A method and system for accelerated magnetic resonance imaging image reconstruction

Abstract

The application belongs to the technical field of image reconstruction, and particularly relates to an image reconstruction method and system for accelerating magnetic resonance imaging. The original data collected by a magnetic resonance device is subjected to undersampling processing, and a preliminary reconstruction is performed in combination with a current frame and adjacent multiple frames of data to obtain a preliminary reconstruction image. The preliminary reconstruction image is input into a first branch and a second branch respectively. The first branch adopts a pre-trained visual basic model to output prior information of structural boundaries and spatial layouts of different anatomical regions in the preliminary reconstruction image. The second branch extracts global structural features and local texture edge features and performs adaptive fusion to obtain backbone features. After the prior information and the backbone features are fused, a detail residual image is predicted and superimposed on the preliminary reconstruction image to obtain a final reconstruction image. Through the combination of adjacent frame undersampling reconstruction and prior guided detail enhancement, a high-quality reconstruction image can still be obtained under a high acceleration multiple, and the acquisition time is effectively shortened.

Description

Technical Field

[0001] This invention belongs to the field of image reconstruction technology, specifically an image reconstruction method and system for accelerating magnetic resonance imaging. Background Technology

[0002] The statements in this section merely refer to the background art related to this invention and do not necessarily constitute prior art.

[0003] Magnetic resonance imaging (MRI) is widely used in medical imaging due to its high resolution and lack of ionizing radiation. However, due to the physical limitations of MRI, acquiring high-quality images requires a long data acquisition time. In some clinical applications, patients need to hold their breath for extended periods during data acquisition, such as in cardiac magnetic resonance imaging (CMR), to reduce interference from lung movements on cardiac imaging; however, breath-holding is particularly difficult for the elderly, young children, and patients with limited respiratory function. Excessive acquisition time not only easily leads to motion artifacts due to involuntary movements of the patient, severely interfering with image quality, but may also result in examination failure or repeated scans, increasing the consumption of medical resources.

[0004] To shorten acquisition time, existing technologies often employ undersampling strategies, which involve acquiring data in k-space at a density lower than required by the Nyquist sampling theorem. However, undersampling introduces aliasing artifacts, leading to structural blurring and loss of detail in the reconstructed image. While traditional compressed sensing reconstruction methods can alleviate this problem to some extent, they suffer from limitations such as slow reconstruction speed, complex parameter tuning, and poor adaptability to high speedup factors. In recent years, deep learning-based reconstruction methods have gradually emerged, but they still struggle to maintain both overall structural fidelity and local detail clarity at high speedup levels. Summary of the Invention

[0005] This invention provides an image reconstruction method and system for accelerating magnetic resonance imaging, which solves the problem of image quality degradation caused by excessively long data acquisition time in existing cardiac magnetic resonance imaging. By utilizing the temporal information of adjacent frames and a priori guided detail enhancement mechanism, the acquisition time is significantly shortened while ensuring image reconstruction quality, thereby improving the clinical applicability of cardiac magnetic resonance imaging.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: The first aspect of this invention discloses an image reconstruction method for accelerating magnetic resonance imaging, comprising the following steps: The raw data acquired by the magnetic resonance imaging device is obtained. Taking the data frame corresponding to the current image to be reconstructed as the center, multiple adjacent data frames are selected, and each selected data frame is undersampled to obtain undersampled data frames. After feature extraction, artifact suppression and data consistency constraints, a preliminary reconstructed image is obtained. Input the preliminary reconstructed images into the first branch and the second branch respectively; In the first branch, a pre-trained visual base model is used to process the preliminary reconstructed image and output the prior information of the structural boundaries and spatial layout of different anatomical regions in the preliminary reconstructed image. In the second branch, global structural features and local texture edge features of the initially reconstructed image are extracted, and the global structural features and local texture edge features are adaptively fused to obtain the backbone features; The prior information output from the first branch is fused with the backbone features output from the second branch to obtain enhanced fusion features. Based on the enhanced fusion features, a detail residual map is predicted, and the detail residual map is superimposed on the preliminary reconstructed image to obtain the final reconstructed image.

[0007] Furthermore, during the initial reconstruction, the sensitivity map of the receiving coil of the magnetic resonance device is estimated based on the undersampled multi-frame data, and the undersampled multi-frame data is converted to the image domain using the sensitivity map to obtain the initial reconstruction features.

[0008] Furthermore, the obtained initial reconstructed features are subjected to multi-scale feature extraction by the encoder to obtain multi-layer coded features; Deep coding features are selected from multi-layer coding features, and artifact suppression processing is applied to obtain artifact-suppressed deep features. The deep features after artifact suppression are restored layer by layer by the decoder and the cue feature map generated by the cue module that matches the current scale, to obtain the reconstructed feature map.

[0009] Furthermore, the data consistency constraint is as follows: the reconstructed feature map obtained after artifact suppression is transformed to the k-space domain using Fourier transform. Combined with the mask used for undersampling, the data consistency constraint of the current frame is updated. Then, it is transformed to the image domain and merged to obtain the preliminary reconstructed image; as shown in the following equation: ; ; in, This represents the updated k-space data. The data at time t is the multi-coil k-space data after undersampling. Represents the consistency constraint weight. Indicates the undersampling mask. This indicates element-wise multiplication. Represents the reference observation k-space, This represents the regularization update term generated by the reconstructed network. This indicates the initial reconstructed image. This indicates a coil merging operation. This indicates that an inverse Fourier transform is performed on undersampled k-space data.

