End-to-end medical MRI image blind super-resolution method and system based on space-frequency parallel attention

Abstract

The invention discloses an end-to-end medical MRI (Magnetic Resonance Imaging) image blind super-resolution method and system based on space-frequency parallel attention. According to the method, CNN and Transform architectures are adopted, and a model synchronously performs spatial domain attention weight distribution and frequency domain attention enhancement on extracted features through a space-frequency parallel attention fusion module; the degradation adaptive estimation module dynamically learns unknown degradation parameters from fusion features based on the global modeling capability of Transform, does not need to manually preset a degradation model, and realizes adaptive fusion of double-domain features through a gating mechanism. Experiments show that the method can effectively utilize prior information and improve the medical MRI blind super-resolution effect.

Description

Technical Field

[0001] This invention belongs to the field of blind super-resolution technology, specifically relating to an end-to-end blind super-resolution method and system for medical MRI images based on spatial-frequency parallel attention. Background Technology

[0002] Magnetic resonance imaging (MRI) images, acquired through medical imaging technology, can visually display the internal tissue structures of the human body, providing a scientific and intuitive basis for disease diagnosis and treatment. Resolution, as a key parameter in medical imaging, directly affects the accuracy of medical diagnosis and treatment outcomes. Because pathological features in the early stages of disease are often subtle, and due to limitations in the performance of current medical equipment, the resolution of MRI images is relatively low, posing a challenge to disease diagnosis. Furthermore, the degradation in actual cases is often complex and unknown, a characteristic that presents a core challenge to blind super-resolution reconstruction of MRI images.

[0003] Currently, there are two main methods for improving the super-resolution of medical MRI images: First, upgrading medical imaging hardware by improving image sensor construction techniques or using larger sensors. However, hardware upgrades are closely related to physics, and technological innovation has a long cycle and requires substantial financial support. Second, using MRI image super-resolution techniques, which do not require hardware upgrades. Medical MRI image super-resolution techniques can be divided into traditional methods and deep learning methods. Traditional methods include interpolation-based methods, total variation regularization, sparse coding, and compressed sensing. However, traditional methods may lead to the loss of fine details, producing blurry and unrealistic results, thus reducing image quality. In contrast, deep learning methods have made significant progress in the field of MRI image super-resolution in recent years. By training neural networks, deep learning technology can learn mapping relationships from a large number of high-resolution and low-resolution image pairs, thereby more effectively recovering detailed information.

[0004] The rapid development of deep learning technology has greatly promoted the research of super-resolution methods for blind MRI images. In recent years, the field of medical MRI image super-resolution has developed rapidly. These methods typically employ convolutional neural networks (CNNs, such as U-Net and ResNet) or Transformer architectures to learn degradation mapping relationships from a large amount of paired low-resolution and high-resolution MRI image data. However, in real-world clinical scenarios, MRI image degradation is easily affected by many factors such as motion blur and noise. When the degradation pattern does not match the training assumptions, model performance drops sharply, resulting in problems such as detail distortion or increased artifacts. Currently, medical MRI blind image super-resolution methods often employ a two-stage architecture of degradation estimation and super-resolution reconstruction: first, image degradation parameters are estimated through an independent module, and then the estimation results are input into a non-blind super-resolution network to complete the reconstruction. However, biases in the degradation estimation module directly lead to a decrease in the accuracy of subsequent reconstruction results. Furthermore, the two-stage design cannot achieve end-to-end optimization, making it difficult to balance overall model adaptability and efficiency. Summary of the Invention

[0005] In view of the technical problems mentioned in the background, the purpose of this invention is to provide an end-to-end blind super-resolution method and system for medical MRI images based on spatial-frequency parallel attention.

[0006] To achieve the objectives of this invention, the technical solution provided by this invention is as follows: First aspect This invention provides an end-to-end blind super-resolution method for medical MRI images based on spatial-frequency parallel attention. By constructing a deep learning network that integrates spatial and frequency domain features, blind super-resolution reconstruction of low-resolution medical MRI images is achieved.

[0007] This method does not rely on prior knowledge of the image degradation process and directly outputs high-resolution results from low-resolution images end-to-end. After end-to-end training is completed, the test MRI images to be super-resolution are fed into the network for super-resolution, and finally the super-resolution results are output.

