Parallel MRI (Magnetic Resonance Imaging) reconstruction method based on scanning-specific triple attention improved residual network
By constructing a parallel MRI reconstruction method based on scan-specific triple attention improved residual network, the problems of long acquisition time and large data requirements in MRI technology are solved, achieving efficient and robust undersampling MRI reconstruction and improving image quality.
Patent Information
- Application Number
- CN202511690027.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-11-18
- Publication Date
- 2026-02-10
AI Technical Summary
Current MRI technology is limited by data sampling constraints, resulting in long acquisition times and the need for large amounts of full-sample data, especially under under-sampling conditions, leading to poor reconstruction quality.
A parallel MRI reconstruction method based on scan-specific triple attention improved residual network is adopted. By constructing triple attention modules (CAMA, PPA, AGCA) and a hybrid loss function, efficient reconstruction is performed using undersampled data from a single scan, combined with adaptive and multi-scale feature extraction.
It achieves efficient MRI reconstruction under undersampling conditions, improves reconstruction quality and robustness, and significantly enhances image detail preservation and artifact suppression.
Smart Images
Figure CN121505097A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The application belongs to the field of magnetic resonance imaging, and particularly relates to a scan-specific triple-attention improved residual network parallel MRI reconstruction method based on unsupervised learning. BACKGROUND
[0002] As an image technology without ionizing radiation, magnetic resonance imaging (MRI) can generate high-resolution images. However, the further development of MRI technology is currently mainly restricted by two factors. On the one hand, due to the fixed limitation of data sampling, the acquisition time of magnetic resonance imaging is slow, which often causes discomfort of patients and may lead to the generation of motion artifacts. On the other hand, supervised deep learning needs a large amount of full-sampling data. Although unsupervised or self-supervised deep learning methods have appeared to solve the limitations of supervised deep learning methods, they still need a large amount of under-sampling database. SUMMARY
[0003] The application aims to solve the problems of the current need for full-sampling data for supervised training and reconstruction and the long time of magnetic resonance imaging. A scan-specific triple-attention improved residual network parallel MRI reconstruction method is introduced, which is specially designed for efficient reconstruction from under-sampling images. It is a scan-specific deep learning method that only uses under-sampling data from a single scan for learning and reconstruction, without the need for a large amount of data set.
[0004] To achieve the above-mentioned purpose, the application comprises the following steps: 1) Obtain full-sampling multi-coil k-space data, and perform multiple operations on the full-sampling multi-coil k-space data through coil sensitivity decoding and encoding, two-dimensional Fourier transform operator, two-dimensional inverse Fourier transform operator, and under-sampling operator for selecting acquisition positions from the entire multi-coil k-space grid to obtain under-sampling multi-coil k-space data. Divide the initial sampling mask into disjoint sets . Wherein, is a fixed set used to define the validation loss. Then, the mask is randomly divided into K rounds; each round produces two disjoint masks , and satisfies . Wherein, is used as the input of the reconstruction network and performs data consistency (DC) operation, is a training mask used to define the training loss function. The under-sampling multi-coil k-space data can obtain high-precision coil sensitivity map and reconstructed image The undersampled multi-coil k-space data, the coil sensitivity map and the corresponding training mask , the validation mask and the initial sampling mask are respectively formed into a training set, a validation set and a test set.
[0005] 2) Constructing a scan-specific triple attention residual parallel MRI reconstruction network (SS-TARNet) based on the reconstruction network model with adaptive and multi-scale feature extraction.
[0006] 3) Constructing a hybrid loss function of the reconstruction network model ; 4) Using the training set and the validation set obtained in step 1) and the hybrid loss function obtained in step 3) to solve the optimal parameters of the reconstruction network model ; 5) Using the test set obtained in step 1) and the optimal parameters of the reconstruction network model obtained in step 4) to perform image reconstruction.