[0010] Furthermore, the prompting module adaptively weights and combines multiple learnable prompting components based on the global statistical information of the input features in the current decoding stage, generating a prompting feature map that matches the current scale.

[0011] Furthermore, in the second branch, global structural features and local texture edge features of the initially reconstructed image are extracted, specifically including: Model the global structural relationships to obtain global structural features; Local texture and edge information are extracted to obtain local texture edge features.

[0012] Furthermore, the global structural features and local texture edge features are adaptively fused. Specifically, an adaptive fusion attention block is used to dynamically allocate the contribution weights of global structural features and local texture edge features based on the prior information and feature state of the current reconstructed region, thereby achieving adaptive fusion.

[0013] Furthermore, the prior information output from the first branch is fused with the backbone features output from the second branch. Specifically, guided by the prior information, the backbone features are fused with the prior information at the feature level to obtain enhanced fused features, as shown in the following formula: ; in, This indicates a priori guided fusion mapping. This indicates enhanced fusion features.

[0014] Furthermore, the detail residual image is superimposed on the preliminary reconstructed image. Specifically, the detail residual image is multiplied by a preset residual fusion coefficient and then added pixel-by-pixel to the preliminary reconstructed image, as shown in the following formula: ; ; in, Represents the residual prediction mapping. Representing the detailed residual plot, Represents the residual fusion coefficient. This represents the final reconstructed image.

[0015] A second aspect of the present invention discloses an image reconstruction system for accelerating magnetic resonance imaging, comprising: Magnetic resonance imaging (MRI) acquisition equipment is used to acquire raw data; A processor, connected to the magnetic resonance imaging acquisition device, is configured to perform the following steps: The raw data acquired by the magnetic resonance imaging device is obtained. Taking the data frame corresponding to the current image to be reconstructed as the center, multiple adjacent data frames are selected, and each selected data frame is undersampled to obtain undersampled data frames. After feature extraction, artifact suppression and data consistency constraints, a preliminary reconstructed image is obtained. Input the preliminary reconstructed images into the first branch and the second branch respectively; In the first branch, a pre-trained visual base model is used to process the preliminary reconstructed image and output the prior information of the structural boundaries and spatial layout of different anatomical regions in the preliminary reconstructed image. In the second branch, global structural features and local texture edge features of the initially reconstructed image are extracted, and the global structural features and local texture edge features are adaptively fused to obtain the backbone features; The prior information output from the first branch is fused with the backbone features output from the second branch to obtain enhanced fusion features. Based on the enhanced fusion features, a detail residual map is predicted, and the detail residual map is superimposed on the preliminary reconstructed image to obtain the final reconstructed image.

[0016] Compared with existing technologies, one or more of the above technical solutions have the following beneficial effects: By employing undersampling to reduce the amount of data collected and shorten the acquisition time from the source, and by combining data from multiple adjacent frames for reconstruction, the structural gaps caused by single-frame undersampling are compensated for using inter-frame temporal information, achieving image reconstruction under high acceleration. Based on this, a pre-trained visual foundation model is introduced through the first branch, utilizing its image understanding capabilities to extract prior information such as the structural boundaries and spatial layout of anatomical regions from the preliminary reconstructed image, providing accurate directional guidance for subsequent detail restoration. The second branch models global structural relationships and local texture edges respectively, and adaptively fuses them to balance overall structural consistency with local detail clarity. By fusing the prior information from the first branch with the core features of the second branch, targeted compensation is made for blurred boundaries and missing details, thus generating high-quality reconstructed images with complete structure, clear edges, and rich texture even under high-acceleration undersampling conditions, improving the clinical applicability of cardiac magnetic resonance imaging. Attached Figure Description

[0017] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0018] Figure 1 A schematic diagram of an image reconstruction process provided for one or more embodiments of the present invention; Figure 2 This is a schematic diagram of cardiac reconstruction results obtained by different methods provided in one or more embodiments of the present invention. Detailed Implementation

[0019] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0020] It should be noted that the following detailed descriptions are exemplary and intended to provide further illustration of the invention. Unless otherwise specified, 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.

[0021] As described in the background section, existing techniques often employ undersampling strategies to shorten acquisition time, i.e., acquiring data in k-space at a density lower than required by the Nyquist sampling theorem. However, undersampling introduces aliasing artifacts, leading to problems such as structural blurring and loss of detail in the reconstructed image. The reason why undersampling reconstruction methods struggle to achieve satisfactory image quality at high speedups is that they typically rely only on limited information from a single frame of k-space data for reconstruction, lacking effective utilization of structural changes between adjacent frames over time. Furthermore, a single network structure cannot simultaneously handle the two different levels of tasks: global structure preservation and local detail restoration, resulting in blurred or distorted reconstructed images in high-frequency regions such as edges and textures.

[0022] To address the aforementioned issues, this solution provides an image reconstruction method and system for accelerating magnetic resonance imaging. A two-stage reconstruction strategy is designed: In the first stage, joint reconstruction is performed using temporal information from multiple adjacent frames centered on the current frame, and structural defects caused by undersampling are dynamically compensated through a cue learning mechanism; In the second stage, a large model prior and a separate global-local detail enhancement mechanism are introduced to finely compensate for the preliminary reconstruction results, ultimately achieving high-quality reconstructed image output.

[0023] The relevant technical terms used in this solution are explained below.

