Magnetic resonance image reconstruction method and device based on cascaded Gaussian sputtering architecture

Abstract

The invention discloses a magnetic resonance image reconstruction method based on a cascade Gaussian sputtering architecture. The method comprises the following steps: taking under-sampled magnetic resonance image domain data as the input of a neural network; iteratively training the input in a neural network formed by a plurality of cascade reconstruction units, executing one-time global residual connection after each cascade reconstruction unit, and finally outputting a reconstructed magnetic resonance image; each cascade reconstruction unit sequentially comprises a Net module for extracting image domain characteristics, an FFT module for converting the image domain characteristics into frequency domain characteristics, a DC module for executing data consistency in a frequency domain and an IFFT module for converting the frequency domain characteristics into the image domain characteristics; wherein the Net module is formed by sequentially connecting a feature extraction network, a Gaussian sputtering network and an image reconstruction network.

Description

Technical Field

[0001] This invention relates to the fields of magnetic resonance accelerated imaging and deep learning, and in particular to a magnetic resonance image reconstruction method and apparatus based on a cascaded Gaussian sputtering architecture. Background Technology

[0002] Compared to other medical imaging methods such as computed tomography (CT), magnetic resonance imaging (MRI) is a radiation-free and safe clinical diagnostic method with advantages such as multi-contrast imaging, multi-planar imaging, and high resolution. However, its imaging speed is relatively slow, especially when acquiring signals in the phase-encoding direction, which requires an extremely long time. During the scan, any physiological movement of the patient can cause a series of problems such as image artifacts and contrast distortion, seriously affecting image quality.

[0003] The key to achieving rapid magnetic resonance imaging (MRI) is shortening the K-space fill time. On one hand, hardware upgrades can accelerate scanning speed, such as increasing the strength of the main magnetic field and the gradient magnetic field. On the other hand, the amount of K-space data required for image reconstruction can be reduced. However, hardware improvements are limited, and drastic changes in gradient encoding can easily cause patient discomfort. Therefore, current mainstream methods focus on reducing the amount of K-space data. According to the Nyquist sampling theorem, when a large amount of K-space (frequency domain) information is lost, the corresponding image becomes blurred, losing many organ and tissue details. Therefore, the challenge lies in finding a suitable reconstruction method that can help recover some detailed features from blurred MRI images.

[0004] Traditional magnetic resonance imaging (MRI) image reconstruction methods mainly include parallel imaging (PI) and compressed sensing (CS). PI uses multiple coils to acquire K-space data and combines coil sensitivity and gradient coding to reconstruct the image. Typical PI methods include sensitivity-encoded imaging (SENSE) and generalized autocalibrating partially parallel acquisitions (GRAPPA). PI's acceleration factor is generally no more than 3, and it has high requirements for coil arrangement. CS breaks the traditional Nyquist sampling theorem, sampling without relying on signal frequency. It only needs to utilize the principal components of the image to remove artifacts. Compared to PI, CS allows for a larger acceleration factor and is therefore widely used in fields such as neural imaging and cardiovascular imaging.

[0005] In recent years, with the development of deep learning (DL), more and more studies have combined computer science (CS) and DL to achieve the prediction of clear images from undersampled, blurry MRI images in a shorter time. However, current methods are mainly based on CNN, Transformer, and Mamba architectures. Although they comprehensively consider local features, global features, and overall computational resources, they all rely on the prediction of individual discrete pixels, lack interpretability, and produce unnatural feature transitions between adjacent organs, which can interfere with the diagnosis of diseases. Summary of the Invention

[0006] The purpose of this invention is to overcome the shortcomings of the prior art and provide a magnetic resonance image reconstruction method and apparatus based on a cascaded Gaussian sputtering architecture. First, features are extracted quickly through a densely connected network. Then, a Gaussian sputtering module is used to model discrete pixels as multiple Gaussian field representations. Finally, a fully connected network is used to map the feature channels to a specified dimension. It is worth noting that cascaded image reconstruction units are also used, which helps to accelerate the convergence speed of the neural network and reduce the network training time.

