A Deep Learning-Based Image Denoising Method and System for Ultra-Low Field Magnetic Resonance Imaging

By using a deep learning denoising network that employs collaborative modeling across the image, frequency, and spatial domains, the problem of noise interference in ultra-low field magnetic resonance imaging systems is solved. This achieves efficient noise suppression and tissue structure restoration, improving image quality and equipment applicability.

CN123118401APending Publication Date: 2026-11-17XI AN JIAOTONG UNIV
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611484706.9
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-09-22
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

The image quality of ultra-low field magnetic resonance imaging systems is severely affected by noise. Existing technologies are difficult to effectively denoise and affect image observation and tissue structure analysis. Traditional methods are prone to edge weakening and over-smoothing under complex noise conditions. Deep learning methods have high computational cost and do not fully utilize the natural connection between magnetic resonance images and the frequency domain.

Method used

A deep learning denoising network based on image domain-frequency domain-spatial domain collaborative modeling is adopted. Through a multi-scale encoder-decoder structure and IKS Block feature processing unit, it combines local features in the image domain, K-space frequency features and spatial Transformer features to perform collaborative modeling and residual reconstruction.

Benefits of technology

While controlling the number of parameters and computational complexity, it improves noise suppression, tissue structure restoration and edge preservation capabilities, reduces the cost of acquiring high-quality paired data, and enhances the processing accuracy and equipment applicability of ultra-low field magnetic resonance images.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123118401A_ABST
    Figure CN123118401A_ABST
Patent Text Reader

Abstract

This invention discloses a deep learning-based image denoising method and system for ultra-low field magnetic resonance imaging (MRI). The method involves acquiring the original ultra-low field MRI image to be denoised and preprocessing it to obtain a standardized MRI image. The standardized MRI image is then input into a pre-trained deep learning denoising network based on image-frequency-spatial domain collaborative modeling to obtain a denoised ultra-low field MRI image. This invention can improve noise suppression, tissue structure restoration, and edge preservation capabilities while controlling the number of parameters and computational complexity, reducing the cost of acquiring high-quality paired data, and enhancing the accuracy, stability, and engineering application capabilities of ultra-low field MRI image processing.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the fields of magnetic resonance imaging (MRI), medical image processing, computer vision, and deep learning, and specifically to an ultra-low-field (ULF) magnetic resonance deep learning image denoising method and system based on image domain-frequency domain-spatial domain collaborative modeling. Background Technology

[0002] Ultra-low field magnetic resonance imaging (MRI) systems offer advantages such as simple magnet structure, good openness, low deployment cost, and suitability for bedside and portable imaging. However, due to factors such as main magnetic field strength, receiver coil sensitivity, acquisition time, environmental electromagnetic interference, and system noise, images obtained from ultra-low field MRI systems typically suffer from low signal-to-noise ratio, significant background random noise, prominent local granular fluctuations, blurred tissue edges, and easy loss of fine structures. These issues, in turn, affect image observation, region of interest measurement, and tissue structure analysis.

[0003] Traditional methods such as mean filtering, median filtering, Gaussian filtering, nonlocal means, and wavelet thresholding mainly suppress noise based on local statistical regularities or preset thresholds. However, they are prone to edge weakening and over-smoothing under complex noise or low contrast conditions. Convolutional neural networks can learn local texture features, but their ability to model long-distance structural relationships and overall frequency distribution is limited. Although Transformers can establish spatial dependencies through self-attention, they consume a lot of computation and memory when directly processing high-resolution images such as 512×512, and simply relying on spatial attention does not fully utilize the natural connection between magnetic resonance images and the frequency domain.

[0004] The original sampling of magnetic resonance signals is located in k-space. Convolution processing in the image domain alone is insufficient to account for both global frequency perturbations and local tissue texture. Adjustment in the frequency domain alone may ignore spatial location and structural relationships. Therefore, it is necessary to collaboratively model local information in the image domain, complex spectral information in the frequency domain, and spatial structural information.

[0005] Furthermore, supervised deep learning denoising relies on a sufficient quantity of spatially strictly corresponding degraded image-reference image pairing data, but repeated acquisition, strict registration, and noise-free reference acquisition of real ultra-low field and high-quality images are quite challenging. Simply adding a single Gaussian noise to a high-quality image is insufficient to simultaneously characterize degradation factors such as intensity variations, reception inhomogeneity, k-space high-frequency loss, resolution degradation, blurring, and Rician noise. Therefore, a lightweight image-frequency-spatial domain collaborative denoising method is urgently needed, along with a repeatable and traceable composite degraded pairing data construction process to improve noise suppression, structure preservation, and applicability to real equipment for ultra-low field magnetic resonance images. Summary of the Invention

[0006] The purpose of this invention is to provide a method and system for denoising ultra-low field magnetic resonance deep learning images, so as to overcome the defects of the existing technology. This invention can improve noise suppression, tissue structure restoration and edge preservation capabilities while controlling the number of parameters and computational complexity, reduce the cost of acquiring high-quality paired data, and improve the accuracy, stability and engineering application capabilities of ultra-low field magnetic resonance image processing.