[0024] k-space data is raw data directly acquired by magnetic resonance imaging (MRI) devices and stored in the form of frequency information. It is not a visible image in itself, but can be understood as the negative data of a photograph. Each data point contains information about a certain frequency component of the entire image. After mathematical transformation (usually inverse Fourier transform), it is converted to the image domain and becomes an image visible to the naked eye.

[0025] Multi-coil data: Magnetic resonance imaging (MRI) devices typically have multiple receiving coils (equivalent to multiple sensors), each independently acquiring k-space data from a different spatial location. Multiple coils provide image information from different angles, and by coil merging, the images generated by each coil can be combined into a single, complete image.

[0026] Undersampling, to shorten acquisition time, only collects data from a portion of the k-space, skipping the rest. It's equivalent to capturing only a part of an image instead of the entire picture. Undersampling results in image quality loss (artifacts). During undersampling, an undersampling mask is used to mark which k-space locations are preserved and which are skipped.

[0027] A coil sensitivity map describes the spatial distribution of the induction intensity of each receiving coil in a magnetic resonance imaging (MRI) device at different spatial locations. Different coils have varying sensing capabilities at the same location, and the sensitivity map records this spatial difference. It is used to subsequently merge data from multiple coils into a single, complete image.

[0028] SAM2 (Segment Anything Model 2), a pre-trained general-purpose vision foundation model developed by Meta (Facebook), excels at extracting information such as object boundaries, region layout, and spatial hierarchy from images. In this scheme, it is used as a "prior knowledge provider" to help determine which regions need enhanced detail recovery.

[0029] An image reconstruction method for accelerating magnetic resonance imaging includes the following steps: The raw data acquired by the magnetic resonance imaging device is obtained. Taking the data frame corresponding to the current image to be reconstructed as the center, multiple adjacent data frames are selected, and each selected data frame is undersampled to obtain undersampled data frames. After feature extraction, artifact suppression and data consistency constraints, a preliminary reconstructed image is obtained. Input the preliminary reconstructed images into the first branch and the second branch respectively; In the first branch, a pre-trained visual base model is used to process the preliminary reconstructed image and output the prior information of the structural boundaries and spatial layout of different anatomical regions in the preliminary reconstructed image. In the second branch, global structural features and local texture edge features of the initially reconstructed image are extracted, and the global structural features and local texture edge features are adaptively fused to obtain the backbone features; The prior information output from the first branch is fused with the backbone features output from the second branch to obtain enhanced fusion features. Based on the enhanced fusion features, a detail residual map is predicted, and the detail residual map is superimposed on the preliminary reconstructed image to obtain the final reconstructed image.

[0030] The specific process is detailed below.

[0031] Step 1: Convert the original .mat multi-coil dataset to .h5 format, obtain the k-space data of the five adjacent frames of multi-coil data centered on the current frame to be reconstructed, apply an undersampling mask to it, and estimate the coil sensitivity map accordingly. Use the sensitivity map to convert the undersampled k-space data into an image domain representation to obtain the initial reconstruction features. Step 2: Input the initial reconstructed features into the encoder for multi-scale feature extraction to obtain multi-layer coded features; Step 3: Based on the context artifact removal module, obtain the artifact-suppressed deep features; input the artifact-suppressed deep features into the decoder, and combine them with the prompting module to perform layer-by-layer feature recovery to obtain the reconstructed feature map; Step 4: Combine the mask to update the consistency constraint of the current k-space, then perform inverse Fourier transform and coil merging to obtain the preliminary reconstructed image; Step 5: For the initial reconstructed image, the first branch is embedded using SAM2 to obtain prior structural and spatial information. Due to its strong general representation capabilities, SAM2 can extract relatively stable target boundaries, region layouts, and spatial hierarchy information from the initial reconstructed image, thus providing structural prior guidance for subsequent detail enhancement. For reconstructed regions with blurred boundaries, missing textures, or lost local high-frequency information, this type of prior information helps improve the accuracy of detail recovery. Step 6: The second branch uses the input detail enhancement network. The structure modeling module is used to model the global structural relationships, and the local detail module is used to extract local texture and edge information. The adaptive fusion attention block is used to adaptively fuse the global structural features and local detail features to obtain the backbone features. Under the guidance of prior information, the backbone features are fused with prior structure and spatial information to obtain the enhanced fusion features. Based on the enhanced fusion features, the detail residual map is predicted, and the detail residual map is residually fused with the preliminary reconstructed image to obtain the final reconstructed image.

