A Diffusion-Based Dual-Domain Enhancement Method for Low-Field MRI Images

Abstract

This invention discloses a dual-domain enhancement method for low-field MRI images based on a diffusion model, aiming to improve the quality and detail of low-field MRI image enhancement by combining deep learning with a diffusion model. The method first decomposes the MRI image in the frequency domain, dividing it into real and imaginary parts. Then, a frequency-domain binary transformation module (FDBT) extracts feature maps of the real and imaginary parts respectively and performs binary transformation. Next, frequency-domain feature fusion combines the features of these two parts, and inverse Fourier transform (IFFT) is used to extract the frequency domain information of the image. After joint optimization in the frequency and image domains, the enhanced image features are finally output. This method performs excellently in denoising and detail enhancement of low-field MRI images, effectively reducing noise and improving image quality. This invention has better image restoration capabilities and stronger generalization ability, maintaining high performance even in unseen real-world data, and possesses high potential for practical applications.

Description

Technical Field

[0001] This invention relates to a dual-domain enhancement method for low-field MRI images based on a diffusion model, belonging to the field of computer image processing. Background Technology

[0002] Low-field magnetic resonance imaging (MRI) equipment is highly practical in resource-scarce environments due to its advantages such as low cost, high mobility, and high safety, and can provide rich tissue and functional information. However, such as Figure 1 As shown, low-field MRI faces challenges such as low image signal-to-noise ratio and high blur, which greatly affects image quality and is detrimental to subsequent diagnosis and treatment.

[0003] In low-field magnetic field environments, the spatial resolution of MRI imaging is significantly reduced, and noise interference is significantly increased. These two factors work together to cause a substantial decrease in image sharpness and contrast, resulting in image quality degradation. This degradation differs from that of ordinary low-quality images; it is a coupling of multiple deteriorating factors, including noise and artifacts caused by low magnetic fields and undersampling. However, typical diffusion models involve adding and removing Gaussian noise, which conflicts with each other. Therefore, typical diffusion models cannot accurately remove artifacts and noise in low-field MRI. Furthermore, while generative model-based network structures improve generation performance, most models lose some original image details and have a certain probability of generating false details, which is unacceptable in clinical applications. Therefore, this invention proposes a dual-domain enhancement method for low-field MRI images based on a diffusion model. Based on sufficient training data pairs, a deep learning-based low-field image enhancement method—Frequency-Informed Diffusion Network—is proposed. Summary of the Invention

[0004] This invention addresses the conflict between Gaussian noise in traditional diffusion models and coupling noise in actual low-field images. It proposes a dual-domain enhancement method for low-field MRI images based on a diffusion model. This invention improves noise removal and detail enhancement in low-field MRI images through joint optimization in the frequency and image domains. The method first decomposes the MRI image in the frequency domain into real and imaginary parts. Then, a Frequency-Domain Binary Transformation Module (FDBT) extracts feature maps of the real and imaginary parts respectively and performs binary processing. Next, frequency-domain feature fusion combines the features of these two parts, and inverse Fourier transform (IFFT) is used to extract the frequency domain information of the image. After joint optimization in the frequency and image domains, the enhanced image features are finally output. In the image domain, this invention simulates the degradation process of low-field MRI images using a degradation simulation method and combines a Residual Denoising Diffusion Model (RDDM) for forward and backward diffusion processes to effectively remove noise from the MRI image. This network employs a U-Net architecture for deep learning, designing forward diffusion and inverse reconstruction processes to achieve efficient denoising and detail enhancement during image restoration. Furthermore, this invention incorporates an Adaptive Fourier Transform (AFF) module, which improves the accuracy and efficiency of image reconstruction by adaptively filtering frequency domain data. This module combines Fourier transform and inverse transform, further enhancing denoising performance and image detail preservation by optimizing frequency domain information. Experimental results demonstrate that this method performs excellently in denoising and detail enhancement of low-field MRI images, effectively reducing noise and improving image quality. Compared to traditional methods, this invention exhibits better image restoration capabilities and stronger generalization ability, maintaining high performance even in unseen real-world data, and possesses significant practical application potential. The entire process of this method is as follows: Figure 2 As shown, a low-field image simulation algorithm is used to generate matching data pairs for network training; an FID Net network is designed to take low-quality MRI images as input and output high-quality MRI images, thus preserving image details while removing noise.