[0007] The objective of this invention is achieved through the following technical solution: A magnetic resonance image reconstruction method based on a cascaded Gaussian sputtering architecture includes: Undersampled magnetic resonance image domain data is used as input to the neural network; The input is iteratively trained in a neural network consisting of several cascaded reconstruction units, and a global residual connection is performed after each cascaded reconstruction unit to finally output the reconstructed magnetic resonance image. Each of the cascaded reconstruction units sequentially includes: a Net module for extracting image domain features, an FFT module for converting image domain features into frequency domain features, a DC module for performing data consistency in the frequency domain, and an IFFT module for converting frequency domain features into image domain features; wherein, the Net module is composed of a feature extraction network, a Gaussian sputtering network, and an image reconstruction network connected in sequence.

[0008] Furthermore, firstly, missing K-space magnetic resonance data and corresponding fully sampled K-space data are collected, and undersampled magnetic resonance image domain data and fully sampled magnetic resonance image domain data are obtained through inverse Fourier transform (IFFT). During iterative training, the prediction target of the neural network is the fully sampled magnetic resonance image domain data. The imaginary and real parts of the undersampled magnetic resonance image domain data are stored separately, with a feature dimension of 2.

[0009] Furthermore, the feature extraction network is a densely connected convolutional neural network, including a feature pre-extraction module, several dense convolutional blocks, a feature fusion module, and residual connections arranged sequentially; wherein, each dense convolutional block uses dilated convolution to expand the receptive field, the convolutional kernel size is 3, the stride is 1, the dilation rate includes 1, 2, 4, and 4, and a channel attention mechanism is introduced in the channel dimension.

[0010] Furthermore, the Gaussian sputtering network represents each pixel in the image features output by the feature extraction network as a learnable two-dimensional Gaussian distribution, and uses the value of each pixel as the initial amplitude of the Gaussian kernel, and restricts it to a non-negative value through a non-negative activation function; when several Gaussian distributions overlap in space, the color features are obtained by alpha mixing, and the continuous feature field is obtained by combining the mean, covariance matrix and position information.

[0011] Furthermore, after the image features output by the feature extraction network, the SegmentationHead function is set to map each pixel to a preset multi-class, with each class corresponding to an initialized Gaussian kernel, and Gumbel-Softmax is used to promote training stability and convergence.

[0012] Furthermore, the image reconstruction network is a multilayer perceptron containing 4 layers, with 256 nodes in the hidden layers, used to map high-dimensional image features to a specified image feature dimension to obtain the predicted magnetic resonance image.

[0013] Furthermore, the DC module performs data consistency between the undersampled magnetic resonance image domain data and the predicted magnetic resonance image in the frequency domain. That is, the originally missing K-space data is directly filled by the predicted value of the Net module, while the originally existing K-space data is replaced by the weighted value of its original value and the predicted value of the Net module after prediction.

[0014] The present invention also provides a magnetic resonance image reconstruction device based on a cascaded Gaussian sputtering architecture, comprising: The input unit is used to input undersampled magnetic resonance image domain data into the neural network for iterative training; The training unit is used to iteratively train the input through several cascaded reconstruction units, and to perform a global residual connection after each cascaded reconstruction unit, ultimately outputting the reconstructed magnetic resonance image. Each of the cascaded reconstruction units comprises, in sequence: The Net module consists of a feature extraction network, a Gaussian sputtering network, and an image reconstruction network connected sequentially, and is used to extract the input image domain features; The FFT module is used to convert image domain features into frequency domain features; The IFFT module is used by the DC module to perform data consistency in the frequency domain and to convert frequency domain features into image domain features.

[0015] Preferably, the present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the magnetic resonance image reconstruction method based on the cascaded Gaussian sputtering architecture.

[0016] Preferably, the present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the magnetic resonance image reconstruction method based on the cascaded Gaussian sputtering architecture.

[0017] Compared with the prior art, the beneficial effects of the technical solution of the present invention are: 1. A cascaded reconstruction unit (Net module—FFT module—DC module—IFFT module—residual) is employed. In the frequency domain, the DC module explicitly constrains measurement consistency, suppressing artifact accumulation. Multi-cascade iterative optimization approximates traditional iterative reconstruction methods, accelerating convergence and stabilizing training. Residual connections alleviate gradient vanishing. This addresses issues such as frequency / image domain mismatch and ghosting artifacts caused by undersampling, as well as the difficulties in training and slow convergence of deep networks.

[0018] The 2.NET module employs a densely connected CNN feature extraction network (containing a feature pre-extraction module, three dense convolutional blocks, a feature fusion module, and residual connections). Dense connections promote feature reuse and gradient flow. Multiple local feature blocks are then fused, preserving details while integrating context. Residual connections further stabilize deep training. This addresses issues such as lost details and structural blurring caused by undersampling, and the difficulty in simultaneously capturing features at different scales.