[0032] Combination Figure 1 The image reconstruction process of this scheme is as follows: the original multi-coil k-space data collected by the magnetic resonance device is acquired, the multi-coil k-space data of 5 adjacent frames are selected with the current frame to be reconstructed as the center, an undersampling mask is applied, and the data is converted to the image domain by inverse Fourier transform. The coil sensitivity map is used to merge the coils to obtain the initial reconstruction features. The initial reconstructed features are sequentially processed through a convolutional layer and several downsampling blocks for multi-scale feature extraction, resulting in multi-layer coded features. The deepest coded features are input into a context artifact removal module for artifact suppression, yielding artifact-suppressed deep features. These features are then progressively recovered through several upsampling blocks, each working in conjunction with a prompting module. The prompting module adaptively weights and combines multiple learnable prompting components based on global statistical information of the input features of the current upsampling block, generating a prompting feature map that matches the current scale to guide feature recovery in the current upsampling block. Each upsampling block also receives coded features of the same scale from the corresponding downsampling block via skip connections. Finally, the reconstructed feature map is output through another convolutional layer, then transformed to the k-space domain by inverse Fourier transform. After data consistency constraint updates using an undersampling mask, it undergoes inverse Fourier transform and root mean square summation (coil merging) to obtain the preliminary reconstructed image. The preliminary reconstructed images are input into the first branch and the second branch respectively; In the first branch, the initially reconstructed image is sequentially processed by SAM2 for image embedding, a convolutional layer, group normalization, Gaussian error linear activation function, and multilayer perceptron, outputting prior structural and spatial information A; the multilayer perceptron includes a Gaussian error linear activation function and two fully connected linear layers located at the input and output ends of the activation function. In the second branch, the initially reconstructed image is sequentially passed through a convolutional layer, group normalization, and Gaussian error linear activation function to obtain shallow features B. Shallow features B are input into a detail enhancement network for enhancement. The detail enhancement network includes a parallel structural modeling module and a local detail module. The structural modeling module is used to model global structural relationships, and the local detail module is used to extract local texture and edge information. The outputs of the two modules are dynamically fused by an adaptive fusion attention module to obtain the backbone features C. The shallow feature B obtained from the second branch is added element-wise to the backbone feature C, and then multiplied element-wise with the prior information A output from the first branch. This multiplication is then further multiplied element-wise with the prior information A to obtain the preliminary fusion feature. The preliminary fusion feature is then concatenated with the backbone feature C along the channel dimension, and then processed sequentially through a convolutional layer, group normalization, and Gaussian error linear activation function. Finally, it is added element-wise with the fusion result before concatenation along the channel dimension (i.e., the backbone feature C) to obtain the enhanced fusion feature. Based on the enhanced fusion feature, a detail residual map is predicted through a convolutional layer. The detail residual map is then added pixel-wise with the preliminary reconstructed image to obtain the final reconstructed image.

[0033] In this embodiment, by jointly utilizing the adjacency information of frames and detail enhancement based on large model priors, undersampled MRI images can be better processed, significantly improving the quality and accuracy of the reconstructed images. By extracting auxiliary information from adjacent frames, combining it with global contextual features within a single frame, and dynamically generating prompts during the decoding and recovery stage to compensate for tissue texture, it helps to recover the structural loss and information blurring caused by undersampling, thereby mitigating aliasing artifacts and structural distortion caused by high-speed undersampling.

[0034] After obtaining the initial reconstructed image, by jointly modeling the global anatomical relationship and local fine-grained features of the image, and introducing prior information provided by the large model to guide the backbone features, the network's ability to process blurred boundaries and minute details can be improved, and targeted compensation for missing high-frequency information can be achieved, thereby generating a high-quality reconstructed image.

[0035] Specifically, a fully sampled MRI dataset is acquired as the MRI images to be processed, containing complete spatial domain information. Several techniques are employed for preprocessing to make it more suitable for subsequent training.

[0036] In step 1, optionally, the obtained multi-coil dataset can be digitized as follows: Step 1.1: Convert the original .mat multi-coil dataset to .h5 format to speed up I / O reading of the dataset and facilitate subsequent training.

[0037] Step 1.2: Centered on the current frame to be reconstructed, select the multi-coil k-space data input corresponding to the next 5 frames, represented as: ;in, This represents the multi-coil k-space data corresponding to the current frame to be reconstructed.

[0038] By introducing five adjacent frames of multi-coil k-space data centered on the current frame to be reconstructed, the modeling ability of structural changes and motion information at adjacent time points can be enhanced by utilizing the temporal continuity of cardiac magnetic resonance sequences. At the same time, the auxiliary information provided by adjacent frames helps to alleviate the structural loss and detail blur caused by undersampling of a single frame.

[0039] Step 1.3: Apply an undersampling mask to the adjacent 5 frames of multi-coil k-space data to obtain undersampled multi-coil k-space data, and estimate the coil sensitivity map based on this.

[0040] Step 1.4: Use the sensitivity map to convert the undersampled multi-coil k-space data into an image domain representation, and perform coil merging to obtain the initial reconstructed features, which can be represented as: ; in, This represents the initial reconstructed features at the current time t. This indicates summing over all coils. Indicates the total number of coils. This indicates that the inverse Fourier transform is performed on the undersampled k-space of the coil. This represents the undersampled k-space data of the c-th coil at time t. This indicates element-wise multiplication. This represents the conjugate of the sensitivity plot corresponding to the c-th coil.

[0041] Step 2: Perform coarse reconstruction on the initial reconstruction features to obtain a preliminary reconstructed image.

[0042] Step 2.1: Initial reconstruction features The input encoder performs multi-scale feature extraction to obtain multi-layer coded features, which can be represented as: ; in, Indicates encoder mapping, , , These represent the coding features at different scales.

[0043] Step 3: Based on the context artifact removal module (CARM), perform artifact suppression processing on the deep encoded features to obtain the artifact-suppressed deep features, which can be represented as: ; in, This indicates the context artifact removal module. This represents the deep features after artifact suppression.

[0044] The Context Artifact Removal (CARM) module in this scheme is positioned between the encoder and decoder to perform contextual modeling and artifact suppression on the deep features output by the encoder. Structurally, its working channel dimension is consistent with the deep feature channels in the network, with 120 channels, and it contains 3 context artifact removal blocks. This design is not a simple general convolution stack, but rather addresses the artifact suppression requirements in magnetic resonance undersampling reconstruction by performing contextual aggregation and residual enhancement on deep features, giving the deep representation stronger structural stability.