[0007] To achieve the above objectives, the present invention adopts the following technical solution: A deep learning-based image denoising method for ultra-low field magnetic resonance imaging includes the following steps: Step 1: Acquire the original ultra-low field magnetic resonance image to be denoised and perform preprocessing to obtain a standardized magnetic resonance image; Step 2: Input the standardized magnetic resonance image into a pre-trained deep learning denoising network based on image domain-frequency domain-spatial domain co-modeling to obtain a denoised ultra-low field magnetic resonance image, specifically including: Step 2.1: Input the standardized magnetic resonance image into the convolutional layer for shallow feature extraction to obtain initial spatial features; Step 2.2: Input the initial spatial features into the multi-scale coding part, and perform step-by-step feature extraction through multiple coding stages of different scales to obtain deep coding features; Step 2.3: Input the deep coding features into the bottleneck feature extraction part to obtain the bottleneck features; Step 2.4: Obtain enhanced decoding features by decoding the bottleneck features at multiple scales; Step 2.5: Input the enhanced decoding features into the convolutional layer to obtain the network prediction residual, and use the residual reconstruction method to obtain the denoised ultra-low field magnetic resonance image.

[0008] Furthermore, step 1 specifically includes: The acquired raw magnetic resonance images are subjected to amplitude conversion to convert them into single-channel magnetic resonance grayscale images:

[0009] in, This indicates an image grayscale conversion operation. This represents the converted single-channel magnetic resonance grayscale image; The size of the single-channel magnetic resonance grayscale image is adjusted:

[0010] in, This indicates an image resizing operation. The image after resizing; The resized image is then normalized to limit the pixel values ​​of the input image to the range of 0 to 1.

[0011] in, and These represent the minimum and maximum grayscale values ​​of the image, respectively. Represents a pixel range constraint function; After the above processing, a standardized magnetic resonance image is obtained. ,in, and These represent the height and width of the input image, respectively.

[0012] Further, step 2.1 specifically involves: standardizing the magnetic resonance image... Input a 3×3 convolutional layer for shallow feature extraction to obtain initial spatial features:

[0013] in, This represents a 3×3 convolution operation. This represents the initial spatial features obtained by applying shallow convolution to the input image; The multi-scale coding part in step 2.2 includes multiple sequentially connected coding stages. Each coding stage consists of an IKS Block feature processing unit and a scale transformation operation, specifically including: In the first-level encoding stage, the basic spatial features are input into the first-level IKS Block feature processing unit:

[0014] in, This represents the output features of the first-level encoding stage. This indicates that the IKS Block feature processing process is performed twice consecutively in the first-level encoding stage; Subsequently, the feature space resolution is reduced through a first 2×2 downsampling operation to obtain scale-compressed features:

[0015] in, This indicates the first scaling operation. These are features after scale compression; The scale-compressed features are then input into the second-level encoding stage:

[0016] in, This represents the output features of the second-level encoding stage. This indicates that the IKS Block feature processing process is performed twice consecutively in the second-level encoding stage. The feature space resolution is reduced by a second 2×2 downsampling operation to obtain deep encoded features:

[0017] in, This indicates the second scaling operation. These are the deep coding features after the second scaling transformation.

[0018] Further, step 2.3 specifically involves: inputting the deep coding features after two scale transformations into the bottleneck feature extraction part to obtain the bottleneck features:

[0019] in, This represents a bottleneck feature processing process consisting of four consecutive IKS blocks. Represents deep semantic features of the network; Step 2.4 specifically includes: First, the bottleneck features are upsampled:

[0020] in, This indicates the first upsampling operation. These are the decoding features after the first upsampling; Subsequently, the decoded features after the first upsampling... Scale features corresponding to the encoding stage Fusion via skip connections:

[0021] in, This indicates a feature concatenation operation. This represents the first-stage decoded features after fusion via skip connections; The first-stage decoded features after fusion are processed by IKS Blocks in two consecutive decoding stages:

[0022] in, This indicates that the IKS Block feature processing is performed twice consecutively during the first-level decoding stage. This is an enhanced decoding feature resulting from two consecutive IKS Block feature processing steps in the first-level decoding stage; Enhanced decoding features Perform a second upsampling:

[0023] in, This indicates the first upsampling operation. These are the decoding features after the second upsampling; The decoded features after the second upsampling Scale features corresponding to the encoding stage Fusion via skip connections:

[0024] in, This represents the second-stage decoded features after fusion via skip connections; The fused second-stage decoded features are then processed by IKS Blocks in two consecutive decoding stages:

[0025] in, This indicates that the IKS Block feature processing process is performed twice consecutively in the second-level decoding stage. This is an enhanced decoding feature resulting from two consecutive IKS Block feature processing steps in the second-level decoding stage; Step 2.5 specifically includes: The enhanced decoding features are input into a 3×3 convolutional layer to obtain the network prediction residuals:

[0026] in, This represents the noise correction residual information obtained from network learning; Subsequently, the final network output is obtained using residual reconstruction:

[0027] in, This represents the input standardized magnetic resonance image. This represents a denoised ultra-low field magnetic resonance image.

[0028] Furthermore, the IKS Block feature processing unit includes an image domain local feature processing unit, a K-space frequency feature processing unit, and a spatial Transformer feature processing unit; The basic spatial features are input into the local feature processing unit of the image domain, and the image domain enhancement features are extracted through convolution operation. The processing flow of the local feature processing unit of the image domain includes layer normalization, pointwise convolution, depthwise separable convolution and nonlinear activation function. The image domain enhancement features are input into the K-space frequency feature processing unit to obtain the frequency-enhanced features. The frequency-enhanced features are input into the space Transformer feature processing unit to obtain the IKS Block output features.

[0029] Furthermore, the step of inputting the basic spatial features into the image domain local feature processing unit and extracting image domain enhancement features through convolution operations specifically includes: First, the basic spatial features are normalized layer by layer:

[0030] in, Presentation layer normalization operation, These are the features after layer normalization; The feature channels are adjusted using 1×1 convolution:

[0031] in, This represents a 1×1 convolution operation, used to implement feature dimension mapping. This indicates the intermediate features after channel adjustment; Local spatial structure information is extracted using 3×3 depthwise convolution.

[0032] in, This indicates a 3×3 depth separable convolution operation. Indicates local spatial structure enhancement features; Enhancing the nonlinear expressive power of the network through the GELU activation function:

[0033] And the feature dimension is restored by 1×1 convolution:

[0034] in, Indicates nonlinear enhancement features, Represents the output of local features in the image domain. This indicates an output mapping convolution operation; Image domain enhancement features are obtained using residual connectivity:

[0035] in, This represents the learnable residual weight parameters. This represents the output features after image domain feature enhancement.

[0036] Further, the step of inputting the image domain enhanced features into the K-space frequency feature processing unit to obtain the frequency-enhanced features specifically includes: Output features after image domain feature enhancement Perform a two-dimensional fast Fourier transform:

[0037] in, This represents a two-dimensional Fast Fourier Transform. This represents the frequency domain characteristics after conversion; Since the Fourier transform result is in complex form, it can be further expressed as:

[0038] in, This represents the real part of the spectrum. Indicates the imaginary part of the spectrum; Concatenate the real and imaginary parts of the spectrum along the channel dimension:

[0039] in, This represents the splicing feature of the real and imaginary parts in the frequency domain; Frequency feature mapping is achieved using 1×1 convolution and the GELU activation function.

[0040] in, This indicates the first channel mapping layer. This indicates the second channel mapping layer. This represents the spectral characteristics after frequency domain feature transformation; The frequency domain information is converted back to the spatial domain using the inverse Fourier transform:

[0041] in, This represents the spatial characteristics after frequency domain enhancement. This represents the inverse fast Fourier transform; Residual fusion is performed between the frequency-enhanced features and the output features after image domain feature enhancement:

[0042] in, Represents the frequency domain fusion weights. This represents the features after fusing K-space frequency information.

[0043] Furthermore, the step of inputting the frequency-enhanced features into the space Transformer feature processing unit to obtain the IKS Block output features specifically includes: The features after fusing K spatial frequency information are then subjected to layer normalization:

[0044] in, This represents the frequency enhancement features after layer normalization. Calculate spatial relationships using a multi-head self-attention mechanism within a window:

[0045] in, This indicates a multi-head self-attention operation in the window; Indicates the attention residual weight; This represents the features after spatial attention enhancement; The spatially attention-enhanced features are input into a gated deep convolutional feedforward network, where convolution operations enhance the local spatial representation capabilities.

[0046] in, , Indicates the gating branch feature; This represents element-wise multiplication; Indicates output mapping convolution; Indicates the feedforward network fusion weights; This represents the features enhanced through spatial modeling. Finally, the IKS Block output features are obtained:

[0047] in, This represents the final feature representation after co-enhancement in the image domain, K-space frequency domain, and spatial domain.