[0019] 3. Dense convolutional blocks utilize dilated convolution, which injects "holes" (zero values) into the standard convolutional kernel, allowing the kernel to cover a larger input area without changing its size. Each operation maintains the image resolution, with dilation rates of 1, 2, 4, and 4. This expands the effective receptive field (equivalent to 3, 5, 9, and 9 respectively) without increasing parameters or losing resolution, enabling the network to simultaneously consider local texture and mid-to-long-range anatomical context, improving coherence and boundary detection.

[0020] 4. The dense convolutional block also employs a channel attention mechanism, which combines channel average pooling with two fully connected layers to generate channel weights and constrain them to [0,1], highlighting important channels and suppressing noisy channels, thereby improving reconstruction quality and stability; and solving problems such as feature channel redundancy and insufficient weights for key anatomical / texture channels.

[0021] 5. The Gaussian sputtering network represents each pixel in the image features output by the feature extraction network as a learnable two-dimensional Gaussian distribution, and uses the value of each pixel as the initial amplitude of the Gaussian kernel, which is constrained to non-negative values ​​by a non-negative activation function. When several Gaussian distributions overlap in space, an alpha mixing method is used to obtain color features, and the mean, covariance matrix, and position information are combined to obtain a continuous feature field. Discrete pixels are transformed into a continuous learnable Gaussian field, and the neighborhood is naturally superimposed by the Gaussian kernel, resulting in smooth and interpretable spatial continuity. Alpha mixing can adaptively fuse at the overlap of multiple Gaussians, increasing visual smoothness.

[0022] 6. After the image features are output by the feature extraction network, the SegmentationHead function is set to map each pixel to a preset multi-class, with each class corresponding to an initialized Gaussian kernel, and then... GumbelSoftmax Promote training stability and convergence.

[0023] 9. The DC module performs data consistency in the frequency domain between the undersampled MRI image domain data and the predicted MRI image. Specifically, the missing K-space data is directly filled in by the predictions from the Net and FFT modules, while the existing K-space data is replaced by a weighted sum of its original value and the predictions from the Net and FFT modules. Explicit constraints on the sampled data in the frequency domain reduce overfitting and artifacts; the weighting mechanism is adjustable between noise reduction and fidelity preservation, balancing detail and stability.

[0024] 10. The combined effect of cascaded reconstruction units and Gaussian sputtering networks: the cascaded units gradually approximate the fully sampled solution, while the Gaussian field provides a continuous feature field; the superposition of the two restores details and natural transitions while ensuring fidelity, shortening training time and improving the final PSNR / SSIM. Attached Figure Description

[0025] Figure 1 A flowchart illustrating the implementation of a magnetic resonance image reconstruction method based on a cascaded Gaussian sputtering architecture according to an embodiment of the present invention is shown.

[0026] Figure 2 A schematic diagram of the overall network of a magnetic resonance image reconstruction method based on a cascaded Gaussian sputtering architecture according to an embodiment of the present invention is shown.

[0027] Figure 3 A schematic diagram of each dense convolutional block in the feature extraction network according to an embodiment of the present invention is shown.

[0028] Figure 4 A schematic diagram of a Gaussian sputtering network according to an embodiment of the present invention is shown.

[0029] Figure 5a and Figure 5b A schematic diagram illustrating experimental results on T2WI data of the IXI dataset according to an embodiment of the present invention is shown. Figure 5a This is a graph showing the results under a Gaussian one-dimensional sampling rate of 30%. Figure 5b This is the result image under Gaussian 2D sampling rate of 30%.

[0030] Figure 6 This is a schematic diagram of the structure of a computer device provided in an embodiment of this application. Detailed Implementation

[0031] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are only for explaining the present invention and are not intended to limit the present invention.

