High-speed MRI reconstruction method based on structured sparse and channel attention image prior
By combining an iterative network with data consistency, structured sparsity, and channel attention image priors, the problem of insufficient reconstruction quality in MRI reconstruction methods under high acceleration factors is solved, and efficient and reliable MRI image reconstruction is achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- NANJING MEDICAL UNIV
- Filing Date
- 2026-01-27
- Publication Date
- 2026-06-02
Smart Images
Figure CN122134880A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of medical image processing technology, specifically to a high-speed MRI reconstruction method based on structured sparsity and channel attention image prior. Background Technology
[0002] Magnetic Resonance Imaging (MRI) is a non-invasive medical imaging technique with unique advantages such as no ionizing radiation, excellent soft tissue resolution, and multi-parameter imaging. It has become an indispensable tool in modern medical diagnosis and is widely used in the diagnosis of diseases in the brain, cardiovascular system, and other areas. However, the slow data acquisition speed of traditional MRI severely limits its clinical application in scenarios such as dynamic imaging and functional MRI. To shorten scan time, K-space undersampling technology has become a core means of accelerating MRI. However, this technology transforms image reconstruction into a highly ill-posed inverse problem, meaning that recovering a fully sampled image from finite K-space data is mathematically inherently pathological.
[0003] Despite the development of K-space undersampling techniques, significant limitations remain. First, K-space undersampling leads to incomplete data, with the information loss becoming more severe under high acceleration factors, resulting in issues such as detail loss and artifact retention, leading to poor MRI reconstruction quality. Traditional methods struggle to effectively fill these information gaps. Second, compressed sensing MRI relies on manually designed sparse bases, which cannot adapt to complex and varied anatomical structures, and is prone to smoothing artifacts and detail loss under high acceleration. Third, end-to-end deep learning methods require a large amount of labeled data to train the model, relying on large-scale, high-quality paired data. However, acquiring high-quality paired MRI data in clinical settings is costly and time-consuming, leading to overfitting and limited generalization ability in scenarios with limited data. Finally, existing iterative unfolding network feature extraction modules are mostly general convolutions, failing to specifically mine the structured sparse features of MRI images. Furthermore, global thresholding operations cannot distinguish the importance of features in different channels, resulting in insufficient preservation of key anatomical structures and difficulty in adapting to complex anatomical structures. Summary of the Invention
[0004] To address the technical problem that existing MRI reconstruction methods cannot simultaneously achieve data efficiency and reconstruction quality, resulting in insufficient reliability of reconstructed MRI images, the present invention aims to provide a high-speed MRI reconstruction method based on structured sparsity and channel attention image prior. The specific technical solution adopted is as follows:
[0005] Acquire MRI image datasets, preprocess MRI image datasets, and generate source processed image datasets;
[0006] The design incorporates multiple basic blocks to construct an iterative network. Each basic block includes a data consistency module, a structured sparse module, and a channel attention deep image prior module.
[0007] The source image dataset is input into the iterative network, and then processed sequentially by the data consistency module, the structured sparsity module, and the channel attention deep image prior module to obtain the consistent image, the sparsity optimized image, and the structure optimized image, respectively.
[0008] The loss function is designed based on an iterative network, and after iterative training with multiple basic blocks, the MRI reconstructed image set is output.
[0009] Preferably, the MRI image dataset is preprocessed to generate the source processed image dataset, specifically as follows:
[0010] The MRI image dataset is converted to K-space by Fourier transform, undersampling is performed by applying a radial sampling mask to obtain undersampled K-space data, and then the source processed image dataset is generated by inverse Fourier transform.
[0011] Preferably, each basic block is connected in series to establish an iterative network.
[0012] Preferably, the source image dataset is input into the iterative network, and processed sequentially by the data consistency module, the structured sparsity module, and the channel attention deep image prior module, respectively obtaining a consistent image, a sparsity-optimized image, and a structure-optimized image, including:
[0013] Based on the data consistency module, a mathematical model for data consistency correction is established for the source image dataset, and a consistent image is output.
[0014] The structured sparse module is used to mine the inherent sparsity properties of consistent images and alleviate the loss of details under high speedup ratio, outputting a sparse-optimized image;
[0015] The channel attention deep image prior module learns the structural patterns of images through a data-driven approach, performs residual correction on sparse optimized images, and outputs structurally optimized images.
[0016] Preferably, a data consistency correction mathematical model is established on the source processed image dataset based on the data consistency module, and a consistent image is output, specifically as follows:
[0017] The data consistency module enables the source processed image dataset to be evaluated to meet the constraints of the measured K-space data, determines the correction direction, and, in conjunction with backpropagation dynamic optimization, establishes a mathematical model for data consistency correction, outputting a consistent image corresponding to the source processed image dataset.
[0018] Preferably, a structured sparse module is used to mine the inherent sparsity properties of the consistent image and alleviate the loss of detail under high speedup ratios, outputting a sparse-optimized image, including:
[0019] Sparse encoding is performed on the consistent image, which is then mapped to a high-dimensional sparse feature space. Sparse structural features are then mined through multi-layer convolution and non-linear activation.
[0020] Design a channel-level learnable threshold mechanism to perform soft thresholding on sparse structural features and apply differentiated sparsity constraints.