[0008] Preferably, the method includes the following steps: Step S1: Simulate the real degradation process of clinical MRI images by using a degradation formula to generate corresponding low-resolution MRI images from high-resolution MRI images, and construct training sample pairs; Step S2: Input the low-resolution MRI image into the model. First, extract the basic features such as the shallow bottom layer edge and texture of the image through the shallow layer extraction module composed of 3x3 convolution. Step S3: Input the shallow features extracted in step S2 into the Spatial-Frequency Parallel Attention (SFPAM) module to achieve synergistic enhancement of spatial and frequency domain features; Step S4: The features output in step S3 are fed into the degradation-aware fusion module DAFM to achieve deep interaction between spatial and frequency domain features. At the same time, the learnable filter dynamically adapts to the frequency domain features of different degradations without the need to preset the degradation model. Step S5: The optimized features from step S4 are fed into the reconstruction module, and a high-resolution MRI image is generated through upsampling. Step S6: Optimize network parameters using the Adam optimizer and loss function; Step S7: Input the test set images into the trained network and output the corresponding super-resolution high-resolution MRI images.

[0009] Preferably, step S1 is as follows: The high-resolution MRI images in the dataset undergo degradation processing, which includes: convolving the MRI images in the dataset with a Gaussian blur kernel, then transforming them to the frequency domain using Fourier transform (FFT), performing a truncation operation in the frequency domain, transforming them back to the spatial domain using inverse Fourier transform, and finally adding additive white Gaussian noise. This process is represented as follows: ; in, For high-resolution MRI images, These are low-resolution MRI images. This represents the convolution operation. Represents the Gaussian blur kernel. Represents the Fourier transform, which transforms an image from the spatial domain to the frequency domain. This represents the inverse Fourier transform, which converts data from the frequency domain to the spatial domain. This indicates a truncation operation. This represents Rice noise.

[0010] Preferably, step S3 is as follows: Step S301: Assume the feature map output by the shallow feature extraction module Size is ,in For training batches, The number of feature map channels. and These are the height and width of the feature map, respectively; Features are input into the channel attention module, where they are divided into... Two branches, one of which is fed into two parallel max-pooling and average-pooling layers to obtain a feature size of... of and The two features are then processed through a weight-sharing multilayer perceptron, added together, and finally activated by a ReLU function to obtain a feature of size . The feature weights, and then combined with Multiplying them together yields a value of 1. Channel attention optimization features This process can be represented as: ; ; ; in, This represents a max pooling operation. This represents an average pooling operation; MLP represents a multilayer perceptron; and ReLU represents the ReLU activation function. Step S302: Optimize the channel attention features output in step S301. First, the input is fed into the spatial domain branch, which consists of nested residual dense blocks (RRDBs). Each RRDB block consists of two residual dense blocks (RDBs). The final output of the spatial domain branch is then compared with the original input features using a weight multiplier of 0.2. Addition achieves residual connection, and this nested structure enhances the ability to capture details of spatial domain features. The process can be represented as follows: ; ; Step S303: Optimize the channel attention features output in step S301. The input is then fed into the frequency domain branch, serving as the input to the frequency domain attention module (FAM). First, the features are normalized using a LayerNorm layer, and then the number of feature channels is adjusted using a 1×1 convolution to obtain the feature values. The process is represented as follows: ; in, This represents a 1×1 convolution. This indicates LayerNorm normalization; Step S304: Then split it according to the channel dimension into , , Then, by using 3×3 depthwise separable convolution to enhance the correlation of local features, we obtain... , , The formula is as follows: ; ; ; ; in, For channel splitting operation, This represents depthwise separable convolution; Step S305: Adopt an image segmentation strategy to divide the image into blocks. , Divide the data into multiple local feature blocks of 8×8 size, and transform the feature blocks into... Feature map , ,in Number of blocks For single-piece dimensions; Step S306: Process the features after block division , Performing a two-dimensional Fast Fourier Transform (FFT) transforms the features from the spatial domain to the frequency domain. By introducing degradation parameters deg_r and deg_i, the impact of degradation on the frequency domain is simulated through product and subtraction, adapting to unknown degradation scenarios in blind super-resolution, and calculating in the frequency domain. and The attention weights are then transformed back to the spatial domain using an inverse fast Fourier transform (IFFT), and rearranged to restore the original feature sizes, resulting in the attention weight map in the frequency domain. The process is represented as follows: ; ; ; in, This represents a two-dimensional Fast Fourier Transform. This represents the inverse fast Fourier transform. This indicates that the inverse fast Fourier transform is in dimensions 4 and 5. This indicates the absolute value operation. Indicates the rearrangement feature size; Step S307: Adjust the frequency domain attention weights and The features are multiplied element-wise, and then the number of channels is restored by a 1×1 convolution to obtain the frequency domain optimized features. The process is represented as follows: ; Step S308: Analyze the spatial domain features output in step S302. and By concatenating along the channel dimension, cross-domain complementary features are obtained. The process is represented as follows: ; Step S309: Cross-domain complementary features output in step S308 Perform average pooling and max pooling operations on the channel dimension respectively, and then pool the features. and The channels are concatenated along the channel dimension, the number of channels is compressed using a 3×3 convolutional layer, and a spatial attention weight map is generated using a sigmoid activation function. This process can be represented as: ; ;

[0011] in, This represents a max pooling operation. This represents an average pooling operation. This represents the Sigmoid activation function; Step S310: Assign spatial attention weights to the graph. Compared with the original input features Element-wise multiplication is performed to enhance key region features, resulting in dual-domain enhanced features. : .