[0032] Example 1 This embodiment provides a magnetic resonance image reconstruction method based on a cascaded Gaussian sputtering architecture, such as... Figure 1 As shown, a large amount of magnetic resonance data with missing K-space was first acquired. K u and the corresponding fully sampled K-space data K f Undersampled magnetic resonance image domain data were obtained by inverse Fourier transform (IFFT). X u and fully sampled magnetic resonance image domain data X f ; to undersample magnetic resonance image domain data X u As input to the neural network, the sampling sequence, sampling trajectory, and sampling rate in the K-space need to be specified during data acquisition. (For example, this experiment used two sampling trajectories: one Gaussian one-dimensional and one Gaussian two-dimensional, with a sampling rate of 30% and a T2WI sampling sequence.) Undersampled magnetic resonance imaging (MRI) data is input into a neural network consisting of several cascaded reconstruction units for iterative training. A global residual connection is performed after each cascaded reconstruction unit, and the reconstructed MRI image is finally output. The training process uses the Adam optimizer with the L1 loss function, sets the number of training epochs to 30, initializes the learning rate to 0.0002, and decays the learning rate evenly over each epoch. The model size is 22.6M.

[0033] Each cascaded reconstruction unit includes, in sequence: a Net module for extracting image domain features, an FFT module for converting image domain features into frequency domain features, a DC module for performing data consistency in the frequency domain, and an IFFT module for converting frequency domain features into image domain features; the cascaded network structure can accelerate the convergence speed of the network.

[0034] The Net module consists of a feature extraction network, a Gaussian sputtering network, and an image reconstruction network connected sequentially. The feature extraction network is used to process undersampled magnetic resonance image domain data. X u Feature extraction is performed to learn information from a higher dimension; the Gaussian sputtering network assigns Gaussian kernels with different distributions to discretized pixels. After continuous optimization and learning, the predicted value at each location can be represented by the superposition of nearby Gaussian fields; the image reconstruction network is used to map complex feature maps to a specified dimension through a multilayer perceptron to obtain the predicted magnetic resonance image; the image reconstruction network is a multilayer perceptron with 4 layers, of which the number of nodes in the hidden layer is 256.

[0035] Figure 2 The diagram illustrates the specific structure of the cascaded reconstruction unit in this embodiment. The approach involves iterating the Net module, FFT module, DC module, IFFT module, and residual connection module once, followed by another iteration of the Net module, FFT module, DC module, IFFT module, and a global residual connection module. This embodiment iterates a total of N=3 times. Before the magnetic resonance image predicted by the Net module is input into the FFT module, it is normalized, and low-frequency information is moved to the center position. The Net module includes a feature extraction network, a Gaussian sputtering network, and an image reconstruction network. The FFT module converts the image into frequency domain data. The DC module performs data consistency in the frequency domain, filling in missing K-space data with the predicted values ​​from the Net and FFT modules, while existing K-space data is replaced by a weighted sum of its original value and the predictions from the Net and FFT modules. The IFFT module converts the frequency domain data into image domain data. The residual connection block prevents gradient vanishing due to excessive network depth.

[0036] Data consistency is often used in deep networks, and its formula is:

[0037] in, M This represents the sampling matrix corresponding to a certain sampling trajectory, where (x, y) represents the image's position coordinates. A value of 0 for a pixel indicates that the corresponding position in k-space was not sampled; conversely, a value of 0 indicates that the corresponding position in k-space was sampled. If a position in k-space was not initially sampled, after passing through the Net, FFT, and DC modules, that unsampled pixel will directly use the prediction data from the Net and FFT modules. Filling: If a certain location in k-space has already been sampled, then after passing through the Net module, FFT module, and DC module, the pixel at that point will be filled with the prediction data from the Net module and FFT module. And the original undersampled k-space data The weighted operation between the two is replaced to prevent noise from affecting network performance during deep network training. Weights ranging from 0 to 1 are used to adjust The extent of their contribution.

[0038] The feature extraction network consists of a densely connected convolutional neural network, which includes a feature pre-extraction module, four dense convolutional blocks, a feature fusion module, and residual connections. Figure 2 The diagram also shows an overall schematic of the feature extraction network according to an embodiment of the present invention. The undersampled magnetic resonance image domain data... X u As input, its input size is W×H×2 ,in, W, H These refer to the image's width and height, respectively, and the imaginary and real parts of the image are stored separately, thus the image's feature dimension is 2. The feature pre-extraction module is a 3×3 convolution that transforms the image's dimension from the original 2 to a higher 64-dimensional feature image, obtaining the feature image. f 0; The feature fusion module refers to obtaining local features after passing through 3 dense convolutional blocks. , , The features of each local feature are Then, these three feature maps are merged along the channel dimension to obtain the overall feature map. f 4 ，f The size of 4 is Then, a 3×3 convolution is used to obtain new features, reducing the feature dimension from 192 to 64, resulting in a new feature map. f 5; Residual connection refers to connecting feature maps f 5. Output of the feature pre-extraction module f Adding zeros prevents gradient vanishing in deeper layers of the network, resulting in the final output features of the feature extraction network. f 6.