[0005] To achieve the above objectives, the technical solution of the present invention is as follows: a dual-domain enhancement method for low-field MRI images based on a diffusion model, the method comprising the following steps:

[0006] Step 1. Simulation Degradation of Low-Field MRI Image Dataset. To better simulate the noise distribution of low-field MR data and improve image restoration performance, a low-field image degradation simulation method is used to degrade high-field images, mimicking the quality of 0.1T MR images. Specifically, the degradation process begins with undersampling in k-space to simulate the inherent loss of detail in low-field images. Next, random Gaussian noise is added to the frequency domain image to introduce a composite noise of Gaussian and Ricean noise to simulate the low signal-to-noise ratio of low-field MR scans. Finally, the images are spatially normalized. Furthermore, since there are significant differences in noise levels between various MR sequences and imaging sites, we randomly changed the standard deviation parameter of Gaussian noise added to the real and virtual images within a certain range. This method enhances the realism and universality of the simulation data, as shown in Equation (1).

[0007] Im = FFT(RN1(Ix) + RN2(Iy)) (1)

[0008] Step 2. Design the Frequency-Informed Diffusion Network (FID-Net) for low-field MRI enhancement. Simulated low-field MRI images are used as training data, and high-field MRI data are used as labels input into the FID-Net enhancement network. The FID-Net network is a two-branch structure based on the Diffusion network. FID-Net is built on the diffusion model, leveraging its powerful pixel-level learning capability to ensure global denoising while improving detail preservation. To better adapt to the diffusion model for MRI enhancement, we combine frequency domain information to guide the process in the image domain. We obtain fused feature maps from the real and imaginary parts of the frequency domain data using binary encoders and frequency domain encoders respectively, and then use a concat operation to integrate them into the middle layer of the UNET to guide the denoising process of the diffusion model.

[0009] Step 3: Perform binary conversion and encoder processing on the original frequency domain image, as shown in equations (2) and (3), to extract the real and imaginary parts of the frequency domain, and perform deep feature map extraction to ensure effective alignment between the frequency domain features and the image domain content.

[0010] Design for Binary Transformation of Millions of Data Points in the Frequency Domain: Binary numbers have only two states, 0 and 1, allowing for direct computation and processing with minimal space requirements. This ensures the accuracy and stability of the data during training. Therefore, we propose using the binary method BT (Binary Transformation) to address the challenge of processing millions of data points in k-space. The specific approach is as follows: The total number of bits after binary transformation is set to 64, with 1 bit for the sign bit. The remaining bits are for the integer and fractional parts, with 32 bits for the integer part and 31 bits for the fractional part. The sign bit is used to handle positive and negative numbers independently and directly added to the final result. Performing binary transformations for both integers and fractions simultaneously on all numbers, instead of processing them one by one in a loop, reduces the algorithm's complexity. Furthermore, we use the `cat` operation to concatenate the sign bit, integer part, and fractional part together, avoiding redundant operations.

[0011] Encoder Design: For FIDNet, extracting and utilizing deep feature maps from the frequency domain is crucial. Therefore, FIDNet employs frequency domain coding. Deep features are extracted from the frequency domain, enabling better capture of low-frequency and high-frequency information, as well as local and global patterns. The system consists of five layers. In the first four layers, each layer contains a 3×3 convolutional operation, a SiLU activation function, and a linear attention module to progressively extract deeper features from the input. To reduce spatial resolution, a 2×2 downsampling operation is applied after each of the first four layers. Initially, the feature map has dimensions H×W×C, where H and W represent height and width, and C is the number of channels. After the downsampling step, the spatial resolution is halved. At the end of the fourth layer, the feature map is reduced to H / 8×W / 8×C', where C' is the number of channels in the fourth layer. The final layer consists of 1×1 convolutions, which refine the feature representation and reduce channel redundancy. The encoder effectively compresses binary features while extracting deep features from the frequency domain.

[0012]

[0013]

[0014] Step 4: To preserve image details, an Adaptive Fourier Transform (AFF) module is proposed. This module adaptively filters the frequency domain data, improving the accuracy and efficiency of image reconstruction. Specifically, given encoder features, we first apply a two-dimensional Fourier transform (2D FFT) along the spatial dimension. Next, the frequency domain features are decomposed into their real parts F... kx and the imaginary part F ky To better focus on detailed information and learn adaptive filters, we provide component F kxand F ky Learnable weight graphs W1 and W2 were constructed, and W1 was compared with F... kx Multiply by W2 and F ky Multiplication is used to apply it. The spectral filter aids training by globally adjusting specific frequencies, while the learned weights adapt to different frequency distributions of the target data. The frequency domain features processed in step 3 are input into the Adaptive Fourier Module (AFF) for fusion. By fusing information from the real and imaginary components, a frequency domain feature map is obtained. The AFF in step 4 is placed in the middle layer of the UNET in the diffusion model. The features of the middle layer are then fused with the adaptive filter and the depth features obtained in step 3 using a concat operation through Fourier transform, thereby further enhancing image details.