[0021] The thresholded sparse structural features are mapped back to the image domain, and sparse decoding is performed by combining residual connections with the consistent image to output a sparse optimized image.
[0022] Preferably, the channel attention depth image prior module learns the image structure rules through a data-driven approach, performs residual correction on the sparse optimized image, and outputs a structure-optimized image, including:
[0023] The channel attention depth image prior module includes a DIP network and a channel attention module;
[0024] The sparse optimized image is scaled using a DIP network and weighted by a channel attention module to output a source structure optimized image.
[0025] A globally learnable step size is introduced to perform residual correction on the sparse optimization image, and the output structure optimization image is combined with the source structure optimization image.
[0026] Preferably, the sparse optimized image is scale-optimized using a DIP network and feature-weighted using a channel attention module to output a source structure-optimized image, including:
[0027] Based on the sparse optimization image, the encoder performs downsampling to abstract global features, and the auxiliary channel attention module sequentially performs global information capture, weight learning and feature weighting to output a weighted feature image.
[0028] The weighted feature image is upsampled by the decoder to restore local details, and high- and low-dimensional information is fused with skip connections to output an optimized source structure image.
[0029] Preferably, the loss function is designed based on an iterative network, and after iterative training with multiple basic blocks, an MRI reconstructed image set is output, specifically:
[0030] A loss function is designed by integrating data consistency loss and symmetry loss using a balanced coefficient. The loss function is used to complete iterative training of multiple basic blocks, and the iterative network outputs a set of MRI reconstructed images.
[0031] To address the aforementioned problems, the present invention also provides an electronic device comprising a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other via the communication bus, and the processor invokes logical instructions in the memory to execute the high-speed MRI reconstruction method based on structured sparse and channel attention image prior as described in any of the preceding claims.
[0032] The present invention has the following beneficial effects:
[0033] 1. An iterative network is constructed based on multiple basic blocks. Through the cooperation of three modules, the source image dataset is processed sequentially. Consistency images are used to ensure data fidelity. Sparse optimization images utilize the structured sparsity characteristics of MRI images for analysis, to differentiate and preserve key anatomical structures and suppress noise. To address the problem of insufficient high-dimensional feature modeling with limited data, a channel attention deep image prior module is used to focus on information-rich regions, reduce artifacts, and obtain structure-optimized images. At the same time, the complementarity of structured sparse prior and deep prior is balanced to avoid feature conflicts and improve structural integrity and detail realism. Finally, to address the overfitting problem with limited data, a loss function is designed to enhance the generalization ability of the iterative network.
[0034] 2. The electronic device provided by this invention has the same beneficial effects as the high-speed MRI reconstruction method based on structured sparsity and channel attention image prior provided by this invention, and will not be described in detail here. Attached Figure Description
[0035] To more clearly illustrate the technical solutions and advantages in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0036] Figure 1 The flowchart illustrates an implementation of a high-speed MRI reconstruction method based on structured sparsity and channel attention image priors, as provided in one embodiment of the present invention.
[0037] Figure 2 A flowchart illustrating the steps of a high-speed MRI reconstruction method based on structured sparsity and channel attention image priors provided in an embodiment of the present invention.
[0038] Figure 3 This is a flowchart illustrating the implementation of each module in the basic block of the high-speed MRI reconstruction method based on structured sparsity and channel attention image priors provided in an embodiment of the present invention.
[0039] Figure 4 A comparison between the high-speed MRI reconstruction method based on structured sparsity and channel attention image prior provided in an embodiment of the present invention and existing reconstruction techniques. Figure 1 ;
[0040] Figure 5 A comparison between the high-speed MRI reconstruction method based on structured sparsity and channel attention image prior provided in an embodiment of the present invention and existing reconstruction techniques. Figure 2 ;
[0041] Figure 6 A comparison between the high-speed MRI reconstruction method based on structured sparsity and channel attention image prior provided in an embodiment of the present invention and existing reconstruction techniques. Figure 3 . Detailed Implementation
[0042] To further illustrate the technical means and effects adopted by the present invention to achieve its intended purpose, the following, in conjunction with the accompanying drawings and preferred embodiments, details the specific implementation, structure, features, and effects of a high-speed MRI reconstruction method based on structured sparsity and channel attention image priors proposed according to the present invention. In the following description, different "one embodiment" or "another embodiment" do not necessarily refer to the same embodiment. Furthermore, specific features, structures, or characteristics in one or more embodiments can be combined in any suitable form.
[0043] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.
[0044] The following description, in conjunction with the accompanying drawings, details a specific scheme for a high-speed MRI reconstruction method based on structured sparsity and channel attention image priors provided by this invention.
[0045] To better illustrate, MRI images, or magnetic resonance imaging (MRI) images, use strong magnetic fields and radio frequency pulses to perform non-invasive, high-resolution imaging of the internal tissue structures of the human body. Reconstruction is performed because the original acquired MRI images are usually frequency domain signals in k-space, which cannot be directly converted into visualized anatomical images. Therefore, reconstruction provides more reliable and informative imaging evidence for clinical practice.
[0046] Existing methods for reconstructing MRI images include compressed sensing MRI, deep learning MRI reconstruction, and reconstruction using iterative unfolding networks. Among these, compressed sensing MRI, or CS-MRI (Compressed Sensing Magnetic Resonance Imaging), is based on sparsity constraints and... The norm optimization problem is transformed into Norm convex optimization problems are solved using algorithms such as gradient descent, threshold shrinkage, and split Bregman iteration, but relying on manual sparse basis makes it difficult to capture complex anatomical structures, and the reconstruction quality is limited under high acceleration.