[0039] Figure 3 A schematic diagram of a dense convolutional block according to an embodiment of the present invention is shown. Each dense convolutional block maintains a consistent structure. To overcome the limitation of small attention range in traditional convolutional neural networks, this embodiment uses dilated convolution, i.e., injecting "holes" (zero values) inside the standard convolutional kernel, thereby allowing the convolutional kernel to cover a larger input region while maintaining its size. For a size of... k × k The convolution kernel has a receptive field size of . k Assuming the inflation rate is r Then its equivalent receptive field size isk +( k -1)×( r -1). If the convolution size k The value is 3, and the step size is... stride The value is 1, and the image resolution remains unchanged in each operation. The dilation rates are 1, 2, 4, and 4, and the corresponding receptive field sizes are 3, 5, 9, and 9, respectively.

[0040] The following derivation shows how to keep the image resolution constant.

[0041] A complete convolution operation requires determining the window size k and the stride. stride and edge fill value padding Assuming the size of the input image is The size after convolution becomes:

[0042]

[0043] If the dilation rate of convolution is dilation stride of dilated convolution stride If the value is 1 and the original window size is k, then the expanded window size is... for Fill its edges. padding The size is Substituting this into the size formula for convolution operation, we get:

[0044]

[0045] As can be seen from the above derivation, the image resolution remains unchanged after each dilation convolution.

[0046] The detailed steps for dense convolution blocks are as follows: The feature map input to each dense convolutional block F 1 size is When subjected to dilation convolutions with dilation rates of 1, 2, 4, and 4, the following results are obtained sequentially. F 2. F 3. F 4 and F All dimensions of 5 are After the fourth dilated convolution, the result will be... F 1. F 2. F 3. F 4 and F 5. Perform channel merging to obtain a size of The feature map, then using 64 Perform a two-dimensional convolution operation on the convolution kernel of the current size. The feature map information is fused, and the feature map is then restored to its original state. Features F 6. Dimensions of the input feature map F 1. Maintain consistent size. This process can be formalized as follows:

[0047] The next step is to apply channel attention mechanisms and global residual connections. F 6. Further improve network performance, which mainly consists of four processes: (1) Perform average pooling on the feature image along the channel dimension so that each channel can be represented by a number; (2) Two fully connected networks are used to assign weights to the feature channels, and an activation function is added after each fully connected network. Finally, the range of the data is controlled between 0 and 1. (3) Multiply this weight value with the input feature image along the channel dimension to highlight the importance of different channels; (4) Combine the feature maps of the channel attention output and input. F 1. Perform addition to complete the global residual connection.

[0048] Gaussian sputtering networks can represent each pixel of an image as a 2D Gaussian distribution, i.e., dependent on the location mean. Covariance Matrix The output of the feature extraction network f Each pixel in the 6 is mapped to a learnable Gaussian field, and a ReLU function is added to prevent negative values. The value of each pixel serves as the initial amplitude of the Gaussian kernel. The variance of the variable in the X direction. The variance of the variable in the Y direction. It represents the covariance between two elements in the X and Y directions.

[0049] Figure 4 A schematic diagram of the Gaussian sputtering network according to this embodiment is shown. After the feature extraction network, the output of the feature extraction network is processed using a segmentation head technique (the SegmentationHead function, whose structure contains only one 3×3 convolution, with an upsampling coefficient of 1, so no upsampling operation is performed). f For each pixel in step 6, a classification prediction is performed, and the output channels of the 3×3 convolution in the SegmentationHead function are set to 100. Therefore, each pixel corresponds to 100 different classes (100 logits), and each class corresponds to a separate initialized Gaussian kernel. GumbelSoftmax This method improves the training effect of the network. (Note:) p Let be any coordinates in the given real space, and let be the center of the Gaussian field. Then the formula for the Gaussian field is: ; In the case of multiple overlapping Gaussian fields, using alpha By mixing, their feature values ​​are combined to obtain the mixed color features. c i .