[0007] In the step 1), the full-sampling multi-coil k-space data is obtained, and the full-sampling multi-coil k-space data is subjected to multiple operations through coil sensitivity decoding and encoding, a two-dimensional Fourier transform operator, a two-dimensional inverse Fourier transform operator, and an undersampling operator for selecting acquisition positions from the entire multi-coil k-space grid to obtain undersampled multi-coil k-space data. The initial sampling mask is divided into disjoint sets . Wherein, is a validation mask, which is a fixed set used to define the validation loss. Then, the mask is randomly divided into K rounds; each round produces two disjoint masks , and satisfies . Wherein, is used as the input of the reconstruction network and performs a data consistency (DC) operation, is a training mask, which is used to define the training loss function. Finally, the undersampled multi-coil k-space data, the coil sensitivity map and the corresponding training mask , the validation mask and the initial sampling mask are respectively formed into a training set, a validation set and a test set. The specific method is: The full-sampling multi-coil k-space data of the knee is obtained from the magnetic resonance imaging instrument, and the full-sampling multi-coil k-space data is subjected to multiple operations through coil sensitivity decoding S -1 and a two-dimensional inverse Fourier transform operator F -1Obtain fully sampled multi-coil image data , Represents the field of complex numbers. Represents the number of pixels in a two-dimensional image. and These are the height and width of the two-dimensional image, respectively; then, an undersampling operator is used. P Coil sensitivity encoding S and two-dimensional Fourier transform operator F Obtain undersampled multi-coil k-space data Defined as , The undersampled k-space data of the coil is represented as , , M This indicates the number of undersampled points in the k-space data of a single coil. The initial sampling mask is obtained by using a uniform selection function. Decomposed into and Then proceed K Round partitioning, each round produces two disjoint subsets that satisfy... Then, the input for training the network can be obtained. and the input of the validation network ,in Finally, the undersampled multi-coil k-space data, coil sensitivity maps, and corresponding training masks are presented. Verification mask and initial sampling mask Form training sets separately Validation set and test set .
[0008] The reconstructed network model in step 2) integrates three complementary attention modules: Channel Aggregation Enhanced Multi-Scale Attention (CAMA), Parallelized Patch-Aware Attention (PPA), and Adaptive Graph Channel Attention (AGCA) to enhance detail preservation, artifact suppression, and semantic representation. The CAA module comprises a Lightweight Channel Aggregation (CA) module and an Efficient Multi-Scale Attention (EMA) module. The CA module adaptively redistributes channel features in a high-dimensional latent space to produce richer information representations. The EMA module performs efficient multi-scale aggregation using parallel 3×3 and 1×1 convolutional branches with grouped features. The EMA module recalibrates channel weights through global information encoding and enables cross-dimensional interactions across branches, modeling short-term and long-term dependencies and capturing pixel-level relationships. Following CA and EMA, gated multiplicative fusion synergistically enhances features, effectively preserving details and suppressing artifacts. The PPA module employs parallel multi-branch feature extraction, including local, global, and sequential convolutional branches, to capture information across scale and semantic levels, while the fusion step jointly preserves details and global context. Then, the PPA module applies channel attention and spatial attention to reweight channels and emphasize salient spatial regions, achieving adaptive enhancement. The synergy between the multi-branch design and dual attention enriches feature learning, improves robustness, and significantly enhances reconstruction quality. The AGCA module treats each channel as a graph node and uses a learnable adjacency matrix and graph convolution to model dependencies between channels, generating adaptive channel weights that highlight informative features while suppressing redundancy. This allows the graph structure to adapt to the data and extracts spatial information more accurately, thereby enhancing feature representation and MRI reconstruction performance across different scanning scenarios. A. Channel Convergence Enhanced Multiscale Attention (CAMA) Input feature tensor The tensors are obtained by passing the efficient multi-scale attention module (EMA) and the channel attention module (CA) respectively. and Finally, tensor After the Sigmod function and Multiply; B. Parallel Sensing Patch Module (PPA) The PPA module generates skip connection features through 1×1 convolutions and extracts multi-scale spatial features using a local-global attention module. In parallel, three levels of 3×3 convolutional layers progressively extract and sum local features. Then, all branch features are fused, and the feature response is enhanced sequentially through channel attention (ECA) and spatial attention (SA) modules. Finally, the output is processed by Dropout, batch normalization, and ReLU activation. This module achieves local receptive field expansion, multi-scale feature aggregation, and attention-guided feature optimization. C. Adaptive Graph Channel Attention Module (AGCA) First, the input feature tensor... Global average pooling is performed and dimensionality is reduced using 1×1 convolution to generate compressed channel descriptors; then, dynamic weight matrices are generated using 1D convolution and Softmax. , and the identity matrix and learnable residual matrix Combined into a hybrid correlation matrix ; through matrix multiplication After achieving inter-channel information exchange, the channel attention weights are increased in dimensionality through ReLU activation and 1×1 convolution, ultimately generating Sigmoid-normalized channel attention weights and incorporating them into the input feature tensor. Multiply.