[0015] Step 5: Convert the frequency domain feature map into image domain information through inverse Fourier transform (IFFT) to obtain the image domain feature map, and combine it with the original image to perform image domain enhancement and improve image quality.

[0016] Step 6: Add the image domain feature map obtained in Step 5 to the middle layer of the diffusion model U-NET for feature fusion, gradually removing noise in the image and restoring image details. The concat operation is used in this step to fuse the deep features obtained in Step 5 with the middle layer of UNET.

[0017] Step 7: The frequency domain data of the degraded image and the real image obtained in Step 1 are converted into image domain data through inverse Fourier transform and input into the dual-domain diffusion model (FIDNet) network. The image is denoised through forward diffusion and inverse denoising. The degraded image and real image mentioned in Step 7 are simulated data pairs obtained using the simulation degradation algorithm for training. Here, the original data is assumed to be frequency domain data, which is then converted into image domain data through Fourier transform. The frequency domain and image domain data are input into the FID-Net network for dual-domain training. The FID-Net network training process proposed in Step 7 uses the simulated MRI paired data generated in Step 1 as training data and the high-field MRI image as the label. The data is input into the network in the format of simultaneous input of the image domain and frequency domain. The network is built on the diffusion model and utilizes its powerful pixel-level learning capability to ensure global denoising effect while improving detail preservation effect. The loss function is obtained by calculating the MSE loss between the low-field enhanced image and the high-field image. The gradient is calculated through the loss value, and the network parameters are updated using the stochastic gradient descent (AdaW) optimization method.

[0018] Step 8: Use the trained model to process real low-field MRI images to obtain high-quality enhanced images, significantly improving image resolution and detail fidelity.

[0019] Experimental results demonstrate that the FID-Net network performs well on both simulated low-field MRI data and real low-field MRI data, effectively enhancing low-field image quality while removing noise, thus proving its potential in practical clinical problems.

[0020] An electronic device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the program, implements the aforementioned method for dual-domain enhancement of low-field MRI images based on a diffusion model.

[0021] A computer-readable storage medium having computer instructions stored thereon, which, when executed by a processor, implement the aforementioned diffusion-based low-field MRI image dual-domain enhancement method.

[0022] Compared with existing technologies, the advantages of this invention are as follows: (1) It proposes a dual-domain enhancement method for low-field MRI images based on a diffusion model, which combines dual optimization of the frequency domain and the image domain. This method combines frequency domain feature fusion with image domain denoising, making full use of the frequency domain information and image domain information of the image, effectively improving the quality of low-field MRI images, especially in terms of noise removal and detail restoration. (2) The enhancement effect is significant and has high application potential. This invention focuses more on and utilizes the frequency domain information in MRI images through a dual-domain feature fusion module and an imaginary Fourier filter. It uses a simulation degradation method that is closer to the real acquired image and designs a dual-domain enhancement network (FIDNet) to optimize the denoising and enhancement effect of low-field MRI images. Experimental results show that this method performs well on both simulated and real data. (3) It restores image details and ensures that the image does not produce distortion or information loss after denoising. The FIDNet network designed by this method can not only effectively remove noise, but also restore the details in the image, ensuring high image quality after denoising, meeting the high standards of medical image processing, and satisfying the actual requirements of doctors for image accuracy. (4) Excellent versatility, requiring no additional parameter adjustments. This method requires no user adjustment of any parameters and can adapt to data generated by different MRI devices, exhibiting strong versatility. Excellent enhancement effects can be achieved regardless of whether the images are acquired by different devices or different types of low-field MRI images. (6) Automated, end-to-end processing method, facilitating integration and application. The method of this invention has highly automated characteristics, can be integrated with the image processing system of an MRI scanner, and can automatically complete image enhancement and denoising processing, greatly improving clinical work efficiency. Attached Figure Description

[0023] Figure 1 This is a comparison image of the imaging quality between low-field and high-field MRI.

[0024] Figure 2 This is a flowchart of the present invention.

[0025] Figure 3 This is a diagram of the FIDNet network structure.