[0050] ; Where α represents the "opacity" or "contribution" of each Gaussian field at its current location. v i This represents the value at each pixel location. More specifically: α=1 indicates that the Gaussian field is completely opaque at this location, completely blocking everything behind it; α=0 indicates that the Gaussian field is completely transparent at this location, contributing nothing to the final color; 0<α<1 indicates that the Gaussian field is partially transparent, allowing some of the color of objects behind it to show through. alpha Blend, and combine the color features of all overlapping Gaussian fields. c i They are weighted and averaged according to their opacity and visibility to obtain a smooth and visually reasonable blending result.

[0051] Within a continuous feature domain, the covariance matrix is ​​known. mean µ i The corresponding value is the normalized feature map coordinate. v i That is, the initial amplitude is v i It through alpha Color characteristics are obtained by mixing c i any position p The feature values ​​at a given location are determined by summing the function values ​​from different Gaussian distributions. Then, these features are mapped onto a learnable Gaussian field using selective Gaussian sputtering, assigning a Gaussian distribution to each pixel.

[0052] This formula can be understood as: given an opacity α, at any position... p actual eigenvalues ​​at (Actual amplitude) v Actual mean μ Actual covariance matrix It is the sum of the features of all nearby Gaussian kernels. Obtained.

[0053] Figure 5 illustrates the experimental results on the IXI dataset according to this embodiment. The experiment was conducted in two cases: one-dimensional Gaussian sampling with 30% sampling and two-dimensional Gaussian sampling with 30% sampling, as shown below. Figure 5a and Figure 5b ,in Figure 5a and Figure 5b The first column represents the gold standard result (predicted target) of the experiment; the second column represents the undersampled magnetic resonance image before reconstruction and its five-fold error map; the third column represents the reconstruction result of the baseline method (Unet method) and its five-fold error map; the fourth column represents the reconstruction result of the magnetic resonance image reconstruction method based on the cascaded Gaussian sputtering architecture in this embodiment after removing the Gaussian sputtering module and its five-fold residual map; the fifth column represents the reconstruction result of the magnetic resonance image reconstruction method based on the cascaded Gaussian sputtering architecture in this embodiment and its five-fold error map. Regardless of whether it is Gaussian one-dimensional 30% sampling or Gaussian two-dimensional 30% sampling, the PSNR / SSIM corresponding to the method of this invention (… Figure 5a and Figure 5b The data (corresponding to the data below the mean error plot) are better than the baseline method. Moreover, by comparing the ablation tests in the fourth and fifth columns, it can be found that the Gaussian sputtering module we used is effective in improving details.

[0054] This application also provides an application scenario: Specifically, the magnetic resonance image reconstruction method based on a cascaded Gaussian sputtering architecture provided in this embodiment can be applied to high-quality image reconstruction during accelerated magnetic resonance imaging in the medical field, obtaining clear clinical data to assist doctors in disease research. Examples include research on nervous system diseases, endocrine system diseases, tumor diseases, and respiratory system diseases. The magnetic resonance image reconstruction method based on a cascaded Gaussian sputtering architecture in this application maintains high image reconstruction quality even at high acceleration rates (low sampling rates). Because the amount of data required for K-space acquisition is small, the acquisition time is shorter, which not only improves patient comfort but also enhances the efficiency of disease diagnosis.

[0055] Example 2 Based on the same inventive concept, this application also provides a magnetic resonance image reconstruction apparatus based on a cascaded Gaussian sputtering architecture, which can be used to implement the method described in the above embodiments, as shown in the following embodiments. Since the principle of the reconstruction apparatus in solving the problem is similar to that of the magnetic resonance image reconstruction method, the implementation of this apparatus can refer to the aforementioned method implementation, and repeated details will not be elaborated further. As used below, the terms "unit" or "module" can refer to a combination of software and / or hardware that implements a predetermined function. Although the system described in the following embodiments is preferably implemented in software, hardware implementation, or a combination of software and hardware, is also possible and contemplated.

[0056] This embodiment provides a specific implementation of a reconstruction device capable of realizing a magnetic resonance image reconstruction method, specifically including the following: The input unit is used to input undersampled magnetic resonance image domain data into the neural network for iterative training; The training unit is used to iteratively train the input through several cascaded reconstruction units, and to perform a global residual connection after each cascaded reconstruction unit, ultimately outputting the reconstructed magnetic resonance image. Each of the cascaded reconstruction units comprises, in sequence: The Net module consists of a feature extraction network, a Gaussian sputtering network, and an image reconstruction network connected sequentially, and is used to extract the input image domain features; The FFT module is used to convert image domain features into frequency domain features; The IFFT module is used by the DC module to perform data consistency in the frequency domain and to convert frequency domain features into image domain features.