[0047] Deep learning-based MRI reconstruction includes architectures such as Convolutional Neural Networks (CNNs), U-Nets, ResNets, Generative Adversarial Networks (GANs), and Transformers. It adopts an end-to-end training mode, but it relies on large-scale, high-quality paired data. When data is scarce, its generalization ability is poor, and some deep learning methods have black box characteristics, which may produce illusory structures that violate physical constraints, affecting the reliability of diagnosis.
[0048] Reconstruction using iterative unfolding networks involves unfolding traditional iterative optimization algorithms such as ISTA (Iterative Shrinkage-Thresholding Algorithm) and ADMM (Alternating Direction Method of Multipliers) into deep network structures, such as ISTA-Net+ (Iterative Shrinkage-Thresholding Algorithm Network Plus), ADMM-Net (Alternating Direction Method of Multipliers Network), and TVI-Net (Total Variation Inverse Network). However, these methods do not fully utilize the structured sparsity of MRI images, the thresholding operation lacks channel adaptability, the single depth prior is prone to overfitting in scenarios with limited data, and the detail recovery and artifact suppression effects are poor under high acceleration.
[0049] Therefore, this invention proposes a high-speed MRI reconstruction method based on structured sparsity and channel attention image prior. By proposing a structured sparsity and channel attention deep image prior network (SSCA-DIPNet), with ISTA-Net+ as the baseline, it integrates a channel adaptive structured sparsity module, a channel attention deep image prior module, a globally learnable fusion mechanism, and a symmetric loss function to achieve high-speed, high-quality MRI reconstruction in scenarios with limited data.
[0050] Please combine Figure 1 and Figure 2 It illustrates the implementation flowchart and step flowchart of a high-speed MRI reconstruction method based on structured sparsity and channel attention image prior provided in the first embodiment of the present invention, the method comprising:
[0051] Step S1: Obtain the MRI image dataset, preprocess the MRI image dataset, and generate the source processed image dataset;
[0052] Step S2: Design multiple basic blocks to build an iterative network. Each basic block includes a data consistency module, a structured sparse module, and a channel attention deep image prior module.
[0053] Step S3: Input the source image dataset into the iterative network, and process it sequentially through the data consistency module, the structured sparsity module, and the channel attention deep image prior module to obtain the consistency image, the sparsity optimized image, and the structure optimized image, respectively.
[0054] Step S4: Design a loss function based on an iterative network, and output an MRI reconstructed image set after iterative training with multiple basic blocks.
[0055] Based on the implementation steps and corresponding methods, an integrated SSCA-DIPNet is proposed, namely a structured sparse and channel attention depth image prior network, to solve the problem of MRI reconstruction under high acceleration and low data conditions.
[0056] As an optional implementation, the MRI image dataset is non-contrast coronary MRA (Magnetic Resonance Angiography), a non-invasive imaging technique that can clearly display the structure of coronary arteries without the injection of contrast agents. Utilizing the inherent flow characteristics of blood and the temporal characteristics of heartbeats, and through specific magnetic resonance sequence design and ECG gating techniques, it achieves accurate assessment of the morphology, course, and wall features of the coronary arteries. In this embodiment, 227 images were sampled and resampled to 256×256 pixels for quantitative and qualitative analysis of the dataset. By shortening the scan time through contextualization, patient comfort can be improved and diagnostic accuracy can be increased.
[0057] Furthermore, in step S1, specifically:
[0058] The MRI image dataset is converted to K-space by Fourier transform, undersampling is performed by applying a radial sampling mask to obtain undersampled K-space data, and then the source processed image dataset is generated by inverse Fourier transform.
[0059] Specifically, MRI imaging requires the acquisition of K-space data. To shorten scan time, K-space undersampling is often used clinically, and its mathematical model can be expressed as: ,in, Represents an MRI image dataset. , Indicates the height of the MRI image; Indicates the width of the MRI image; This represents a masked two-dimensional Fourier transform; Indicates the undersampling mask. ; This represents the actual undersampled K-space data collected. ; Indicates noise. Based on undersampled K-space data The source image dataset is generated by inverse Fourier transform, which serves as the initial value for subsequent iterative operations. .
[0060] To better illustrate this, from a mathematical perspective, For a non-full-rank linear operator, its inverse operator The absence of prior constraints makes the reconstruction problem ill-posed. Therefore, to solve ill-posed problems, accelerated MRI reconstruction can be transformed into a regularized optimization problem of "data fidelity + prior constraints", namely:
[0061]
[0062] in, This represents the optimized MRI image dataset; Indicates data fidelity item, Represents the norm square operation; , Both represent regularization parameters, and ; This represents the prior information of the structured coefficients; This represents the prior of the depth image.
[0063] It can be noted that the data fidelity item This is used to ensure that the reconstruction results are consistent with the measured data at the sampled K-space location, which is the physical basis for guaranteeing the reliability of the reconstruction; structured sparse prior. By mining the sparse distribution characteristics of image features, redundant information and noise interference are suppressed, which conforms to the regularization optimization logic of image reconstruction; depth image prior Learning higher-order structural patterns in MRI images using a data-driven approach; regularization parameters , This is used to balance the weights of various constraints, achieving a dynamic adaptation between physical consistency and structural rationality.