[0012] Preferably, step S4 is as follows: Step S401: Develop the dual-domain enhanced features output from step S3. The input degradation adaptive estimation module DSEM enables deep interaction between spatial and frequency domain features. First, the features are normalized using a LayerNorm layer to eliminate numerical differences across different channels and spatial locations, providing a stable feature distribution for subsequent domain-specific processing. The formula is as follows: ; Step S402: Process the normalized features The window is divided into 8×8 blocks, and then the number of feature channels is adjusted by 1×1 convolution. The blocks are then divided into 3 groups on average according to the channel dimension to obtain features Q, K, and V. Spatial attention weights within the window are calculated using multi-head attention. Finally, the window attention features are rearranged to restore the original size. The formula is as follows: ; ; ; ; in, Indicates the rearrangement feature size. This is a channel splitting operation, where C is the number of channels. Scaling factor This represents the softamax normalization function; Step S403: Process the normalized features Inputting a 1×1 convolution and a 3×3 depthwise separable convolution yields the enhanced result. , will feature Transform from the spatial domain to the frequency domain, and learn a learnable frequency domain weight in the frequency domain. The high-frequency characteristics of the degradation adaptation are then converted to the frequency domain. The formula is: ; ; ; in, For learnable frequency domain filtering parameters, This represents a two-dimensional Fast Fourier Transform. This represents the inverse fast Fourier transform; Step S404: Enhance the high-frequency features and Dot product enables a preliminary interaction between features in the frequency domain and the spatial domain. The formula is as follows: ; Step S405: First, and Feature channels are concatenated to obtain cross-domain concatenated features, which are then input into two sets of gated convolutions. Pooling, convolution, and a sigmoid function are used to obtain the gate weights. Finally, element-wise multiplication is used to achieve dynamic weighted fusion, resulting in preliminary fused features. The formula is: ; ; ; Step S406: ... The input MLP is subjected to nonlinear enhancement along the channel dimension, and then added to the residual of the original normalized features to obtain the final features after degenerate perceptual fusion. The formula is: .

[0013] Preferably, step S5 is as follows: Step S501: Calculate the depth features output in step S4. Upsampling is performed based on a super-resolution scaling factor of 4, using two stacked 1×1 convolutions and a pixel rearrangement (PixelShuffle). The 1×1 convolutions expand the number of channels by a factor of 4, as shown in the formula: ; ; Step S502: Convert the upsampled feature channel count into the input channel count of the MRI image using a 3×3 convolution to obtain the final output: .

[0014] Second aspect This invention provides an end-to-end blind super-resolution system for medical MRI images based on spatial-frequency parallel attention, used to execute the aforementioned end-to-end blind super-resolution method for medical MRI images. By constructing a deep learning network that integrates spatial and frequency domain features, blind super-resolution reconstruction of low-resolution medical MRI images is achieved.

[0015] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention, based on a CNN and Transformer architecture, comprises a shallow feature extraction module, a spatial-frequency parallel attention fusion module, a degradation-aware fusion module, and a high-resolution image reconstruction module. The shallow feature extraction module extracts shallow features from the input low-resolution MRI image, capturing the image's basic texture and structural information. The spatial-frequency parallel attention fusion module simultaneously assigns spatial domain attention weights and enhances frequency domain attention on the extracted features. The degradation adaptive estimation module, leveraging the global modeling capabilities of the Transformer, dynamically learns unknown degradation parameters from the fused features, eliminating the need for manually preset degradation models, and achieves adaptive fusion of dual-domain features through a gating mechanism. The high-resolution image reconstruction module combines the degradation estimation results with the fused features, using a sub-pixel convolutional layer (PixelShuffle) to enlarge the feature map size, and further corrects reconstruction errors through residual connections, outputting a high-resolution MRI image. This network can be completed within an end-to-end framework, effectively avoiding the propagation of degradation estimation errors while simultaneously preserving spatial domain details and suppressing frequency domain noise. After model training, the clinical low-resolution MRI test images to be processed can be fed into the network for super-resolution, finally outputting high-resolution MRI image results.