[0057] In one exemplary embodiment, a computer device is described, which may be a server or a terminal, and its specific structure can be referred to. Figure 6 The device mainly consists of a processor, memory, input / output (I / O) interfaces, and a communication interface. The processor, memory, and I / O interfaces are interconnected via a system bus, while the communication interface connects to the bus system through the I / O interface. The processor serves as the core for computation and control; the memory encompasses non-volatile storage media and RAM; the former stores the operating system, applications, and database, while the latter provides the environment for the operation of the former two. The database is specifically used to store undersampled magnetic resonance data. The I / O interface handles information exchange between the processor and external devices, and the communication interface enables network communication with external terminal devices. The device executes a specific program through the processor to implement a magnetic resonance image reconstruction method based on a cascaded Gaussian sputtering architecture. Those skilled in the art will understand that… Figure 6The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0058] Preferably, embodiments of this application also provide a specific implementation of an electronic device capable of implementing all steps in the magnetic resonance image reconstruction method based on a cascaded Gaussian sputtering architecture described in the above embodiments. The electronic device specifically includes the following: Processor, memory, communications interface, and bus; The processor, memory, and communication interface communicate with each other via a bus; the communication interface is used to realize information transmission between server-side devices, metering devices, and user-side devices.

[0059] The processor is used to call the computer program in memory. When the processor executes the computer program, it implements all the steps in the magnetic resonance image reconstruction method based on the cascaded Gaussian sputtering architecture in the above embodiments.

[0060] Embodiments of this application also provide a computer-readable storage medium capable of implementing all steps of the magnetic resonance image reconstruction method based on cascaded Gaussian sputtering architecture described in the above embodiments. The computer-readable storage medium stores a computer program that, when executed by a processor, implements all steps of the magnetic resonance image reconstruction method based on cascaded Gaussian sputtering architecture described in the above embodiments.

[0061] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and related data (such as data used for analysis, stored data and displayed data, etc.) involved in this application are all processed with the explicit authorization of the user or the full authorization of the relevant parties, and the collection, use and processing of all data strictly comply with applicable laws, regulations and industry standards.

[0062] It will be understood by those skilled in the art that all or part of the steps of the methods described in the above embodiments can be implemented by a computer program driving related hardware. Such programs are typically stored in a non-volatile computer-readable storage medium and, when executed, include the flows defined in the foregoing method embodiments.

[0063] Furthermore, the memories, databases, or other storage media mentioned in the embodiments of this application may cover at least one class of non-volatile and volatile memories. Examples of non-volatile storage media include, but are not limited to: read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical storage devices, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetoresistive memory (MRAM), ferroelectric memory (FRAM), phase-change memory (PCM), graphene memory, etc. Volatile memories include forms such as random access memory (RAM) or external cache memory, wherein RAM can be further divided into types such as static random access memory (SRAM) and dynamic random access memory (DRAM).

[0064] Regarding databases, the types of databases involved in this application may include at least one of relational databases and non-relational databases. Non-relational databases may also include distributed databases based on blockchain, and are not limited to the examples mentioned above.

[0065] Regarding the processor, the processor referred to in the various embodiments of this application may be a general-purpose processor, a central processing unit (CPU), a graphics processing unit (GPU), a digital signal processor (DSP), a programmable logic device, or a data processing logic unit built based on the principle of quantum computing, etc., and is not limited to these.

[0066] The technical features disclosed in the various embodiments of this application can be arbitrarily combined according to actual needs. To maintain the brevity of the description, not all possible combinations of features are listed here, but as long as these combinations do not have any contradictions in technical logic, they should all be considered within the scope of this specification.

[0067] It should be noted that this document illustrates the principles and implementation methods of this application through specific implementation examples. The above embodiments are only used to help understand the core ideas and methods of this application. For those skilled in the art, appropriate adjustments and changes can be made to the specific implementation forms and application scope without departing from the principles of this application. Therefore, the content of this specification should not be construed as a limitation of this application.

[0068] This invention is not limited to the embodiments described above. The above description of specific embodiments is intended to illustrate and explain the technical solutions of this invention. The specific embodiments described above are merely illustrative and not restrictive. Without departing from the spirit and scope of the claims, those skilled in the art can make many specific modifications based on the teachings of this invention, and these modifications all fall within the scope of protection of this invention.