[0026] Figure 4 Here is a flowchart of the FIDNet training process.

[0027] Figure 5 The image shows the algorithm test results. Detailed Implementation

[0028] The present invention will be further illustrated below with reference to specific embodiments. It should be understood that these embodiments are for illustrative purposes only and are not intended to limit the scope of the invention. After reading the present invention, any modifications of the present invention in various equivalent forms by those skilled in the art will fall within the scope defined by the appended claims.

[0029] Example 1: See Figures 1-5 A dual-domain enhancement method for low-field MRI images based on a diffusion model, the method comprising the following steps:

[0030] Step 1. Simulation Degradation of Low-Field MRI Image Dataset. To better simulate the noise distribution of low-field MR data and improve image restoration performance, a low-field image degradation simulation method is used to degrade high-field images, mimicking the quality of 0.1T MR images. Specifically, the degradation process begins with undersampling in k-space to simulate the inherent loss of detail in low-field images. Next, random Gaussian noise is added to the frequency domain image to introduce a composite noise of Gaussian and Ricean noise to simulate the low signal-to-noise ratio of low-field MR scans. Finally, the images are spatially normalized. Furthermore, since there are significant differences in noise levels between various MR sequences and imaging sites, we randomly changed the standard deviation parameter of Gaussian noise added to the real and virtual images within a certain range. This method enhances the realism and universality of the simulation data, as shown in Equation (1).

[0031] Im = FFT(RN1(Ix) + RN2(Iy)) (1)

[0032] Step 2. Design the Frequency-Informed Diffusion Network (FID-Net) for low-field MRI enhancement. Simulated low-field MRI images are used as training data, and high-field MRI data are used as labels input into the FID-Net enhancement network. The FID-Net network is a two-branch structure based on the Diffusion network. FID-Net is built on the diffusion model, leveraging its powerful pixel-level learning capability to ensure global denoising while improving detail preservation. To better adapt to the diffusion model for MRI enhancement, we combine frequency domain information to guide the process in the image domain. We obtain fused feature maps from the real and imaginary parts of the frequency domain data using binary encoders and frequency domain encoders respectively, and then use a concat operation to integrate them into the middle layer of the UNET to guide the denoising process of the diffusion model.

[0033] Step 3: Perform binary conversion and encoder processing on the original frequency domain image, as shown in equations (2) and (3), to extract the real and imaginary parts of the frequency domain, and perform deep feature map extraction to ensure effective alignment between the frequency domain features and the image domain content.

[0034] Design for Binary Transformation of Millions of Data Points in the Frequency Domain: Binary numbers have only two states, 0 and 1, allowing for direct computation and processing with minimal space requirements. This ensures the accuracy and stability of the data during training. Therefore, we propose using the binary method BT (Binary Transformation) to address the challenge of processing millions of data points in k-space. The specific approach is as follows: The total number of bits after binary transformation is set to 64, with 1 bit for the sign bit. The remaining bits are for the integer and fractional parts, with 32 bits for the integer part and 31 bits for the fractional part. The sign bit is used to handle positive and negative numbers independently and directly added to the final result. Performing binary transformations for both integers and fractions simultaneously on all numbers, instead of processing them one by one in a loop, reduces the algorithm's complexity. Furthermore, we use the `cat` operation to concatenate the sign bit, integer part, and fractional part together, avoiding redundant operations.

[0035] Encoder Design: For FIDNet, extracting and utilizing deep feature maps from the frequency domain is crucial. Therefore, FIDNet employs frequency domain coding. Deep features are extracted from the frequency domain, enabling better capture of low-frequency and high-frequency information, as well as local and global patterns. The system consists of five layers. In the first four layers, each layer contains a 3×3 convolutional operation, a SiLU activation function, and a linear attention module to progressively extract deeper features from the input. To reduce spatial resolution, a 2×2 downsampling operation is applied after each of the first four layers. Initially, the feature map has dimensions H×W×C, where H and W represent height and width, and C is the number of channels. After the downsampling step, the spatial resolution is halved. At the end of the fourth layer, the feature map is reduced to H / 8×W / 8×C', where C' is the number of channels in the fourth layer. The final layer consists of 1×1 convolutions, which refine the feature representation and reduce channel redundancy. The encoder effectively compresses binary features while extracting deep features from the frequency domain.