[0016] The present invention employs an end-to-end deep learning network architecture, which deeply combines the local feature capture advantage of CNN with the global degradation modeling capability of Transformer. It extracts local fine features of low-resolution images through CNN and models the global feature distribution through Transformer, accurately estimating unknown degradation patterns under different scanning scenarios, thus solving the shortcomings of single architectures in local detail or global degradation modeling.

[0017] This invention features a spatial-frequency parallel attention fusion module (SFPAM) that enables synergistic enhancement of spatial and frequency domain features. The spatial domain branch enhances the detail capture capability of spatial domain features through residual nested dense blocks (RRDB), focusing on lesion areas and organ edges. The frequency domain branch enhances high-frequency details in the frequency domain through operations such as Fourier transform. The combination of the two can simultaneously preserve the spatial structure information and frequency domain detail information of the image, avoiding the detail loss problem caused by single-domain processing.

[0018] This invention develops a degradation-aware fusion module, DAFM, which can learn filters to dynamically adapt to the frequency domain features of different degradations without the need for a pre-set degradation model. By utilizing learnable frequency domain filtering parameters and a gating mechanism, it can adaptively adjust the feature fusion method according to different degradation conditions, improving the model's adaptability to unknown degradation scenarios and maintaining good super-resolution performance even in complex clinical MRI image degradation scenarios.

[0019] In summary, this invention achieves deep interaction and fusion of features through multiple attention mechanisms—channel attention, spatial window attention, and frequency domain attention—closely combining global structure capture in the spatial domain with high-frequency detail enhancement in the frequency domain. This avoids the problem of losing subtle lesion textures in single spatial domain processing and solves the defect of amplifying noise in pure frequency domain operation. Furthermore, the degradation-aware fusion module DAFM is designed to dynamically adapt to different degradation modes through learnable filtering parameters, effectively avoiding the shortcoming of "degradation estimation error propagation" in traditional two-stage methods. Attached Figure Description

[0020] Figure 1 This is a schematic diagram of the method flow provided in an embodiment of the present invention; Detailed Implementation

[0021] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0022] This invention provides an end-to-end blind super-resolution method for medical MRI based on spatial-frequency parallel attention. The implementation process is as follows: a low-resolution MRI image is input into the network. First, shallow features are extracted by convolution, followed by a deep feature extraction module consisting of a spatial-frequency domain parallel attention module and a degradation adaptive fusion module. Finally, a reconstruction module is used. After training, the test MRI image to be super-resolution is fed into the network for super-resolution, and the super-resolution result is output. This invention performs blind super-resolution of medical MRI images by combining spatial and frequency domain information. The spatial-frequency domain parallel attention module learns low-frequency and high-frequency, global and local information, while the degradation adaptive fusion module dynamically learns unknown degradation parameters from the fused features and further fuses spatial and frequency domain features.

[0023] like Figure 1 As shown, this invention provides an end-to-end blind super-resolution method for medical MRI based on spatial-frequency parallel attention, comprising the following steps: Step S1: Simulate the real degradation process of clinical MRI images by using a degradation formula to generate corresponding low-resolution MRI images from high-resolution MRI images, and construct training sample pairs; Specifically, Step S101: Before inputting data into the model, the original image needs to be preprocessed and augmented. In the experiment, the image resolution was uniformly constrained to 512×512. A corresponding low-resolution image of 64×64 was generated through degradation processing. Then, data augmentation operations such as random flipping, random cropping, and random rotation were performed on the original image to avoid overfitting during network training and enhance the generalization ability of the generator. Finally, the pixel value range of the image was converted to [-1, 1].

[0024] Step S102: Perform degradation processing on the high-resolution MRI images in the dataset. Specifically, this process includes: convolving the MRI images in the dataset with a Gaussian blur kernel, then transforming them to the frequency domain using Fourier transform (FFT), performing truncation in the frequency domain, then transforming them back to the spatial domain using inverse Fourier transform, and finally adding additive white Gaussian noise. This process is represented as:

[0025] in, For high-resolution MRI images, These are low-resolution MRI images. This represents the convolution operation. Represents the Gaussian blur kernel. Representing the Fourier transform, it transforms the image from the spatial domain to the frequency domain. This represents the inverse Fourier transform, which converts data from the frequency domain to the spatial domain. This indicates a truncation operation. This represents Rice noise.

[0026] Step S103: This application conducted two sets of degradation experiments. The kernel width σ of the Gaussian kernel used was set to a range of [0.2, 4.0], and the size was set to 21 × 21. One set was set to an isotropic Gaussian kernel without noise; the other set was set to Rice noise with an anisotropic Gaussian kernel and a noise level range of [0, 30]. The anisotropic Gaussian kernel is characterized by a Gaussian probability density function N(0, ∑) (with a covariance matrix Σ with zero mean and variation). The covariance matrix Σ is determined by two random eigenvalues ​​λ1, λ2∈U(0.2, 4) and a random rotation angle θ∈U(0, π).