[0045] During the initial reconstruction, the encoder, context artifact removal module, cue decoder, and consistency constraint update unit are used to perform multi-scale feature extraction, artifact suppression, layer-by-layer recovery, and sampling consistency correction on the undersampled data, thereby obtaining the initial reconstructed image.

[0046] During the initial undersampling process and sensitivity map estimation, this scheme considers effectively providing a temporal continuity context, therefore selecting five adjacent frames. Applying an undersampling mask and estimating the coil sensitivity map based on the data from these five adjacent frames provides a physical consistency constraint for subsequent image domain reconstruction.

[0047] Table 1 illustrates that the five adjacent frames selected in this scheme represent a relatively ideal result.

[0048] Table 1. Performance metrics of the model at different frame rates

[0049] Step 4: Update the k-space based on mask constraints, and merge it with the coil after inverse Fourier transform to achieve preliminary image reconstruction.

[0050] Step 4.1: Input the artifact-suppressed deep features into the decoder, and combine them with the prompt block to perform layer-by-layer feature recovery, obtaining the reconstructed feature map, which can be represented as: ; ; in, Indicates the first The prompt features generated by the layer prompt module Indicates the first Layer prompt module mapping, Indicates the number after recovery Layer decoding features, Indicates the first Layer decoding restores the mapping. This represents the skip connection feature at the corresponding scale.

[0051] The PromptBlock module in this scheme is used for layer-by-layer feature recovery during the decoding stage. Its function is to adaptively weight and combine multiple learnable cue components based on global statistical information of the input features at the current decoding stage, generating a cue feature map that matches the current scale. This cue feature map then participates in the recovery process along with the current decoded features and skip connection features at the corresponding scale. The internal structure includes: a learnable cue parameter library, a linear mapping unit for generating cue weights, and a convolutional unit for fusing cue features.

[0052] The prompting module is configured in a multi-scale manner, with a prompt length of 5 and prompt feature dimensions of 72, 48, and 24, corresponding to the prompting guidance requirements of different decoding levels; the prompt space dimensions are 16, 32, and 64, respectively. The purpose of this structure is to enable the decoder to introduce structural priors and contextual priors that match the current scale during the reconstruction process, thereby more effectively suppressing undersampling artifacts and improving structural reconstruction capabilities.

[0053] Step 4.2: Combine the mask to perform consistency constraint update on the current k-space and obtain the preliminary reconstructed image, which can be represented as: ; ; in, This represents the updated k-space data. The data at time t is the multi-coil k-space data after undersampling. Represents the consistency constraint weight. Indicates the undersampling mask. This indicates element-wise multiplication. Represents the reference observation k-space, This represents the regularization update term generated by the reconstructed network. This indicates the initial reconstructed image. This indicates a coil merging operation. This indicates that an inverse Fourier transform is performed on undersampled k-space data.

[0054] The reconstructed feature map obtained in step 4.1 is transformed to the k-space domain using a Fourier transform to obtain the network-predicted k-space data for the current frame. In the frequency domain, a consistency constraint update is performed based on the current k-space prediction result, the reference observed k-space, the undersampled mask, and the consistency constraint weights. The prediction bias at the sampled positions is corrected, and the missing information at the undersampled positions is compensated by the regularization update term generated by the reconstruction network, thus obtaining the updated k-space data. The sampled positions are used to maintain consistency with the original observed data, and the undersampled positions are used to preserve the network's ability to recover missing frequency domain information. The updated k-space data is then subjected to an inverse Fourier transform and coil merging to obtain the preliminary reconstructed image.

[0055] Step 5: Extract the prior structure and spatial information of the preliminary reconstructed image, input the preliminary reconstructed image into SAM2 for image embedding representation. Output the prior structure and spatial information, which is used to guide subsequent refined reconstruction, and can be represented as: ; in, Represents the embedding map of SAM2. Represents prior structural and spatial information. This indicates a preliminary image reconstruction.

[0056] SAM2, as a pre-trained visual foundation model, possesses strong general representation capabilities. It can extract relatively stable structural boundaries, regional layouts, and spatial hierarchy information from initially reconstructed images, thus forming structural and spatial priors. For common problems in undersampled magnetic resonance reconstruction tasks, such as blurred boundaries, missing textures, and loss of local high-frequency information, prior information can provide effective guidance for subsequent detailed reconstruction, thereby improving detail recovery capabilities and structural fidelity.

[0057] This scheme, based on freezing SAM2, sets up a prior adaptation structure after the SAM2 output to make the general visual embeddings output by SAM2 applicable to magnetic resonance reconstruction tasks. The prior adaptation structure includes an embedding projection layer, a conditional modulation layer, and a fusion layer. The embedding projection layer is used to perform channel mapping on the high-dimensional embedding features output by SAM2; the conditional modulation layer is used to generate modulation parameters based on the mapped embedding features and conditionally modulate the image backbone features; the fusion layer is used to concatenate and convolve the modulated image features and the embedding features to generate prior guiding features for use by the detail enhancement module.