[0036]

[0037]

[0038] Step 4: To preserve image details, an Adaptive Fourier Transform (AFF) module is proposed. This module adaptively filters the frequency domain data, improving the accuracy and efficiency of image reconstruction. Specifically, given encoder features, we first apply a two-dimensional Fourier transform (2D FFT) along the spatial dimension. Next, the frequency domain features are decomposed into their real parts F... kx and the imaginary part F ky To better focus on detailed information and learn adaptive filters, we provide component F kx and F ky Learnable weight graphs W1 and W2 were constructed, and W1 was compared with F... kx Multiply by W2 and F ky Multiplication is used to apply it. The spectral filter aids training by globally adjusting specific frequencies, while the learned weights adapt to different frequency distributions of the target data. The frequency domain features processed in step 3 are input into the Adaptive Fourier Module (AFF) for fusion. By fusing information from the real and imaginary components, a frequency domain feature map is obtained. The AFF in step 4 is placed in the middle layer of the UNET in the diffusion model. The features of the middle layer are then fused with the adaptive filter and the depth features obtained in step 3 using a concat operation through Fourier transform, thereby further enhancing image details.

[0039] Step 5: Convert the frequency domain feature map into image domain information through inverse Fourier transform (IFFT) to obtain the image domain feature map, and combine it with the original image to perform image domain enhancement and improve image quality.

[0040] Step 6: Add the image domain feature map obtained in Step 5 to the middle layer of the diffusion model U-NET for feature fusion, gradually removing noise in the image and restoring image details. The concat operation is used in this step to fuse the deep features obtained in Step 5 with the middle layer of UNET.

[0041] Step 7: The frequency domain data of the degraded image and the real image obtained in Step 1 are converted into image domain data through inverse Fourier transform and input into the dual-domain diffusion model (FIDNet) network. The image is denoised through forward diffusion and inverse denoising. The degraded image and real image mentioned in Step 7 are simulated data pairs obtained using the simulation degradation algorithm for training. Here, the original data is assumed to be frequency domain data, which is then converted into image domain data through Fourier transform. The frequency domain and image domain data are input into the FID-Net network for dual-domain training. The FID-Net network training process proposed in Step 7 uses the simulated MRI paired data generated in Step 1 as training data and the high-field MRI image as the label. The data is input into the network in the format of simultaneous input of the image domain and frequency domain. The network is built on the diffusion model and utilizes its powerful pixel-level learning capability to ensure global denoising effect while improving detail preservation effect. The loss function is obtained by calculating the MSE loss between the low-field enhanced image and the high-field image. The gradient is calculated through the loss value, and the network parameters are updated using the stochastic gradient descent (AdaW) optimization method.

[0042] Step 8: Use the trained model to process real low-field MRI images to obtain high-quality enhanced images, significantly improving image resolution and detail fidelity.

[0043] Experimental results demonstrate that the FID-Net network performs well on both simulated low-field MRI data and real low-field MRI data, effectively enhancing low-field image quality while removing noise, thus proving its potential in practical clinical problems.

[0044] Example 2: Figure 2 As shown, this invention provides a dual-domain enhancement method for low-field MRI images based on a diffusion model. The enhancement test results are as follows. Figure 5 As shown, the specific steps are as follows:

[0045] Step 1: Align the physical dimensions of the numerical phantom according to the characteristics of the low-field MRI image.

[0046] Step 2: Add artifacts and noise to the frequency domain data to more realistically simulate actual scanning conditions.

[0047] Step 3: Inverse Fourier transform the frequency domain data into image domain data.

[0048] Step 4: Use simulated low-field and high-field dual-domain data as training data to train the FIDNet network.

[0049] Step 5: Input real low-field MRI data and output high-quality MRI data.

[0050] Effectiveness evaluation:

[0051] This invention discloses a dual-domain enhancement method for low-field MRI images based on a diffusion model. The test results of this method are as follows: Figure 5 As shown in the figure, the image enhancement effect is illustrated in two sets of test images, including simulated degraded data and real low-field MRI data. The test procedure is as follows: low-field MRI image data is input into the model, and the deep learning model is activated to perform dual-domain image enhancement processing. After processing, the system automatically outputs a high-quality denoised and enhanced MRI image. By comparing the image with the original image, the significant improvement in denoising effect and detail restoration is verified. The algorithm runs fast, processing images in a short time, removing noise and preserving image details, greatly enhancing image quality and facilitating diagnosis and treatment decisions by doctors.