[0027] Step S2: Input the low-resolution MRI image into the model. First, extract the basic features such as the shallow bottom layer edge and texture of the image through a shallow extraction module composed of 3 x 3 convolutions. Step S3: Input the shallow features extracted in step S2 into the Spatial-Frequency Parallel Attention (SFPAM) module to achieve synergistic enhancement of spatial and frequency domain features; Specifically, Step S301: Assume the feature map output by the shallow feature extraction module Size is ,in For training batches, The number of feature map channels. and These represent the height and width of the feature map, respectively. Features are input into the channel attention module, where they are divided into... Two branches, one of which is fed into two parallel max-pooling and average-pooling layers to obtain a feature size of... of and The two features are then processed through a weight-sharing multilayer perceptron, added together, and finally activated by a ReLU function to obtain a feature of size . The feature weights, and then combined with Multiplying them together yields a value of 1. Channel attention optimization features This process can be represented as:

[0028]

[0029]

[0030] in, This represents a max pooling operation. This represents an average pooling operation, MLP represents a multilayer perceptron, and ReLU represents the ReLU activation function.

[0031] Step S302: Optimize the channel attention features output in step S301. First, the input is fed into the spatial domain branch, which consists of Residual Nested Dense Blocks (RRDBs). Each RRDB consists of two Residual Dense Blocks (RDBs). The final output of the spatial domain branch is then multiplied by 0.2 and mixed with the original input features. Addition achieves residual connection, and this nested structure enhances the ability to capture details of spatial domain features. The process can be represented as follows:

[0032]

[0033] Step S303: Optimize the channel attention features output in step S301. The input is then fed into the frequency domain branch, serving as the input to the frequency domain attention module (FAM). First, the features are normalized using a LayerNorm layer, and then the number of feature channels is adjusted using a 1×1 convolution to obtain the final feature count. The process is represented as follows:

[0034] in, This represents a 1×1 convolution. This indicates LayerNorm normalization.

[0035] Step S304: Then split it according to the channel dimension into , , Then, by using 3×3 depthwise separable convolution to enhance the correlation of local features, we obtain... , , The formula is as follows:

[0036]

[0037]

[0038]

[0039] in, For channel splitting operation, This represents depthwise separable convolution.

[0040] Step S305: Adopt an image segmentation strategy to divide the image into blocks. , Divide the data into multiple local feature blocks of 8×8 size, and transform the feature blocks into... Feature map , ,in Number of blocks For single-piece dimensions; Step S306: Process the features after block division , Performing a two-dimensional Fast Fourier Transform (FFT) transforms the features from the spatial domain to the frequency domain. By introducing degradation parameters deg_r and deg_i, the impact of degradation on the frequency domain is simulated through product and subtraction, adapting to unknown degradation scenarios in blind super-resolution, and calculating in the frequency domain. and The attention weights are then transformed back to the spatial domain using an inverse fast Fourier transform (IFFT), and rearranged to restore the original feature sizes, resulting in the attention weight map in the frequency domain. The process is represented as follows:

[0041]

[0042]

[0043] in, This represents a two-dimensional Fast Fourier Transform. This represents the inverse fast Fourier transform. This indicates that the inverse fast Fourier transform is in dimensions 4 and 5. This indicates the absolute value operation. This indicates the size of the rearrangement feature.

[0044] Step S307: Adjust the frequency domain attention weights and The features are multiplied element-wise, and then the number of channels is restored by a 1×1 convolution to obtain the frequency domain optimized features. The process is represented as follows:

[0045] Step S308: Analyze the spatial domain features output in step S302. and By concatenating along the channel dimension, cross-domain complementary features are obtained. The process is represented as follows:

[0046] Step S309: Cross-domain complementary features output in step S308 Perform average pooling and max pooling operations on the channel dimension respectively, and then pool the features. and The channels are concatenated along the channel dimension, the number of channels is compressed using a 3×3 convolutional layer, and a spatial attention weight map is generated using a sigmoid activation function. This process can be represented as:

[0047]

[0048]

[0049] in, This represents a max pooling operation. This represents an average pooling operation. This represents the Sigmoid activation function.