[0058] Step 6: Enhance the details of the preliminary reconstructed image under the guidance of prior information to obtain the final reconstructed image: Step 6.1: Model the global structural relationships using the Structural Modeling Module (SMB), extract local texture and edge information using the Local Details Module (LDB), and fuse them using Adaptive Fusion Attention Blocks (AFAB) to obtain the backbone features, which can be represented as: ; ; in, This represents the structural modeling module. This indicates a local detail module. This represents the Adaptive Fusion Attention Block (AFAB). Indicates the main characteristics, This indicates a preliminary image reconstruction.

[0059] The Structural Modeling Module (SMB) in this scheme is used for global structural relationship modeling in the detail enhancement stage, primarily serving the structural prior extraction and overall contour enhancement of the initial reconstruction results. This module enhances long-distance dependency representation capabilities and highlights the overall spatial layout and structural continuity of the target image by performing deep feature transformation and structural relationship modeling on the coarsely reconstructed image. The module adopts a structure based on convolutional feature extraction combined with state modeling units, with 112 channels, 4 state building modules, and 16 state dimensions. This design is suitable for scenarios in magnetic resonance reconstruction with high requirements for global structural consistency, providing stable structural constraints for subsequent local detail enhancement.

[0060] The Local Detail (LDB) module in this scheme is mainly used to extract local texture and edge information, focusing on compensating for the loss of local high-frequency details caused by undersampling. This module does not primarily focus on global structure modeling, but instead enhances edge response, texture restoration, and local region detail compensation capabilities through local convolutional paths. This LDB module works in parallel with the structure modeling module, structurally employing a combination of local convolutional extraction and residual output, with 112 channels, ultimately outputting a single-channel detail residual map.

[0061] The purpose of this module is to enable the network to restore the overall structure without ignoring local textures and edge details, thereby avoiding an overly smooth reconstruction result.

[0062] The Adaptive Fusion Attention Block (AFAB) in this scheme is used to dynamically fuse global structural features and local detail features. This module does not simply concatenate the two types of features directly, but rather adaptively adjusts the contribution ratio of different feature branches based on prior information and the current feature state through mechanisms such as conditional modulation, feature fusion, and gated residuals. The core structure of this module includes: a conditional modulation unit; a feature fusion unit; and a gated residual output unit. The prior embedding branch has 80 projection channels, the backbone has 112 channels, and the gate value is set to 0.1. This design allows the network to dynamically determine whether to prioritize global structure or local texture based on the structural complexity and detail richness of different regions, thereby improving the adaptability of the fusion and the precision of the reconstruction results.

[0063] Step 6.2, based on prior structural and spatial information Guided by this, the core features are fused with prior information to obtain enhanced fused features, which can be represented as: ; in, This indicates a priori guided fusion mapping. Indicates enhanced fusion features, Indicates the main characteristics.

[0064] Step 6.3: Predict the detail residual map based on the enhanced fusion features, and perform residual fusion with the preliminary reconstructed image to obtain the final reconstructed image, which can be represented as: ; ; in, Represents the residual prediction mapping. Representing the detailed residual plot, Represents the residual fusion coefficient. This represents the final reconstructed image.

[0065] The above process is achieved through an image reconstruction system for accelerated magnetic resonance imaging, which includes a data input module, an undersampling and sensitivity map estimation module, a preliminary reconstruction module, a priori embedding module, a detail enhancement and fusion module, and an output module.

[0066] The data input module is used to input raw multi-coil magnetic resonance data, convert multi-coil data in .mat format to .h5 format, and extract multi-coil k-space data from the five adjacent frames centered on the current frame to be reconstructed.

[0067] The undersampling processing and sensitivity map estimation module applies an undersampling mask to five adjacent frames of multi-coil k-space data to obtain undersampling multi-coil k-space data and the corresponding mask, and estimates the coil sensitivity map based on the undersampling multi-coil k-space data and the corresponding mask.

[0068] The preliminary reconstruction module converts undersampled multi-coil k-space data into initial reconstruction features in the image domain based on the sensitivity map, and performs preliminary reconstruction on the initial reconstruction features. The preliminary reconstruction module includes an encoder, a context artifact removal module, a cue decoder, and a consistency constraint update unit. The encoder is used to extract multi-scale features from the initial reconstruction features, the context artifact removal module is used to suppress artifact interference in deep features, the cue decoder is used to perform layer-by-layer feature recovery in combination with cue features, and the consistency constraint update unit is used to perform consistency constraint update on the sampled positions in k-space in combination with a mask, and outputs the preliminary reconstructed image.

[0069] The prior embedding module is used to perform SAM2 image embedding on the initially reconstructed image to obtain prior structural and spatial information.

[0070] The detail enhancement and fusion module enhances the details of the initially reconstructed image and achieves feature fusion under the guidance of prior structural and spatial information. The detail enhancement and fusion module includes a structural modeling module, a local detail module, an adaptive fusion attention module, and a residual prediction unit. Among them, the structural modeling module is used to model the global structural relationship, the local detail module is used to extract local texture and edge information, the adaptive fusion attention block is used to adaptively fuse global structural features and local detail features, and the residual prediction unit is used to output a detail residual map.

[0071] The output module is used to perform residual fusion between the detail residual map and the preliminary reconstructed image to obtain the final reconstructed image and output it.

[0072] Furthermore, during model training, appropriate data preprocessing and data augmentation methods can be applied to the input data according to different data types and reconstruction task requirements to improve the model's adaptability to different sampling conditions and image distributions.

[0073] During training, model parameters can be optimized using a single loss function or a combination of multiple loss functions to improve model training stability and final reconstruction performance.