[0064] Furthermore, in step S2, each basic block is connected in series to establish an iterative network.
[0065] The explanation is as follows: Based on the regularization optimization problem of "data fidelity + prior constraints", in order to achieve the dual goals of high speedup and data efficiency, the SSCA-DIPNet framework based on iterative expansion is used. That is, multiple basic blocks are set up and connected in series to build an iterative network. It aims to solve the regularization optimization problem through the synergistic mechanism of physical constraint anchoring, utilization of inherent attributes and feature efficiency improvement, so as to solve the ill-posed reconstruction problem and data dependency bottleneck under high speedup.
[0066] Specifically, in this embodiment, the iterative network... Iterative updates are achieved using identical and concatenated basic blocks. Each basic block is concatenated with a data consistency module, a structured sparse module, and a channel attention deep image prior module, forming a closed loop of physical constraints and dual prior optimization. Thus, the overall iterative relationship corresponds to:
[0067]
[0068] in, Indicates the current number The optimization results of each round; Indicates the first Nonlinear transformation operators for each basic block; Indicates the adjacent first Image estimation values obtained through multiple iterations; This represents the actual undersampled K-space data collected; This represents the undersampling mask.
[0069] It should be noted that the initial value for the iteration is the zero-filled reconstruction result, i.e., undersampled K-space data. The source image dataset is generated by inverse Fourier transform. ,go through After rounds of iteration, the final reconstructed image is output, i.e. This is the MRI reconstruction image obtained subsequently.
[0070] Please see Figure 3 Furthermore, step S3 includes:
[0071] Step S31: Based on the data consistency module, establish a data consistency correction mathematical model for the source processed image dataset and output a consistent image.
[0072] It should be noted that the data consistency module is the mathematical implementation of the physical constraints of K-space measurement in the iterative process, corresponding to the data fidelity term in the optimization problem.
[0073] Furthermore, in step S31, specifically:
[0074] The data consistency module enables the source processed image dataset to be evaluated to meet the constraints of the measured K-space data, determines the correction direction, and, in conjunction with backpropagation dynamic optimization, establishes a mathematical model for data consistency correction, outputting a consistent image corresponding to the source processed image dataset.
[0075] Specifically, the core function of the data consistency module is to ensure that the image estimation input in the current base block meets the constraints of the measured K-space data, avoiding the reconstruction results from deviating from physical laws due to aggressive undersampling at high speedup ratios, and providing reliable initial input for subsequent prior modeling; in this embodiment, the first... The mathematical model for data consistency correction is explained using the following basic blocks:
[0076]
[0077] in, Indicates the first The consistency image output by the data consistency module in each basic block; Indicates the adjacent first Image estimation values obtained through multiple iterations; Represents the learnable data consistency weights. ; Indicates the inverse Fourier transform; This represents a masked two-dimensional Fourier transform; Indicates the undersampling mask; This represents the actual undersampled K-space data collected.
[0078] It can be explained that, This represents the residual of the current image estimate in K-space, after inverse Fourier transform. Mapping back to the image domain, for Provides targeted correction directions; learnable data consistency weights. Dynamic optimization through backpropagation is used to balance the degree of preservation of the current image structure with the actual K-space data.
[0079] Step S32: Use a structured sparse module to mine the inherent sparsity properties of the consistent image and alleviate the loss of details under high speedup ratio, and output a sparse optimized image.
[0080] It is explained that the structured sparse module is used for end-to-end implementation of structured sparse priors. It reduces the dependence of high speedup ratios on large-scale data by mining the inherent sparse properties of MRI images, while mitigating the loss of details under high speedup ratios. The module adopts a progressive process of "high-dimensional feature encoding - channel adaptive thresholding - sparse feature decoding" to adapt to the feature distribution characteristics of high speedup ratios.
[0081] Further, step S32 includes:
[0082] Step S321: Sparsely encode the consistent image and map it to a high-dimensional sparse feature space. Mine sparse structural features through multi-layer convolution and non-linear activation.
[0083] Specifically, the consistency image output by the data consistency module Mapping to a high-dimensional sparse feature space allows for better capture of key, discriminative information in the image while reducing the negative impact of redundant features. Through multi-layer convolution and non-linear activation—specifically, a sliding window approach—local perception of the consistent image is achieved, extracting local features at different scales and orientations. Furthermore, non-linear transformation capabilities are introduced, enabling the structured sparse module to learn more complex non-linear mapping relationships. This process aims to mine the sparse structural features of the image; the corresponding calculation formula is as follows:
[0084]
[0085] in, Represents the initial feature mapping; This represents the first-stage encoded features; This represents the second-stage coding features; Represents the feature map convolution kernel, , This indicates the number of output channels and the number of input channels. Indicates the height and width of the convolution kernel; Represents a consistent image; , , Both represent bias terms; Indicates the activation function; , Both represent feature refinement convolution kernels. .
[0086] Preferably, in this embodiment, the number of channels It can be adjusted according to the actual situation; optionally, convolutional coding can also be replaced by tensor decomposition to enhance multi-dimensional sparsity capture.