Claims

1. A magnetic resonance image reconstruction method based on a cascaded Gaussian sputtering architecture, characterized in that, include: Undersampled magnetic resonance image domain data is used as input to the neural network; The input is iteratively trained in a neural network consisting of several cascaded reconstruction units, and a global residual connection is performed after each cascaded reconstruction unit to finally output the reconstructed magnetic resonance image. Each of the cascaded reconstruction units sequentially includes: a Net module for extracting image domain features, an FFT module for converting image domain features into frequency domain features, a DC module for performing data consistency in the frequency domain, and an IFFT module for converting frequency domain features into image domain features; wherein, the Net module is composed of a feature extraction network, a Gaussian sputtering network, and an image reconstruction network connected in sequence.

2. The magnetic resonance image reconstruction method based on a cascaded Gaussian sputtering architecture according to claim 1, characterized in that, First, missing K-space magnetic resonance data and corresponding fully sampled K-space data are collected. Then, undersampled and fully sampled magnetic resonance image domain data are obtained through inverse Fourier transform (IFFT). During iterative training, the prediction target of the neural network is the fully sampled magnetic resonance image domain data. The imaginary and real parts of the undersampled magnetic resonance image domain data are stored separately, with a feature dimension of 2.

3. The magnetic resonance image reconstruction method based on a cascaded Gaussian sputtering architecture according to claim 1, characterized in that, The feature extraction network is a densely connected convolutional neural network, which includes a feature pre-extraction module, several dense convolutional blocks, a feature fusion module and residual connections arranged in sequence. Each dense convolutional block uses dilated convolution to expand the receptive field. The kernel size is 3, the stride is 1, and the dilation rate includes 1, 2, 4 and 4. A channel attention mechanism is introduced in the channel dimension.

4. The magnetic resonance image reconstruction method based on a cascaded Gaussian sputtering architecture according to claim 1, characterized in that, The Gaussian sputtering network represents each pixel in the image features output by the feature extraction network as a learnable two-dimensional Gaussian distribution, and uses the value of each pixel as the initial amplitude of the Gaussian kernel, and restricts it to a non-negative value through a non-negative activation function; when several Gaussian distributions overlap in space, the color features are obtained by alpha mixing, and the mean, covariance matrix and position information are combined to obtain a continuous feature field.

5. The magnetic resonance image reconstruction method based on a cascaded Gaussian sputtering architecture according to claim 4, characterized in that, After the image features output by the feature extraction network, the SegmentationHead function is set to map each pixel to a preset multi-class, each class corresponds to an initialized Gaussian kernel, and Gumbel-Softmax is used to promote training stability and convergence.

6. The magnetic resonance image reconstruction method based on a cascaded Gaussian sputtering architecture according to claim 1, characterized in that, The image reconstruction network is a multilayer perceptron, which contains 4 layers and 256 nodes in the hidden layer. It is used to map high-dimensional image features to a specified image feature dimension to obtain the predicted magnetic resonance image.

7. The magnetic resonance image reconstruction method based on a cascaded Gaussian sputtering architecture according to claim 1, characterized in that, The DC module performs data consistency between the undersampled magnetic resonance image domain data and the predicted magnetic resonance image in the frequency domain. That is, the missing K-space data is directly filled by the predicted value of the Net module, while the existing K-space data is replaced by the weighted value of its original value and the predicted value of the Net module after prediction.

8. A magnetic resonance image reconstruction device based on a cascaded Gaussian sputtering architecture, characterized in that, include: The input unit is used to input undersampled magnetic resonance image domain data into the neural network for iterative training; The training unit is used to iteratively train the input through several cascaded reconstruction units, and to perform a global residual connection after each cascaded reconstruction unit, ultimately outputting the reconstructed magnetic resonance image. Each of the cascaded reconstruction units comprises, in sequence: The Net module consists of a feature extraction network, a Gaussian sputtering network, and an image reconstruction network connected sequentially, and is used to extract the input image domain features; The FFT module is used to convert image domain features into frequency domain features; The IFFT module is used by the DC module to perform data consistency in the frequency domain and to convert frequency domain features into image domain features.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the magnetic resonance image reconstruction method based on the cascaded Gaussian sputtering architecture as described in any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the computer program implements the steps of the magnetic resonance image reconstruction method based on a cascaded Gaussian sputtering architecture as described in any one of claims 1 to 6.

Images