[0050] Step S310: Assign spatial attention weights to the graph. Compared with the original input features Element-wise multiplication is performed to enhance key region features, resulting in dual-domain enhanced features. :

[0051] Step S4: The features output in step S3 are fed into the degradation-aware fusion module (DAFM) to achieve deep interaction between spatial and frequency domain features. At the same time, the frequency domain features with different degradations are dynamically adapted through learnable filtering without the need to preset a degradation model. Step S401: Develop the dual-domain enhanced features output from step S3. The input degradation adaptive estimation module (DSEM) enables deep interaction between spatial and frequency domain features. First, the features are normalized using a LayerNorm layer to eliminate numerical differences across different channels and spatial locations, providing a stable feature distribution for subsequent domain-specific processing. The formula is as follows:

[0052] Step S402: Process the normalized features The window is divided into 8×8 blocks, and then the number of feature channels is adjusted by 1×1 convolution. The blocks are then divided into 3 groups according to the channel dimension to obtain features Q, K, and V. Spatial attention weights within the window are calculated using multi-head attention. Finally, the window attention features are rearranged to restore the original size. The formula is as follows:

[0053]

[0054]

[0055]

[0056] in, Indicates the rearrangement feature size. This is a channel splitting operation, where C is the number of channels. Scaling factor This represents the softamax normalization function.

[0057] Step S403: Process the normalized features Inputting a 1×1 convolution and a 3×3 depthwise separable convolution yields the enhanced result. , will feature Transform from the spatial domain to the frequency domain, and learn a learnable frequency domain weight in the frequency domain. The high-frequency characteristics of the degradation adaptation are then converted to the frequency domain. The formula is:

[0058]

[0059]

[0060] in, For learnable frequency domain filtering parameters, This represents a two-dimensional Fast Fourier Transform. This represents the inverse fast Fourier transform.

[0061] Step S404: Enhance the high-frequency features and Dot product enables a preliminary interaction between features in the frequency domain and the spatial domain. The formula is as follows:

[0062] Step S405: First, and Feature channels are concatenated to obtain cross-domain concatenated features, which are then input into two sets of gated convolutions. Pooling, convolution, and a sigmoid function are used to obtain the gate weights. Finally, element-wise multiplication is used to achieve dynamic weighted fusion, resulting in preliminary fused features. The formula is:

[0063]

[0064]

[0065] Step S406: ... The input MLP is subjected to nonlinear enhancement along the channel dimension, and then added to the residual of the original normalized features to obtain the final features after degenerate perceptual fusion. The formula is:

[0066] Step S5: The optimized features from step S4 are fed into the reconstruction module, and a high-resolution MRI image is generated through upsampling. Step S501: Calculate the depth features output in step S4. Upsampling is performed using a super-resolution scaling factor of 4, through two stacked 1×1 convolutions and PixelShuffle, where the 1×1 convolution expands the number of channels to 4 times, as shown in the formula:

[0067]

[0068] Step S502: Convert the upsampled feature channel number into the input channel number of the MRI image using a 3×3 convolution.

[0069] Step S6: Optimize network parameters using the Adam optimizer and joint loss function; All experiments used a uniform training setup: the network was trained using the Adam optimizer, and first-order gradient estimation was performed. Set to 0.9, second-order gradient estimation Set to 0.999. Their initial learning rate is set to... A stepped decay mechanism was used to adjust the learning rate. The batch size was set to 64, and the L1 loss function was used.

[0070] Step S7: Feed the test set images into the trained network for testing to obtain the complete super-resolution image.

[0071] To demonstrate the effectiveness and adaptability of this invention, experiments were conducted using the IXI dataset to verify the effectiveness of the algorithm presented in this chapter. The evaluation metrics for the test results included two items: Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity (SSIM). Table 1 compares the evaluation metrics of this invention with other blind super-resolution methods on the IXI dataset using isotropic experiments. Table 2 compares the evaluation metrics of this invention with other restoration methods on the IXI dataset using anisotropic Gaussian kernels and Ricean noise at a noise level of 5; Table 3 compares the evaluation metrics of this invention with other restoration methods on the IXI dataset using anisotropic Gaussian kernels and Ricean noise at a noise level of 10; Table 4 compares the evaluation metrics of this invention with other restoration methods on the IXI dataset using anisotropic Gaussian kernels and Ricean noise at a noise level of 20; Table 5 compares the evaluation metrics of this invention with other restoration methods on the IXI dataset using anisotropic Gaussian kernels and Ricean noise at a noise level of 30.

[0072] Table 1

[0073] Table 2

[0074] Table 3

[0075] Table 4

[0076] Table 5

[0077] Finally, it should be noted that the above embodiments are merely illustrative and explanatory of the present invention, and are not intended to limit the present invention to the scope of the described embodiments. Furthermore, those skilled in the art will understand that the present invention is not limited to the above embodiments, and many more variations and modifications can be made based on the teachings of the present invention, all of which fall within the scope of protection claimed by the present invention.