[0087] It can be explained that the second stage of encoding features This refers to the final sparse structural features obtained; feature mapping convolution kernel. This is used to map a single-channel image to multi-channel features; the bias term is used to eliminate feature distribution offset; through the activation function... Nonlinearity is introduced to accommodate complex structures.
[0088] Step S322: Design a channel-level learnable threshold mechanism to perform soft thresholding on sparse structural features and apply differentiated sparsity constraints.
[0089] Specifically, since traditional methods use a global threshold, which cannot adapt to the feature importance of different channels, a channel-level learnable threshold mechanism is designed to independently perform a soft thresholding operation on each feature channel, thereby achieving differentiated sparsity constraints. This means that key anatomical structures are preserved and noise is suppressed in a differentiated manner. The corresponding calculation formula is as follows:
[0090]
[0091] in, Indicates the first Features after thresholding of each channel; Represents the set of all channels after thresholding; Indicates the number of channels; Indicates the second-stage coding features The One channel; Indicates the first The learnable threshold of each channel, ; , Both indicate soft threshold operation; This represents absolute value operations.
[0092] It can be explained that soft thresholding operation , Used to suppress values below the learnable threshold Redundant features are retained if they exceed a learnable threshold. Key structural features.
[0093] Step S323: Map the thresholded sparse structural features back to the image domain, combine the residual connection with the consistent image for sparse decoding, and output the sparse optimized image.
[0094] Specifically, the structured sparse module thresholds the sparse features. Mapping back to the image domain and combining the effective information from the residual connection multiplexing data consistency module, i.e., reusing the consistent image, avoids introducing additional artifacts during the decoding process. The corresponding calculation formula is:
[0095]
[0096]
[0097] in, This indicates initial decoding activation; Indicates the final decoded features; Indicates the first Sparse-optimized images output by the structured sparse modules in each basic block; Indicates the activation function; , Both represent decoding convolution kernels. ; Represents the set of all channels after thresholding; , , Both represent learnable bias parameters; Represents a consistent image; Represents the feature map convolution kernel, .
[0098] It can be explained that the feature mapping convolution kernel Used to map multi-channel features back to a single-channel image; learnable bias parameters are used to eliminate feature distribution offsets in the convolution output.
[0099] Step S33: The channel attention deep image prior module learns the image structure rules through a data-driven approach, performs residual correction on the sparse optimization image, and outputs a structure-optimized image.
[0100] To address the shortcomings of sparse priors in modeling higher-order structures, a deep image prior module with channel attention is introduced to better focus on high-information regions in scenarios with small samples and high speedup ratios; this module is used to learn the structural patterns of MRI images through a data-driven approach and optimize the sparse images. Perform residual correction.
[0101] Further, step S33 includes:
[0102] The channel attention depth image prior module includes a DIP network and a channel attention module;
[0103] To explain, the DIP network, or Deep Image Prior, in this embodiment adopts a multi-scale architecture of encoder-decoder + skip connections to focus on information-rich regions and reduce artifacts; the channel attention module, or CA (Channel Attention) module, optimizes the feature selection logic for high speedup by capturing global information, learning weights, and weighting features to improve the utilization rate of key features and reduce noise interference under small sample and high speedup conditions.
[0104] Step S331: The sparse optimized image is scaled by the DIP network and features are weighted by the channel attention module to output the source structure optimized image.
[0105] It can be explained that in the DIP network, the scale design is optimized for high speedup ratios. This is achieved by downsampling to abstract global features, upsampling to restore local details, and skip connections to fuse high and low dimensional information to avoid loss of details under high speedup ratios.
[0106] To better illustrate, the calculation formula for the channel attention module is as follows:
[0107]
[0108]
[0109] in, Indicates CA module operation; This represents the channel weight vector learned by the CA module; Feature map representing the input channel attention module; This represents element-wise multiplication; Indicates global average pooling; , Both represent convolution kernels; , This represents the bias term corresponding to the two convolution kernels; This represents the activation function.
[0110] It can be noted that the feature map of the input channel attention module To obtain the relevant feature maps through the aforementioned steps during actual operation, therefore The output feature map is the one weighted by the channel attention module; while global average pooling... Used to capture feature maps Global channel information in the convolution kernel; convolutional kernel for Convolution kernel; activation function Use the sigmoid function to control the output range within a certain range. between.
[0111] As an alternative implementation, in subsequent applications, the channel attention module can be upgraded to spatial-channel dual attention to focus on the lesion area; and the fusion mechanism can be replaced with region adaptive attention fusion, dynamically adjusting the weights of different priors to achieve balance.
[0112] Further, step S331 includes:
[0113] Step S3311: Based on the sparse optimization image, the encoder performs downsampling to abstract global features. The auxiliary channel attention module sequentially performs global information capture, weight learning, and feature weighting, and outputs a weighted feature image.
[0114] Specifically, the calculation formula for downsampling is:
[0115]
[0116] in, , These represent the first and second parts of the encoder. Scale and the first Scale-weighted feature image; , Both represent the downsampling scale index; This indicates the operation of the channel attention module; Indicates the activation function; Indicates the downsampling convolution kernel; Indicates the encoder's first... The scale bias term.
[0117] It can be explained that during the downsampling process, That is, the initial value of the downsampling operation is the sparse optimization image. Preferably, the downsampling convolution kernel With a high speedup, the number of channels is set to 32 to balance representation capability and computation, achieving feature dimensionality reduction and abstraction.