[0009] The hybrid loss function in step 3) It consists of two parts, including Norm and Norm; the two components are combined by averaging to form the final mixed loss function. This design can simultaneously preserve the overall structure and restore details in tasks such as medical image reconstruction. Mathematically, this can be expressed as: ; in, It is a predicted value. It is the actual value.
[0010] In step 4), the optimal parameters of the triple attention-improved residual parallel MRI reconstruction network model for a specific scan are obtained using the Adam optimizer in deep learning. The network is trained and validated using the training and validation sets generated in step 1), and the mixed loss function in step 3) is minimized. Obtain the optimal parameters , This represents the network parameters that achieve the optimal undersampling reconstruction effect in a triple attention-improved residual parallel MRI reconstruction network model for a specific scan.
[0011] In step 5), the optimal parameters of the reconstructed network model obtained in step 4) are used, based on the test set obtained in step 1). Image reconstruction can be performed to obtain the reconstructed image. Finally, The reconstructed image obtained in step 1). Amplitude fusion is performed to obtain the final reconstructed image. ,Right now: ; in, This indicates that the modulus of the two reconstructed images are taken respectively.
[0012] Compared with existing technologies, the beneficial technical effects of this invention are as follows: This invention proposes a parallel MRI reconstruction method based on a scan-specific triple attention improved residual network. This method integrates three efficient attention modules, realizing multi-scale fusion and detail restoration between images. Furthermore, this invention uniquely utilizes the reconstruction results obtained from two different networks for amplitude fusion output, resulting in better reconstruction effects and thus greatly improving reconstruction performance in undersampled MRI scenarios. Attached Figure Description
[0013] Fig. 1 This is the overall network architecture diagram of the present invention; Fig. 2 This section shows the reconstructed images and their error maps for different methods on the knee dataset, using a 5×1 DRU (speed factor of 5, sampling mode of one-dimensional random sampling). The first row contains four images: the fully sampled image and the reconstructed images obtained through three different reconstruction networks, respectively. The second row contains four images: the sampling mask for 5×1 DRU and the error maps corresponding to the reconstructed images in the first row, respectively. Detailed Implementation
[0014] The following embodiments, in conjunction with the accompanying drawings, will further illustrate the present invention. The described embodiments are merely some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.
[0015] Example 1: As Figs. 1-2 As shown, the embodiments of the present invention include the following steps: Step 1: Acquire fully sampled multi-coil k-space data. Then, perform multiple operations on the fully sampled multi-coil k-space data using coil sensitivity decoding and encoding, a 2D Fourier transform operator, a 2D inverse Fourier transform operator, and an undersampling operator to select sampling positions from the entire multi-coil k-space grid to obtain undersampled multi-coil k-space data. Set the initial sampling mask... Divide into disjoint sets .in, The verification mask is a fixed set used to define the verification loss. Subsequently, the mask... conduct K The rounds are randomly assigned; each round generates two disjoint masks. and satisfy .in, Used as input to rebuild the network and perform data consistency (DC) operations. The training mask is used to define the training loss function. High-precision coil sensitivity maps and reconstructed images can be obtained by passing undersampled multi-coil k-space data through the IMJENSE network. Finally, the undersampled multi-coil k-space data, coil sensitivity map, and corresponding training mask are combined. Verification mask and initial sampling mask The specific methods for forming the training set, validation set, and test set are as follows: Full-sample multi-coil k-space data of the knee is acquired from a magnetic resonance imaging (MRI) scanner, and decoded using coil sensitivity. S -1 and the two-dimensional inverse Fourier transform operator F -1 Obtain fully sampled multi-coil image data , Represents the field of complex numbers. Represents the number of pixels in a two-dimensional image. and These are the height and width of the two-dimensional image, respectively; then, an undersampling operator is used. P Coil sensitivity encoding S and two-dimensional Fourier transform operator F Obtain undersampled multi-coil k-space data Defined as , The undersampled k-space data of the coil is represented as , , M This indicates the number of undersampled points in the k-space data of a single coil. The initial sampling mask is obtained by using a uniform selection function. Decomposed into and Then proceed K Round partitioning, each round produces two disjoint subsets that satisfy... Then, the input for training the network can be obtained. and the input of the validation network ,in Finally, the undersampled multi-coil k-space data, coil sensitivity maps, and corresponding training masks are presented. Verification mask and initial sampling mask Form training sets separately Validation set and test set .
