-th trainable image.

[0147] A third aspect of the present invention provides a computer-readable storage medium, which includes an MRI reconstruction method program based on multi-scale soft threshold operation. When the MRI reconstruction method program based on multi-scale soft threshold operation is executed by a processor, the steps of an MRI reconstruction method based on multi-scale soft threshold operation as described in any one of the above items are implemented.

[0148] The present invention discloses an MRI reconstruction method, system, and storage medium based on multiscale soft thresholding. By employing multiscale feature extraction, local and global feature fusion, inter-iteration feature transfer, and improved soft thresholding, the method effectively addresses key challenges in existing MRI reconstruction methods. Compared to existing technologies, the method maintains high-quality reconstruction even at high acceleration factors; significantly reduces image blur and detail loss; better suppresses noise and artifacts; improves the clarity and diagnostic value of reconstructed images; and maintains high computational efficiency, making it suitable for clinical applications.

[0149] In the several embodiments provided in this application, it should be understood that the disclosed devices and methods can be implemented in other ways. The device embodiments described above are merely schematic. For example, the division of the units is merely a logical function division. In actual implementation, there may be other division methods, such as: multiple units or components can be combined, or can be integrated into another system, or some features can be ignored or not executed. In addition, the coupling, direct coupling, or communication connection between the components shown or discussed can be through some interfaces, and the indirect coupling or communication connection of the devices or units can be electrical, mechanical or other forms.

[0150] The units described above as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units; they may be located in one place or distributed across multiple network units; some or all of the units may be selected according to actual needs to achieve the purpose of the scheme of this embodiment.

[0151] In addition, all functional units in the embodiments of the present invention may be integrated into one processing unit, or each unit may be separately used as a unit, or two or more units may be integrated into one unit; the above-mentioned integrated units may be implemented in the form of hardware or in the form of hardware plus software functional units.

[0152] Those skilled in the art will appreciate that all or part of the steps of the above-mentioned method embodiments may be implemented by hardware associated with program instructions, and the aforementioned program may be stored in a computer-readable storage medium. When the program is executed, the program executes the steps of the above-mentioned method embodiments. The aforementioned storage medium includes various media that can store program codes, such as mobile storage devices, read-only memories (ROMs), random access memories (RAMs), magnetic disks, or optical disks.

[0153] Alternatively, if the above-mentioned integrated unit of the present invention is implemented in the form of a software functional module and sold or used as an independent product, it can also be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the embodiment of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes a number of instructions for enabling a computer device (which can be a personal computer, server, or network device, etc.) to execute all or part of the methods described in each embodiment of the present invention. The aforementioned storage medium includes: various media that can store program codes, such as mobile storage devices, ROM, RAM, magnetic disks or optical disks.

Claims

1. An MRI reconstruction method based on multi-scale soft thresholding operation, characterized in that: The following steps are involved: Acquire full-sampled MRI data and perform simulated undersampling to obtain simulated undersampling data; Inputting the simulated under-sampled data into a reconstruction network for iterative reconstruction, which includes initializing the reconstruction network to input the simulated under-sampled data; and Perform iterative reconstruction, including extraction of omnidirectional features, auxiliary feature transfer and multi-scale feature fusion, as well as loss calculation and model optimization; After a preset number of iterative reconstructions, the final reconstructed image is output.

2. The MRI reconstruction method based on multi-scale soft thresholding according to claim 1, characterized in that: The step of acquiring the fully sampled MRI data and performing simulated undersampling to obtain the simulated undersampling data specifically includes: Acquire fully sampled MRI data, specifically complete k-space data as the basis for training and testing; The fully sampled MRI data is simulated under-sampled based on a preset sampling method to generate the simulated under-sampled data, wherein the sampling method includes at least one of random sampling and Cartesian sampling.