[0118] Step S3312: The weighted feature image is upsampled by the decoder to restore local details, and high- and low-dimensional information is fused with skip connections to output an optimized source structure image.
[0119] Specifically, the calculation formula for upsampling is as follows:
[0120]
[0121] in, , These represent the numbers in the decoder. Scale and the first Weighted feature image recovered at scale; Indicates the upsampling convolution kernel; Indicates the decoder's first Scale bias term; Indicates the first in the encoder Scale-weighted feature image; This indicates a channel-level skip connection.
[0122] It can be explained that, under high acceleration ratios, channel-level jump connections... It can more effectively fuse the structural features of the encoder with the abstract features of the decoder, avoiding edge blurring caused by upsampling; upsampling convolution kernel The splicing features are then compressed to 32 channels for simplification and optimization, enabling detailed restoration.
[0123] Next, after downsampling and upsampling, the source structure optimized image is output, and the corresponding calculation formula is:
[0124]
[0125] in, Indicates the first The source structure optimization image output by the channel attention depth image prior module in each basic block; Indicates the output convolution kernel; Indicates the first in the decoder Weighted feature image recovered at scale; This indicates the bias term.
[0126] It can be explained that the source structure is optimized in the image. It contains finer anatomical details and lower artifacts; during operation, the decoder's final output is... That is, an index scale of 1 corresponds to the last upsampling.
[0127] Step S332: Introduce a globally learnable step size to perform residual correction on the sparse optimization image, and combine the source structure optimization image to output the structure optimization image.
[0128] Specifically, to balance the contributions of sparse modeling and deep priors, a globally learnable step size is introduced to perform residual correction on the sparse optimized image. This balances the complementarity of structured sparse priors and deep priors, avoids feature conflicts, and improves structural integrity and detail realism. The corresponding calculation formula is as follows:
[0129]
[0130] in, Indicates the first The structure-optimized image output by the channel attention depth image prior module in each basic block; Represents a sparse optimized image; Represents the source structure optimized image; Indicates the globally learnable step size. .
[0131] It can be explained that the globally learnable step size, i.e. The constraint is non-negative to avoid iteration instability caused by negative step size.
[0132] Furthermore, in step S4, specifically:
[0133] A loss function is designed by integrating data consistency loss and symmetry loss using a balanced coefficient. The loss function is used to complete iterative training of multiple basic blocks, and the iterative network outputs a set of MRI reconstructed images.
[0134] Specifically, the data consistency loss is directly related to the data fidelity term in the optimization problem. It measures the pixel-level difference between the final reconstructed result and the ground truth image, and is a core constraint to ensure reconstruction accuracy. The corresponding calculation formula is as follows:
[0135]
[0136] in, This indicates a loss of data consistency. Indicates the total number of image samples; Indicates the total number of pixels in the image; Indicates the first Image samples The final reconstruction result after rounds of iterations; Indicates the first The ground truth image corresponding to each image sample; Represents the norm squaring operation.
[0137] Symmetric loss is used to constrain the consistency of the encoding, thresholding, and decoding processes in structured sparse modules, avoiding excessive feature distortion or model overfitting, and indirectly achieving the regularization effect of structured sparse priors. Specifically, it addresses the overfitting problem with few samples by constraining the consistency of encoding-decoding features, thus enhancing the generalization ability of the iterative network. The corresponding calculation formula is:
[0138]
[0139]
[0140] in, Indicates symmetric loss; This indicates the total number of elements in the feature map; Indicates the final decoded features; Represents the initial feature mapping; Indicates the number of iterations; Indicates the first Symmetric loss per iteration; Represents the norm squaring operation.
[0141] It can be explained that the initial feature mapping is the first... The original encoded features of each basic block, and their relationship with the final decoded features. The difference directly reflects the degree of feature distortion in the encoding-decoding process; the total number of elements in the feature map This is used to normalize the symmetric loss of a single basic block. , and thus symmetrical loss This is the average symmetry loss of all basic blocks, used to balance the constraint strength of each iteration.
[0142] Next, by integrating two types of constraints—data fidelity and sparsity consistency—that is, by using a balanced coefficient to integrate data consistency loss and symmetry loss in the loss function design, all learnable parameters are optimized simultaneously to ensure that the iterative network balances reconstruction accuracy and generalization ability. The corresponding calculation formula is as follows:
[0143]
[0144] in, Represents the loss function; This indicates a loss of data consistency. Indicates symmetric loss; This represents the balance coefficient.
[0145] It can be explained that in the calculation of the loss function, the balance coefficient The weight ratio of symmetry loss and data consistency loss is used to adjust the data fidelity and sparsity consistency. Optionally, in actual operation, perceptual loss can also be added to improve the subjective visual effect.
[0146] Understandably, an iterative network is built based on multiple basic blocks. Through the cooperation of three modules, the source image dataset is processed sequentially. Consistent images are used to ensure data fidelity. Sparse optimized images utilize the structured sparsity characteristics of MRI images for analysis, to differentiate and preserve key anatomical structures and suppress noise. To address the problem of insufficient high-dimensional feature modeling with limited data, a channel attention deep image prior module is used to focus on information-rich regions, reduce artifacts, and obtain structure-optimized images. At the same time, the complementarity of structured sparse prior and deep prior is balanced to avoid feature conflicts and improve structural integrity and detail realism. Finally, to address the overfitting problem with limited data, a loss function is designed to enhance the generalization ability of the iterative network.