[0016] Step 2: Construct a scan-specific triple attention-based improved residual parallel MRI reconstruction network (SS-TARNet), a reconstruction network model with adaptive and multi-scale feature extraction. This reconstruction network integrates three complementary attention modules: Channel Aggregation Enhanced Multi-Scale Attention (CAMA), Parallelized Patch-Aware Attention (PPA), and Adaptive Graph Channel Attention (AGCA) to enhance detail preservation, artifact suppression, and semantic representation. The CAA module comprises a lightweight channel aggregation module (CA) and an efficient multi-scale attention module (EMA). The CA module adaptively redistributes channel features in a high-dimensional latent space to produce richer information representations. The EMA module performs efficient multi-scale aggregation using parallel 3×3 and 1×1 convolutional branches with grouped features. The EMA module recalibrates channel weights through global information encoding and enables cross-dimensional interactions across branches, modeling short-term and long-term dependencies and capturing pixel-level relationships. After CA and EMA, gated multiplicative fusion synergistically enhances features, effectively preserving details and suppressing artifacts. The PPA module employs parallel multi-branch feature extraction, including local, global, and sequential convolutional branches, to capture information across scales and semantic levels, while the fusion step jointly preserves details and global context. The PPA module then applies channel attention and spatial attention to reweight channels and emphasize salient spatial regions, achieving adaptive enhancement. The synergy between the multi-branch design and dual attention enriches feature learning, improves robustness, and significantly enhances reconstruction quality. The AGCA module treats each channel as a graph node and uses a learnable adjacency matrix and graph convolution to model inter-channel dependencies, generating adaptive channel weights that highlight informative features while suppressing redundancy. This adapts the graph structure to the data and extracts spatial information more accurately, thereby enhancing feature representation and MRI reconstruction performance across different scanning scenarios. A. Channel Convergence Enhanced Multiscale Attention (CAMA) Input feature tensor The tensors are obtained by passing the efficient multi-scale attention module (EMA) and the channel attention module (CA) respectively. and Finally, tensor After the Sigmod function and Multiply; B. Parallel Sensing Patch Module (PPA) The PPA module generates skip connection features through 1×1 convolutions and extracts multi-scale spatial features using a local-global attention module. In parallel, three levels of 3×3 convolutional layers progressively extract and sum local features. Then, all branch features are fused, and the feature response is enhanced sequentially through channel attention (ECA) and spatial attention (SA) modules. Finally, the output is processed by Dropout, batch normalization, and ReLU activation. This module achieves local receptive field expansion, multi-scale feature aggregation, and attention-guided feature optimization. C. Adaptive Graph Channel Attention Module (AGCA) First, the input feature tensor... Global average pooling is performed and dimensionality is reduced using 1×1 convolution to generate compressed channel descriptors; then, dynamic weight matrices are generated using 1D convolution and Softmax. , and the identity matrix and learnable residual matrix Combined into a hybrid correlation matrix ; through matrix multiplication After achieving inter-channel information exchange, the channel attention weights are increased in dimensionality through ReLU activation and 1×1 convolution, ultimately generating Sigmoid-normalized channel attention weights and incorporating them into the input feature tensor. Multiply.
[0017] Step 3: Hybrid Loss Function It consists of two parts, including Norm and Norm; the two components are combined by averaging to form the final mixed loss function. This design can simultaneously preserve the overall structure and restore details in tasks such as medical image reconstruction. Mathematically, this can be expressed as: ; in, It is a predicted value. It is the actual value.
[0018] Step 4: Solving for the optimal parameters of the triple attention-based improved residual parallel MRI reconstruction network model for specific scans. The Adam optimizer in deep learning is used, and the network is trained and validated using the training and validation sets generated in step 1). This is achieved by minimizing the mixed loss function in step 3). Obtain the optimal parameters , This represents the network parameters that achieve the optimal undersampling reconstruction effect in a triple attention-improved residual parallel MRI reconstruction network model for a specific scan.