[0074] Through the above implementation method, joint reconstruction of undersampled multi-coil k-space data of five adjacent frames can be achieved, and image detail enhancement can be completed under the guidance of prior structure and spatial information, thereby improving the structural fidelity, texture clarity and overall reconstruction quality of the reconstructed image.

[0075] To verify the effectiveness of this solution, four NVIDIA 5880 (48GB) GPUs and the popular deep learning library PyTorch were used for verification.

[0076] During the validation process, the initial reconstruction learning rate was set to 1e-4 and the number of training epochs was set to 50. The final reconstruction learning rate was set to 2e-5 and the number of training epochs was set to 30. The AdamW optimizer was used to update the model parameters.

[0077] To evaluate the performance of this solution, the evaluation metrics used include: Metrics include Peak Signal-to-Noise Ratio (PSNR), which indicates how sharp an image is; The Structural Similarity Index (SSIM) is used to describe the structural similarity of images; Normalized mean square error (NMSE) measures the difference between the reconstructed image and the real cardiac MRI image.

[0078] These metrics allow for a comprehensive evaluation of the performance of the proposed image reconstruction method. Figure 2 The reconstruction results of the heart image and the corresponding error graph are presented when the speedup is 10, and the heart reconstruction results obtained by different methods are compared. Figure 2 This display shows the overall reconstruction results of various methods (SwinIR, AMIR, PromptIR, AdaRPNet, our method, and GT), the reconstruction results of local cardiac regions for each method, and the corresponding error plots, illustrating the differences between the reconstruction results and the actual results.

[0079] Figure 2 (a) shows the overall reconstruction results of the SwinIR method, the local cardiac region reconstruction results of the method, and the corresponding error map of the method; Figure 2 (b) shows the overall reconstruction results of the AMIR method, the local cardiac region reconstruction results of the method, and the corresponding error map. Figure 2(c) shows the overall reconstruction results of the PromptIR method, the local cardiac region reconstruction results of the method, and the corresponding error map. Figure 2 (d) shows the overall reconstruction results of the AdaRPNet method, the local cardiac region reconstruction results of the method, and the corresponding error map. Figure 2 (e) shows the overall reconstruction results of this method, the local cardiac region reconstruction results of this method, and the corresponding error map of this method; Figure 2 (f) shows the overall reconstruction results of the GT method, the local cardiac region reconstruction results of the method, and the corresponding error map.

[0080] Tables 2, 3, and 4 present the performance index scores of different methods under different conditions.

[0081] Table 2. Performance metrics when the acceleration factor is 4 times

[0082] Table 3. Performance metrics at an acceleration factor of 8 times

[0083] Table 4. Performance metrics at an acceleration factor of 10x

[0084] The existing algorithms used for comparison in Tables 2, 3, and 4 include: SwinIR stands for Image Restoration Using Swin Transformery.

[0085] AMIR: All-in-one medical image restoration via task-adaptive routing.

[0086] AdaRPNet stands for Adaptive all-in-one image restoration via frequency mining and modulation.

[0087] PromptIR stands for Prompting for All-in-One Blind Image Restoration, which is a cue-learning method for all-in-one blind image restoration.

[0088] Ground Truth (GT) is a non-degradable image obtained by standard reconstruction after full sampling by a magnetic resonance imaging (MRI) device. It serves as a benchmark for evaluating the reconstruction quality of other methods.

[0089] This paper calculates the performance metrics of each method at 4x, 8x, and 10x speedup factors. Although CMR image reconstruction tasks are more challenging at 8x and 10x speedup, the proposed method consistently outperforms existing methods across all metrics, demonstrating the best performance scores.

[0090] In summary, traditional CMR typically requires 30 minutes to an hour or more. Traditional cardiac MRI usually demands long acquisition times for high-quality imaging. The proposed method, by jointly reconstructing five adjacent multi-coil k-space data frames centered on the current frame to be reconstructed, maintains good image reconstruction quality even under high acceleration factors. This reduces the sampling volume, shortens image acquisition time, and improves MRI efficiency while ensuring imaging quality. In practical applications, the acquisition time can be reduced to less than 15 minutes.

[0091] Artifacts are a significant factor affecting diagnostic accuracy in cardiac magnetic resonance imaging (CMR) image reconstruction. This approach first employs a cue-based preliminary reconstruction module to perform multi-scale feature extraction, contextual artifact removal, cue recovery, and consistency constraint updates on undersampled data, resulting in a relatively complete preliminary reconstructed image. Subsequently, prior structural and spatial information is combined to enhance details and compensate for residuals, effectively suppressing artifacts caused by undersampling and improving the recovery of image edges and texture details. This leads to clearer and more reliable reconstructed images, contributing to improved diagnostic accuracy in cardiac magnetic resonance imaging (CMR).

[0092] Because the scan time is reduced, high-quality images can be obtained in a shorter time, thus reducing the need for patients to hold their breath. This significantly improves patient cooperation, allowing more patients to complete the examination smoothly. The shortened acquisition time and increased patient comfort have expanded the application of CMR to include patients with difficulty controlling their breathing and children.