Claims

1. An end-to-end blind super-resolution method for medical MRI images based on spatial-frequency parallel attention, characterized in that, By constructing a deep learning network that integrates spatial and frequency domain features, blind super-resolution reconstruction of low-resolution medical MRI images can be achieved.

2. The end-to-end blind super-resolution method for medical MRI images based on spatial-frequency parallel attention according to claim 1, characterized in that, Includes the following steps: Step S1: Simulate the real degradation process of clinical MRI images by using a degradation formula to generate corresponding low-resolution MRI images from high-resolution MRI images, and construct training sample pairs; Step S2: Input the low-resolution MRI image into the model, and first extract the shallow features of the image through a shallow extraction module composed of 3x3 convolutions; Step S3: Input the shallow features extracted in step S2 into the Spatial-Frequency Parallel Attention (SFPAM) module to achieve synergistic enhancement of spatial and frequency domain features; Step S4: The features output in step S3 are fed into the degradation-aware fusion module DAFM to achieve deep interaction between spatial and frequency domain features. At the same time, the learnable filter dynamically adapts to the frequency domain features of different degradations without the need to preset the degradation model. Step S5: The optimized features from step S4 are fed into the reconstruction module, and a high-resolution MRI image is generated through upsampling. Step S6: Optimize network parameters using the Adam optimizer and loss function; Step S7: Input the test set images into the trained network and output the corresponding super-resolution high-resolution MRI images.

3. The end-to-end blind super-resolution method for medical MRI images based on spatial-frequency parallel attention according to claim 2, characterized in that, Step S1 is as follows: The high-resolution MRI images in the dataset undergo degradation processing, which includes: convolving the MRI images in the dataset with a Gaussian blur kernel, then transforming them to the frequency domain using Fourier transform (FFT), performing a truncation operation in the frequency domain, transforming them back to the spatial domain using inverse Fourier transform, and finally adding additive white Gaussian noise. This process is represented as follows: ； in, For high-resolution MRI images, These are low-resolution MRI images. This represents the convolution operation. Represents the Gaussian blur kernel. Represents the Fourier transform, which transforms an image from the spatial domain to the frequency domain. This represents the inverse Fourier transform, which converts data from the frequency domain to the spatial domain. This indicates a truncation operation. This represents Rice noise.

4. The end-to-end blind super-resolution method for medical MRI images based on spatial-frequency parallel attention according to claim 3, characterized in that, Step S3 is as follows: Step S301: Assume the feature map output by the shallow feature extraction module Size is ,in For training batches, The number of feature map channels. and These are the height and width of the feature map, respectively; Features are input into the channel attention module, where they are divided into... Two branches, one of which is fed into two parallel max-pooling and average-pooling layers to obtain a feature size of... of and The two features are then processed through a weight-sharing multilayer perceptron, added together, and finally activated by a ReLU function to obtain a feature of size . The feature weights, and then combined with Multiplying them together yields a value of 1. Channel attention optimization features This process can be represented as: ； ； ； in, This represents a max pooling operation. This represents an average pooling operation; MLP represents a multilayer perceptron; and ReLU represents the ReLU activation function. Step S302: Optimize the channel attention features output in step S301. First, the input is fed into the spatial domain branch, which consists of nested residual dense blocks (RRDBs). Each RRDB block consists of two residual dense blocks (RDBs). The final output of the spatial domain branch is then compared with the original input features using a weight multiplier of 0.