[0019] Step 5: Utilize the test set obtained in step 1) and the optimal parameters of the reconstructed network model obtained in step 4). Image reconstruction can be performed to obtain the reconstructed image. Finally, The reconstructed image obtained in step 1). Amplitude fusion is performed to obtain the final reconstructed image. ,Right now: ; in, This indicates that the modulus of the two reconstructed images are taken respectively.
[0020] The methods proposed in this invention, including IMJENSE, E2EVN, and others, were tested on a test dataset using a one-dimensional random undersampling mask (1DRU) and a two-dimensional Poisson disk mask (2DPU), with an acceleration factor of AF=5. The PSNR and SSIM values of the reconstructed images are shown in Table 1. Compared to the comparative methods IMJENSE and E2EVN, the proposed method significantly improves the quantization metrics PSNR and SSIM (except that the PSNR of the proposed method is slightly lower than that of IMJENSE when set to 5×1DRU). This invention offers a significant improvement in undersampling image reconstruction and has great potential for MRI reconstruction.
[0021] Table 1: Quantitative comparison of reconstruction performance of different methods on the knee dataset.
[0022] From Table 1, we can obtain: In the comparison of OURS and IMJENSE: when set to 5×2 DPU, the PSNR of the method of the present invention is about 1.0555 higher and the SSIM is about 0.0056 higher. When set to 5×1 DRU: the PSNR is slightly lower than IMJENSE, but the SSIM is about 0.0076 higher. In the comparison of OURS and E2EVN: when set to 5×2 DPU, the PSNR of the method of the present invention is about 11.1089 higher and the SSIM is about 0.0643 higher. When set to 5×1 DRU: the PSNR is about 0.2569 higher and the SSIM is about 0.004 higher.
[0023] This invention presents a network model with adaptive, multi-scale feature extraction based on a scan-specific triple attention-based improved residual parallel MRI reconstruction network. Furthermore, this invention uniquely employs adaptive modules and a multi-scale attention mechanism for adaptive multi-scale feature fusion, thereby significantly improving reconstruction performance in undersampled MRI scenarios.
[0024] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.
Claims
1. A parallel MRI reconstruction method based on scan-specific triple attention improved residual networks, characterized in that, Includes the following steps: 1) Acquire fully sampled multi-coil k-space data. Perform multiple operations on the fully sampled multi-coil k-space data using coil sensitivity decoding and encoding, a two-dimensional Fourier transform operator, a two-dimensional inverse Fourier transform operator, and an undersampling operator that selects the sampling position from the entire multi-coil k-space grid to obtain undersampled multi-coil k-space data. Then, use the initial sampling mask... Divide into disjoint sets ,in, For the verification mask, a fixed set is used to define the verification loss. Subsequently, the mask is... conduct K The rounds are randomly assigned; each round generates two disjoint masks. and satisfy ,in, Used as input for network reconstruction and to perform data consistency operations. The training mask, used to define the training loss function, is used to obtain high-precision coil sensitivity maps and reconstructed images from undersampled multi-coil k-space data through the IMJENSE network. The undersampled multi-coil k-space data, coil sensitivity map, and corresponding training mask are used. Verification mask and initial sampling mask The training set, validation set, and test set are respectively formed; 2) Construct SS-TARNet, a scan-specific triple attention-based improved residual parallel MRI reconstruction network, a reconstruction network model with adaptive and multi-scale feature extraction capabilities. 3) Constructing a hybrid loss function for the reconstruction network model ; 4) Using the training and validation sets obtained in step 1) and the hybrid loss function obtained in step 3) Solve for the optimal parameters of the reconstructed network model. ; 5) Using the test set obtained in step 1), reconstruct the optimal parameters of the network model obtained in step 4). Perform image reconstruction.
2. The parallel MRI reconstruction method based on scan-specific triple attention improved residual network according to claim 1, characterized in that: The specific method for step 1) is as follows: Full-sample multi-coil k-space data of the knee is acquired from a magnetic resonance imaging (MRI) scanner, and decoded using coil sensitivity. S -1 and the two-dimensional inverse Fourier transform operator F -1 Obtain fully sampled multi-coil image data , Represents the field of complex numbers. Represents the number of pixels in a two-dimensional image. and These are the height and width of the two-dimensional image, respectively; then, an undersampling operator is used. P Coil sensitivity encoding S and two-dimensional Fourier transform operator F Obtain undersampled multi-coil k-space data Defined as , The undersampled k-space data of the coil is represented as , , M This indicates the number of undersampled points in the k-space data of a single coil. The initial sampling mask is obtained by using a uniform selection function. Decomposed into and Then proceed K Round partitioning, each round produces two disjoint subsets that satisfy... Then, the input for training the network can be obtained. and the input of the validation network ,in Finally, the undersampled multi-coil k-space data, coil sensitivity maps, and corresponding training masks are presented. Verification mask and initial sampling mask Form training sets separately Validation set and test set .