[0093] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for accelerating image reconstruction in magnetic resonance imaging, characterized in that, Includes the following steps: The raw data acquired by the magnetic resonance imaging device is obtained. Taking the data frame corresponding to the current image to be reconstructed as the center, multiple adjacent data frames are selected, and each selected data frame is undersampled to obtain undersampled data frames. After feature extraction, artifact suppression and data consistency constraints, a preliminary reconstructed image is obtained. Input the preliminary reconstructed images into the first branch and the second branch respectively; In the first branch, a pre-trained visual base model is used to process the preliminary reconstructed image and output the prior information of the structural boundaries and spatial layout of different anatomical regions in the preliminary reconstructed image. In the second branch, global structural features and local texture edge features of the initially reconstructed image are extracted, and the global structural features and local texture edge features are adaptively fused to obtain the backbone features; The prior information output from the first branch is fused with the backbone features output from the second branch to obtain enhanced fusion features. Based on the enhanced fusion features, a detail residual map is predicted through a convolutional layer. The detail residual map is then superimposed on the preliminary reconstructed image to obtain the final reconstructed image. Specifically, the data consistency constraint is as follows: the reconstructed feature map obtained after artifact suppression is transformed to the k-space domain by Fourier transform, and combined with the mask used for undersampling, the data consistency constraint of the current frame data is updated, and then transformed to the image domain and merged to obtain the preliminary reconstructed image. Specifically, the global structural features and local texture edge features are adaptively fused. This is achieved by using an adaptive fusion attention block, which dynamically allocates the contribution weights of global structural features and local texture edge features based on the prior information and feature state of the current reconstructed region.

2. The image reconstruction method for accelerated magnetic resonance imaging as described in claim 1, characterized in that, Feature extraction, specifically, involves estimating the sensitivity map of the receiving coil of the magnetic resonance device based on the undersampled multi-frame data during the initial reconstruction, and using the sensitivity map to convert the undersampled multi-frame data to the image domain to obtain the initial reconstruction features.

3. The image reconstruction method for accelerated magnetic resonance imaging as described in claim 1, characterized in that, Artifact suppression, specifically: The initial reconstructed features obtained during the preliminary reconstruction are processed by the encoder to extract multi-scale features, resulting in multi-layer coded features; Deep coding features are selected from multi-layer coding features, and artifact suppression processing is applied to obtain artifact-suppressed deep features. The deep features after artifact suppression are restored layer by layer by the decoder and the cue feature map generated by the cue module that matches the current scale, to obtain the reconstructed feature map.

4. The image reconstruction method for accelerated magnetic resonance imaging as described in claim 3, characterized in that, The prompting module adaptively weights and combines multiple learnable prompting components based on the global statistical information of the input features in the current decoding stage, generating a prompting feature map that matches the current scale.

5. The image reconstruction method for accelerated magnetic resonance imaging as described in claim 1, characterized in that, Data consistency constraints are shown in the following formula: ； ； in, This represents the updated k-space data. The data at time t is the multi-coil k-space data after undersampling. Represents the consistency constraint weight. Indicates the undersampling mask. This indicates element-wise multiplication. Represents the reference observation k-space, This represents the regularization update term generated by the reconstructed network. This indicates the initial reconstructed image. This indicates a coil merging operation. This indicates that an inverse Fourier transform is performed on undersampled k-space data.

6. The image reconstruction method for accelerated magnetic resonance imaging as described in claim 1, characterized in that, In the second branch, global structural features and local texture edge features of the initially reconstructed image are extracted, specifically including: Model the global structural relationships to obtain global structural features; Local texture and edge information are extracted to obtain local texture edge features.

7. The image reconstruction method for accelerated magnetic resonance imaging as described in claim 1, characterized in that, The prior information output from the first branch is fused with the backbone features output from the second branch. Specifically, guided by the prior information, the backbone features and the prior information are fused at the feature level to obtain the enhanced fused features, as shown in the following formula: ； in, To guide the fusion mapping a priori To enhance fusion features, For prior structural and spatial information.

8. The image reconstruction method for accelerated magnetic resonance imaging as described in claim 1, characterized in that, The detailed residual image is superimposed on the preliminary reconstructed image. Specifically, the detailed residual image is multiplied by a preset residual fusion coefficient and then added pixel by pixel to the preliminary reconstructed image, as shown in the following formula: ； ； in, Representing the detailed residual plot, Represents the residual prediction mapping. To enhance fusion features, This represents the final reconstructed image. This indicates the initial reconstructed image. This represents the residual fusion coefficient.

9. An image reconstruction system for accelerating magnetic resonance imaging, used to implement the image reconstruction method as described in any one of claims 1-8, characterized in that, include: Magnetic resonance imaging (MRI) acquisition equipment is used to acquire raw data; A processor, connected to the magnetic resonance imaging acquisition device, is configured to perform the following steps: The raw data acquired by the magnetic resonance imaging device is obtained. Taking the data frame corresponding to the current image to be reconstructed as the center, multiple adjacent data frames are selected, and each selected data frame is undersampled to obtain undersampled data frames. After feature extraction, artifact suppression and data consistency constraints, a preliminary reconstructed image is obtained. Input the preliminary reconstructed images into the first branch and the second branch respectively; In the first branch, a pre-trained visual base model is used to process the preliminary reconstructed image and output the prior information of the structural boundaries and spatial layout of different anatomical regions in the preliminary reconstructed image. In the second branch, global structural features and local texture edge features of the initially reconstructed image are extracted, and the global structural features and local texture edge features are adaptively fused to obtain the backbone features; The prior information output from the first branch is fused with the backbone features output from the second branch to obtain enhanced fusion features. Based on the enhanced fusion features, a detail residual map is predicted, and the detail residual map is superimposed on the preliminary reconstructed image to obtain the final reconstructed image.

Images