2. Addition achieves residual connection, and this nested structure enhances the ability to capture details of spatial domain features. The process can be represented as follows: ； ； Step S303: Optimize the channel attention features output in step S301. The input is then fed into the frequency domain branch, serving as the input to the frequency domain attention module (FAM). First, the features are normalized using a LayerNorm layer, and then the number of feature channels is adjusted using a 1×1 convolution to obtain the feature values. The process is represented as follows: ； in, This represents a 1×1 convolution. This indicates LayerNorm normalization; Step S304: Then split it according to the channel dimension into , , Then, by using 3×3 depthwise separable convolution to enhance the correlation of local features, we obtain... , , The formula is as follows: ； ； ； ； in, For channel splitting operation, This represents depthwise separable convolution; Step S305: Adopt an image segmentation strategy to divide the image into blocks. , Divide the data into multiple local feature blocks of 8×8 size, and transform the feature blocks into... Feature map , ,in Number of blocks For single-piece dimensions; Step S306: Process the features after block division , Performing a two-dimensional Fast Fourier Transform (FFT) transforms the features from the spatial domain to the frequency domain. By introducing degradation parameters deg_r and deg_i, the impact of degradation on the frequency domain is simulated through product and subtraction, adapting to unknown degradation scenarios in blind super-resolution, and calculating in the frequency domain. and The attention weights are then transformed back to the spatial domain using an inverse fast Fourier transform (IFFT), and rearranged to restore the original feature sizes, resulting in the attention weight map in the frequency domain. The process is represented as follows: ； ； ； in, This represents a two-dimensional Fast Fourier Transform. This represents the inverse fast Fourier transform. This indicates that the inverse fast Fourier transform is in dimensions 4 and 5. This indicates the absolute value operation. Indicates the rearrangement feature size; Step S307: Adjust the frequency domain attention weights and The features are multiplied element-wise, and then the number of channels is restored by a 1×1 convolution to obtain the frequency domain optimized features. The process is represented as follows: ； Step S308: Analyze the spatial domain features output in step S302. and By concatenating along the channel dimension, cross-domain complementary features are obtained. The process is represented as follows: ； Step S309: Cross-domain complementary features output in step S308 Perform average pooling and max pooling operations on the channel dimension respectively, and then pool the features. and The channels are concatenated along the channel dimension, the number of channels is compressed using a 3×3 convolutional layer, and a spatial attention weight map is generated using a sigmoid activation function. This process can be represented as: ； ； ； in, This represents a max pooling operation. This represents an average pooling operation. This represents the Sigmoid activation function; Step S310: Assign spatial attention weights to the graph. Compared with the original input features Element-wise multiplication is performed to enhance key region features, resulting in dual-domain enhanced features. : 。 5. The end-to-end blind super-resolution method for medical MRI images based on spatial-frequency parallel attention according to claim 4, characterized in that, Step S4 is as follows: Step S401: Develop the dual-domain enhanced features output from step S3. The input degradation adaptive estimation module DSEM enables deep interaction between spatial and frequency domain features. First, the features are normalized using a LayerNorm layer to eliminate numerical differences across different channels and spatial locations, providing a stable feature distribution for subsequent domain-specific processing. The formula is as follows: ； Step S402: Process the normalized features The window is divided into 8×8 blocks, and then the number of feature channels is adjusted by 1×1 convolution. The blocks are then divided into 3 groups on average according to the channel dimension to obtain features Q, K, and V. Spatial attention weights within the window are calculated using multi-head attention. Finally, the window attention features are rearranged to restore the original size. The formula is as follows: ； ； ； ； in, Indicates the rearrangement feature size. This is a channel splitting operation, where C is the number of channels. Scaling factor This represents the softamax normalization function; Step S403: Process the normalized features Inputting a 1×1 convolution and a 3×3 depthwise separable convolution yields the enhanced result. , will feature Transform from the spatial domain to the frequency domain, and learn a learnable frequency domain weight in the frequency domain. The high-frequency characteristics of the degradation adaptation are then converted to the frequency domain. The formula is: ； ； ； in, For learnable frequency domain filtering parameters, This represents a two-dimensional Fast Fourier Transform. This represents the inverse fast Fourier transform; Step S404: Enhance the high-frequency features and Dot product enables a preliminary interaction between frequency domain and spatial domain features; the formula is: ； Step S405: First, and Feature channels are concatenated to obtain cross-domain concatenated features, which are then input into two sets of gated convolutions. Pooling, convolution, and a sigmoid function are used to obtain the gate weights. Finally, element-wise multiplication is used to achieve dynamic weighted fusion, resulting in preliminary fused features. The formula is: ； ； ； Step S406: ... The input MLP is subjected to nonlinear enhancement along the channel dimension, and then added to the residual of the original normalized features to obtain the final features after degenerate perceptual fusion. The formula is: 。 6. The end-to-end blind super-resolution method for medical MRI images based on spatial-frequency parallel attention according to claim 5, characterized in that, Step S5 is as follows: Step S501: Calculate the depth features output in step S4. Upsampling is performed based on a super-resolution scaling factor of 4, using two stacked 1×1 convolutions and a pixel rearrangement (PixelShuffle). The 1×1 convolutions expand the number of channels by a factor of 4, as shown in the formula: ； ； Step S502: Convert the upsampled feature channel count into the input channel count of the MRI image using a 3×3 convolution to obtain the final output: 。 7. An end-to-end blind super-resolution system for medical MRI images based on spatial-frequency parallel attention, used to perform the end-to-end blind super-resolution method for medical MRI images based on spatial-frequency parallel attention as described in any one of claims 1-6, characterized in that, By constructing a deep learning network that integrates spatial and frequency domain features, blind super-resolution reconstruction of low-resolution medical MRI images can be achieved.

Images