[0147] As explained, in this invention, the structured sparse module employs a progressive process of high-dimensional feature encoding, channel adaptive thresholding, and sparse feature decoding. This process utilizes learnable channel-level thresholds to achieve differentiated feature processing and enhance the preservation of key anatomical structures. Specifically, the encoding stage extracts sparse structural features through multi-layer convolution, the thresholding stage performs soft thresholding independently on each feature channel, and the decoding stage combines residual connections to map back to the image domain.
[0148] In the channel attention deep image prior module, a multi-scale architecture of encoder-decoder + jump connection is adopted, which embeds the channel attention module. The encoder downsamples to abstract global features, the decoder upsamples to recover local details, and the jump connection fuses high- and low-dimensional information. The CA module captures global channel information through global average pooling, learns channel weights through 1×1 convolution and activation function, dynamically calibrates feature importance, and focuses on information-rich regions. In addition, a globally learnable stride is introduced in this module to perform residual updates on structure-optimized images and sparse-optimized images, balancing the contributions of structured sparse priors and deep image priors. Through the collaborative operation of the two priors, the structured sparse prior mines structured sparse features, while the deep image prior focuses on key channels and multi-scale details to solve the problems of detail loss and artifacts under high acceleration.
[0149] Finally, in the design of the loss function, data consistency loss and symmetry loss are integrated. Data consistency loss ensures the consistency between the reconstruction result and the measured K-space data, while symmetry loss constrains the feature consistency of the encoding-decoding process in the structured sparse module, preventing overfitting and improving the model's generalization ability.
[0150] Please combine Figures 4-6 To better illustrate this, the iterative network proposed in this invention, namely the target model SSCA-DIPNet, is compared with existing methods such as the baseline ISTA-Net+, the ablation model SS-Net (SSM only), and CA-DIPNet (CA-DIP only) to verify the reliability of the target model provided in this invention. Specifically, Table 1 shows the quantitative evaluation results of the reconstruction quality of the non-comparison coronary MRA test set.
[0151] Table 1. Quantitative assessment results of reconstruction quality in the non-contrast coronary MRA test set.
[0152]
[0153] It can be noted that in Table 1, AF, or Acceleration Factor, is used to obtain the quantitative assessment results (mean ± standard deviation) of reconstruction quality in the non-contrast coronary MRA test set by adjusting its value. The relevant indicators are PSNR (Peak Signal-to-Noise Ratio) and SSIM (Structural Similarity Index Measure). Figures 4-6 In the images, the acceleration factors AF are 20, 10, and 5, respectively. The evaluation results corresponding to the first row of images are PSNR and SSIM, respectively. The second row of images corresponds to a magnified view showing the reconstructed details. The third row of images corresponds to the error images. Through subjective visual evaluation, it can be found that the reconstruction quality of the target model is the best in the magnified view of the reconstructed details, that is, the fidelity of the reproduction of fine texture details is improved and the undersampling artifacts are significantly reduced. For the error images, the corresponding images fed back by the target model are better than other comparison methods in terms of texture rendering clarity, naturalness, and structural integrity. Therefore, judging from the evaluation results and the feedback of related result images, the data results of the target model are better.
[0154] When the target model proposed in this invention is implemented in real-world settings, it can be used in environments such as cardiovascular imaging, dynamic imaging and functional MRI, and pediatric and elderly patient examinations. In cardiovascular imaging, it enables rapid scanning of non-contrast coronary MRA, reducing patient scanning time, improving examination comfort, and assisting in coronary heart disease screening. In dynamic imaging and functional MRI, such as cardiac dynamic MRI and diffusion-weighted imaging (DWI), it solves the motion artifact problem caused by the slow scanning speed of traditional methods. In pediatric and elderly patient examinations, it shortens scanning time and reduces the risk of imaging failure due to patient agitation.
[0155] The specific implementation methods are as follows: First, the post-processing software integrated into the MRI equipment serves as a high-speed reconstruction module, directly connecting to the K-space data acquired by the equipment to output high-quality reconstructed images; second, it is developed as an independent medical image reconstruction software, suitable for MRI image post-processing in hospital radiology departments and medical imaging centers; third, it is compatible with MRI scanners from different manufacturers and supports K-space data formats of various scanning sequences.
[0156] A second embodiment of the present invention provides an electronic device, which includes a processor, a communication interface, a memory, and a communication bus. The processor, the communication interface, and the memory communicate with each other through the communication bus. The processor calls logical instructions in the memory to execute a high-speed MRI reconstruction method based on structured sparsity and channel attention image prior as described in any embodiment of the present invention.
[0157] When it is in operation, it needs to use a high-speed MRI reconstruction method based on structured sparsity and channel attention image prior. Therefore, whether the device and program data are integrated or different hardware is configured to produce a function with similar effect to that of the present invention, it is within the protection scope of the present invention. The device has the same beneficial effect as the aforementioned high-speed MRI reconstruction method based on structured sparsity and channel attention image prior, and will not be described in detail here.