[0052] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any transformations or substitutions that can be conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A dual-domain enhancement method for low-field MRI images based on a diffusion model, characterized in that, The following are the specific steps for enhancing low-field MRI images using dual-domain information, namely image domain and frequency domain information: Step 1: Degradation simulation is performed on the frequency domain images of high-field MRI to generate degraded low-field MRI images, which are then used as training data to form paired data pairs. Step 2: Decompose the MRI frequency domain image into real and imaginary parts to obtain the original representation of the frequency domain image. Step 3: Perform binary and encoder processing on the real and imaginary part images processed in Step 2 to extract the real and imaginary part information in the frequency domain, and perform deep feature map extraction to ensure effective alignment between the frequency domain features and the image content. Step 4: Input the frequency domain features processed in Step 3 into the Adaptive Fourier Transform (AFF) module for fusion. By fusing information from the real and imaginary components, a frequency domain feature map is obtained. Step 5: Convert the frequency domain feature map into image domain information using Inverse Fourier Transform (IFFT) to obtain the image domain feature map. Combine this with the original image to perform image domain enhancement and improve image quality. Step 6: Add the image domain feature map obtained in Step 5 to the middle layer of the diffusion model U-NET for feature fusion, gradually removing noise from the image and restoring image details. Step 7: Perform inverse Fourier transform on the frequency domain data of the degraded image and the real image obtained in Step 1 to convert them into image domain data. Input this data into the dual-domain diffusion model FIDNet network, and perform denoising processing on the image through forward diffusion and inverse denoising processes. Step 8: Use the trained model to process real low-field MRI images to obtain high-quality enhanced images, significantly improving image resolution and detail fidelity. In step 7, the degraded image and the real image are trained using simulated data pairs obtained from the simulation degradation algorithm. Here, the original data is assumed to be frequency domain data, which is then Fourier transformed to obtain image domain data. Both the frequency domain and image domain data are input into the FID-Net network for dual-domain training. During the FID-Net network training process, the simulated MRI paired data generated in step 1 is used as training data, and the high-field MRI image is used as the label. The data is input into the network in a format where both the image domain and frequency domain are input simultaneously. The network is built on the diffusion model, which utilizes its powerful pixel-level learning capability to ensure global denoising while improving detail preservation. The loss function is obtained by calculating the MSE loss between the low-field enhanced image and the high-field image. The gradient is calculated using the loss value, and the network parameters are updated using the stochastic gradient descent optimization method AdaW.

2. The method for dual-domain enhancement of low-field MRI images based on a diffusion model according to claim 1, characterized in that, The degradation simulation method mentioned in step 1 first undersamples the high-field image in the K-space to simulate the loss of detail in the low-field image. Then, Ricean noise is added to the image domain to simulate the low signal-to-noise ratio of the low-field image. Finally, the image is spatially normalized. in, and These represent the real and imaginary parts of the original K-space image, respectively. and These represent undersampling the real and imaginary parts of the image and adding Rice noise with different standard deviations, respectively. The Fourier transform represents the conversion of K-space data into image domain data. This represents the image after simulation degradation.

3. The method for dual-domain enhancement of low-field MRI images based on a diffusion model according to claim 1, characterized in that, Step 3 introduces binary and encoder processing. Binary processing means converting millions of real and imaginary part data in the frequency domain into binary data for easier manipulation. The encoder aims to further extract deep feature maps from the obtained binary data through multi-layer convolution, attention mechanisms, and downsampling operations. in, and These represent the real and imaginary parts of the original K-space image, respectively. Represents binary processing. Represents encoder, and These are deep feature maps representing the real and imaginary parts, respectively.

4. The method for dual-domain enhancement of low-field MRI images based on a diffusion model according to claim 1, characterized in that, The adaptive Fourier module (AFF) proposed in step 4 is placed in the middle layer of the UNET in the diffusion model. The features of the middle layer are transformed by Fourier transform and the feature maps of the real and imaginary parts obtained by the adaptive filter are fused with the real and imaginary depth features obtained in step 3 through the concat operation to further enhance the image details.

5. The method for dual-domain enhancement of low-field MRI images based on a diffusion model according to claim 1, characterized in that, Step 6 proposes to fuse the deep features obtained in step 5 with the features of the middle layer of UNET using the concat operation.

6. 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 program, it implements a low-field MRI image dual-domain enhancement method based on a diffusion model as described in any one of claims 1 to 5.

7. A computer-readable storage medium storing computer instructions thereon, characterized in that, When executed by a processor, the computer instructions implement a low-field MRI image dual-domain enhancement method based on a diffusion model as described in any one of claims 1-5.

Images