3. The MRI reconstruction method based on multi-scale soft thresholding according to claim 1, characterized in that: Initializing the reconstruction network to input the simulated under-sampling data specifically includes: Initialize the preset step size ρ and regularization parameter λ of the gradient descent update module; Initialize convolution weights, soft threshold parameters, and other learnable parameters; The simulated under-sampled data is input into the reconstruction network to initialize the reconstructed image.

4. The MRI reconstruction method based on multi-scale soft thresholding according to claim 1, characterized in that: In iterative reconstruction, the extraction of all-round features specifically includes: After initializing the reconstruction network and inputting the simulated under-sampled data, performing local feature extraction based on a preset preliminary feature extraction block, including extracting convolution features; and Global feature extraction is performed based on a preset global feature extraction module, which includes downsampling to obtain low-frequency features, and enhancing global features based on variance modulation to perform feature aggregation to obtain the global features.

5. The MRI reconstruction method based on multi-scale soft thresholding according to claim 4, characterized in that: In iterative reconstruction, auxiliary feature transfer specifically includes: Extract auxiliary features before soft thresholding Extract auxiliary features after soft thresholding Pass auxiliary features z through cascade and convolutional layers k , in, Represents the concatenation of two inputs.

6. According to the MRI reconstruction method based on multi-scale soft thresholding operation in claim 5, in the iterative reconstruction, the multi-scale feature fusion specifically comprises: The multi-scale reconstruction block based on the encoder-decoder structure extracts multi-scale features. The formula of the encoding stage is as follows: Among them, Down represents the downsampling operation, and Conv represents the convolution operation on the downsampled features. Represents input data, and Represent the results of the first downsampling and the second downsampling respectively; The extracted multi-scale features are scaled and concatenated, and a sparse representation of the image is obtained through soft thresholding. The image is then downsampled twice as a skip connection to fuse the features of the decoder stage. In the decoder stage, the encoder output is skip-connected with the sparse representation and then upsampled twice to obtain the final output fusion features.

7. An MRI reconstruction system based on multi-scale soft threshold operation, characterized in that: The system comprises a memory and a processor, wherein the memory comprises an MRI reconstruction method program based on a multi-scale soft threshold operation, and when the MRI reconstruction method program based on a multi-scale soft threshold operation is executed by the processor, the following steps are implemented: Acquire full-sampled MRI data and perform simulated undersampling to obtain simulated undersampling data; Inputting the simulated under-sampled data into a reconstruction network for iterative reconstruction, which includes initializing the reconstruction network to input the simulated under-sampled data; and Perform iterative reconstruction, including extraction of omnidirectional features, auxiliary feature transfer and multi-scale feature fusion, as well as loss calculation and model optimization; After a preset number of iterative reconstructions, the final reconstructed image is output.

8. The MRI reconstruction system based on multi-scale soft thresholding operation according to claim 7, characterized in that: The step of acquiring the fully sampled MRI data and performing simulated undersampling to obtain the simulated undersampling data specifically includes: Acquire fully sampled MRI data, specifically complete k-space data as the basis for training and testing; The fully sampled MRI data is simulated under-sampled based on a preset sampling method to generate the simulated under-sampled data, wherein the sampling method includes at least one of random sampling and Cartesian sampling.

9. The MRI reconstruction system based on multi-scale soft thresholding operation according to claim 7, characterized in that: Initializing the reconstruction network to input the simulated under-sampling data specifically includes: Initialize the preset step size ρ and regularization parameter λ of the gradient descent update module; Initialize convolution weights, soft threshold parameters, and other learnable parameters; The simulated under-sampled data is input into the reconstruction network to initialize the reconstructed image.

10. A computer-readable storage medium, characterized in that The computer-readable storage medium includes an MRI reconstruction method program based on multi-scale soft threshold operation. When the MRI reconstruction method program based on multi-scale soft threshold operation is executed by a processor, the steps of an MRI reconstruction method based on multi-scale soft threshold operation as described in any one of claims 1 to 6 are implemented.

Images