[0158] It should be noted that the order of the above embodiments of the present invention is merely for descriptive purposes and does not represent the superiority or inferiority of the embodiments. The processes depicted in the accompanying drawings do not necessarily require a specific or sequential order to achieve the desired result. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.
[0159] The various embodiments in this specification are described in a progressive manner. The same or similar parts between the various embodiments can be referred to each other. Each embodiment focuses on describing the differences from other embodiments.
Claims
1. A high-speed MRI reconstruction method based on structured sparsity and channel attention image prior, characterized in that, The method includes: Acquire MRI image datasets, preprocess MRI image datasets, and generate source processed image datasets; The design incorporates multiple basic blocks to construct an iterative network. Each basic block includes a data consistency module, a structured sparse module, and a channel attention deep image prior module. The source image dataset is input into the iterative network, and then processed sequentially by the data consistency module, the structured sparsity module, and the channel attention deep image prior module to obtain the consistent image, the sparsity optimized image, and the structure optimized image, respectively. The loss function is designed based on an iterative network, and after iterative training with multiple basic blocks, the MRI reconstructed image set is output.
2. The high-speed MRI reconstruction method based on structured sparsity and channel attention image prior as described in claim 1, characterized in that, Preprocessing the MRI image dataset to generate the source processed image dataset, specifically: The MRI image dataset is converted to K-space by Fourier transform, undersampling is performed by applying a radial sampling mask to obtain undersampled K-space data, and then the source processed image dataset is generated by inverse Fourier transform.
3. The high-speed MRI reconstruction method based on structured sparsity and channel attention image prior as described in claim 1, characterized in that, Each basic block is connected in series to build an iterative network.
4. The high-speed MRI reconstruction method based on structured sparsity and channel attention image prior as described in claim 3, characterized in that, The source image dataset is input into the iterative network, and processed sequentially by the data consistency module, the structured sparsity module, and the channel attention deep image prior module, resulting in consistent images, sparsity-optimized images, and structure-optimized images, respectively. Based on the data consistency module, a mathematical model for data consistency correction is established for the source image dataset, and a consistent image is output. The structured sparse module is used to mine the inherent sparsity properties of consistent images and alleviate the loss of details under high speedup ratio, outputting a sparse-optimized image; The channel attention deep image prior module learns the structural patterns of images through a data-driven approach, performs residual correction on sparse optimized images, and outputs structurally optimized images.
5. The high-speed MRI reconstruction method based on structured sparsity and channel attention image prior as described in claim 4, characterized in that, Based on the data consistency module, a mathematical model for data consistency correction is established for the source image dataset, and a consistent image is output, specifically: The data consistency module enables the source processed image dataset to be evaluated to meet the constraints of the measured K-space data, determines the correction direction, and, in conjunction with backpropagation dynamic optimization, establishes a mathematical model for data consistency correction, outputting a consistent image corresponding to the source processed image dataset.
6. The high-speed MRI reconstruction method based on structured sparsity and channel attention image prior as described in claim 4, characterized in that, A structured sparse module is used to mine the inherent sparsity properties of consistent images and alleviate detail loss under high speedup ratios, outputting a sparse-optimized image, including: Sparse encoding is performed on the consistent image, which is then mapped to a high-dimensional sparse feature space. Sparse structural features are then mined through multi-layer convolution and non-linear activation. Design a channel-level learnable threshold mechanism to perform soft thresholding on sparse structural features and apply differentiated sparsity constraints. The thresholded sparse structural features are mapped back to the image domain, and sparse decoding is performed by combining residual connections with the consistent image to output a sparse optimized image.
7. The high-speed MRI reconstruction method based on structured sparsity and channel attention image prior as described in claim 4, characterized in that, The channel attention deep image prior module learns image structure patterns through a data-driven approach, performs residual correction on sparse optimized images, and outputs structure-optimized images, including: The channel attention depth image prior module includes a DIP network and a channel attention module; The sparse optimized image is scaled using a DIP network and weighted by a channel attention module to output a source structure optimized image. A globally learnable step size is introduced to perform residual correction on the sparse optimization image, and the output structure optimization image is combined with the source structure optimization image.
8. The high-speed MRI reconstruction method based on structured sparsity and channel attention image prior as described in claim 7, characterized in that, The sparse optimized image is scaled using a DIP network and weighted by a channel attention module to output a source structure optimized image, including: Based on the sparse optimization image, the encoder performs downsampling to abstract global features, and the auxiliary channel attention module sequentially performs global information capture, weight learning and feature weighting to output a weighted feature image. The weighted feature image is upsampled by the decoder to restore local details, and high- and low-dimensional information is fused with skip connections to output an optimized source structure image.
9. The high-speed MRI reconstruction method based on structured sparsity and channel attention image prior as described in claim 1, characterized in that, The loss function is designed based on an iterative network. After iterative training with multiple basic blocks, the output MRI reconstructed image set is as follows: A loss function is designed by integrating data consistency loss and symmetry loss using a balanced coefficient. The loss function is used to complete iterative training of multiple basic blocks, and the iterative network outputs a set of MRI reconstructed images.
10. An electronic device, characterized in that, The device includes a processor, a communication interface, a memory, and a communication bus. The processor, the communication interface, and the memory communicate with each other through the communication bus. The processor calls logical instructions in the memory to execute the high-speed MRI reconstruction method based on structured sparse and channel attention image prior as described in any one of claims 1 to 9.