3. The parallel MRI reconstruction method based on scan-specific triple attention improved residual network according to claim 1, characterized in that, The reconstructed network model in step 2) integrates three complementary attention modules: Channel Aggregation Enhanced Multi-Scale Attention (CAMA), Parallelized Patch-Aware Attention (PPA), and Adaptive Graph Channel Attention (AGCA). The CA module includes a lightweight channel aggregation module (CA) and an efficient multi-scale attention module (EMA). The CA module adaptively redistributes channel features in a high-dimensional latent space. The EMA module uses parallel 3×3 and 1×1 convolutional branches with grouped features to perform efficient multi-scale aggregation, recalibrates channel weights through global information encoding, and achieves cross-dimensional interaction across branches, modeling short-term and long-term dependencies and capturing pixel-level relationships. After CA and EMA, gated multiplicative fusion collaboratively enhances the features. The PPA module employs parallel multi-branch feature extraction, including local, global, and sequential convolutional branches, to capture information across scales and semantic levels. The fusion step jointly preserves details and global context. Then, channel attention and spatial attention are applied to reweight channels and emphasize salient spatial regions, achieving adaptive enhancement. The AGCA module treats each channel as a graph node and uses a learnable adjacency matrix and graph convolution to model the dependencies between channels, generating adaptive channel weights. A. Channel Convergence Enhanced Multiscale Attention CAMA Input feature tensor Tensors are obtained by passing the efficient multi-scale attention module (EMA) and the channel attention module (CA) respectively. and Finally, tensor After the Sigmod function and Multiply; B. Parallel Sensing Patch Module (PPA) The PPA module generates skip connection features through 1×1 convolutions, while simultaneously extracting multi-scale spatial features using a local-global attention module. In parallel, three levels of 3×3 convolutional layers progressively extract local features and sum them. Then, all branch features are fused, and the feature response is enhanced sequentially through channel attention (ECA) and spatial attention (SA) modules. Finally, the output is activated by Dropout, batch normalization, and ReLU. C. Adaptive Graph Channel Attention Module (AGCA) First, the input feature tensor... Global average pooling is performed and dimensionality is reduced using 1×1 convolution to generate compressed channel descriptors; then, dynamic weight matrices are generated using 1D convolution and Softmax. , and the identity matrix and learnable residual matrix Combined into a hybrid correlation matrix ; through matrix multiplication After achieving information exchange between channels, the channel attention weights are generated by ReLU activation and 1×1 convolution to increase their dimensionality. Finally, Sigmoid-normalized channel attention weights are generated and multiplied with the input features.
4. The parallel MRI reconstruction method based on scan-specific triple attention improved residual network according to claim 1, characterized in that, The hybrid loss function in step 3) It consists of two parts, including Norm and The norm, the two components are combined by averaging to form the final mixed loss function. This can be expressed mathematically as: ； in, It is a predicted value. It is the actual value.
5. The parallel MRI reconstruction method based on scan-specific triple attention improved residual network according to claim 1, characterized in that, In step 4), the optimal parameters of the reconstructed network model are solved. The Adam optimizer from deep learning is used to train and validate the network using the training and validation sets generated in step 1), and the mixed loss function in step 3) is minimized. Obtain the optimal parameters , This represents the network parameters that achieve the optimal undersampling reconstruction effect in a triple attention-improved residual parallel MRI reconstruction network model for a specific scan.
6. The parallel MRI reconstruction method based on scan-specific triple attention improved residual network according to claim 5, characterized in that, Using the test set obtained in step 1) and the optimal parameters of the reconstructed network model obtained in step 4). Image reconstruction can be performed to obtain the reconstructed image. Finally, The reconstructed image obtained in step 1). Amplitude fusion is performed to obtain the final reconstructed image. ,Right now: ； in, This indicates that the modulus of the two reconstructed images are taken respectively.