[0048] Furthermore, the training method for the deep learning denoising network based on image domain-frequency domain-spatial domain co-modeling includes the following steps: High-field magnetic resonance images are acquired as training target reference images and preprocessed. Composite degradation processing is performed on the preprocessed training target reference image to obtain a composite degradation image without random noise; Noise simulation is performed based on the aforementioned composite degradation image without random noise to obtain an ultra-low field magnetic resonance simulation degradation image; A paired training dataset is constructed based on the ultra-low field magnetic resonance simulated degradation images and the corresponding training target reference images; Based on the paired training dataset, a deep learning denoising network based on image domain-frequency domain-spatial domain co-modeling is forward propagated to obtain a predicted denoised image. The joint loss is calculated based on the predicted denoised image and the training target reference image, and the network parameters are updated accordingly. The optimal network parameters are determined based on the iterative training results, resulting in a trained deep learning denoising network based on collaborative modeling of the image domain, frequency domain, and spatial domain.

[0049] A deep learning-based image denoising system for ultra-low field magnetic resonance imaging includes: Data acquisition and preprocessing module: used to acquire the original ultra-low field magnetic resonance image to be denoised and preprocess it to obtain a standardized magnetic resonance image; Denoising module: Used to input standardized magnetic resonance images into a pre-trained deep learning denoising network based on image domain-frequency domain-spatial domain co-modeling to obtain denoised ultra-low field magnetic resonance images; The noise reduction module specifically includes: Shallow feature extraction module: used to input standardized magnetic resonance images into convolutional layers for shallow feature extraction to obtain initial spatial features; Deep extraction module: used to input the initial spatial features into the multi-scale coding part, and perform step-by-step feature extraction through multiple coding stages of different scales to obtain deep coded features; Bottleneck Feature Extraction Module: This module takes deep-encoded features as input to the bottleneck feature extraction section to obtain the bottleneck features. Decoding module: Used to decode bottleneck features through multi-scale decoding to obtain enhanced decoded features; Processing module: Used to input enhanced decoding features into convolutional layers, obtain network prediction residuals, and use residual reconstruction to obtain denoised ultra-low field magnetic resonance images.

[0050] Compared with the prior art, the present invention has the following beneficial technical effects: This invention constructs ultra-low field approximate degradation-reference paired data with traceable degradation parameters by applying intensity variations, low-frequency bias fields, k-space high-frequency loss, resolution degradation, spatial blurring, and complex domain Rician noise to high-quality magnetic resonance reference images. This simulates the complex degradation in ultra-low field magnetic resonance imaging, reducing the difficulty of acquiring and registering real paired data. Compared with methods that only process in a single feature domain, this invention uses local feature processing in the image domain, complex spectrum modulation in the frequency domain, and window space feature processing to collaboratively model tissue edges, spectral distribution, and spatial structural relationships. By combining multi-scale encoder-decoder, skip connections, and global residual learning, it can improve noise suppression, edge preservation, and detail recovery capabilities while controlling the number of model parameters and computational complexity, reducing over-smoothing and loss of structural information. This improves the image processing accuracy, model applicability, and engineering application capabilities of bedside, open, and portable ultra-low field magnetic resonance systems. Attached Figure Description

[0051] The accompanying drawings are provided to further understand the invention and constitute a part of this invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0052] Figure 1 This is a schematic diagram of the IKSFormer-light network structure and the core module IKS Block in the ultra-low field magnetic resonance deep learning image denoising method of the present invention; wherein, (a) is a schematic diagram of the overall network structure of IKSFormer-light, and (b) is a schematic diagram of the internal structure of IKS Block.

[0053] Figure 2 The above is a comparison of the visual denoising results of different ablation models on the paired test set in the embodiments of the present invention. (a), (b) and (c) represent three different paired test samples, respectively. The upper row of each group is the complete image, and the lower row is the magnified image of the corresponding local area. The red, green and orange boxes and their vertical downward arrows represent the magnified local areas in the three test samples, respectively.

[0054] Figure 3 This is a comparison chart of quantitative indicators of different ablation models on a paired test set in this embodiment of the invention. Among them: (a) is the comparison result of peak signal-to-noise ratio (PSNR); (b) is the comparison result of structural similarity index (SSIM); (c) is the comparison result of root mean square error (RMSE); and (d) is the comparison result of high frequency error norm (HFEN).

[0055] Figure 4 The image above shows the denoising results and residual distribution of a 64 mT real device magnetic resonance image in this embodiment of the invention.

[0056] Figure 5 This is an analysis diagram of the signal-to-noise ratio (SNR) and background noise distribution of real device magnetic resonance images based on the region of interest (ROI) in an embodiment of the present invention. In the diagrams (a) to (d), the analysis results of four groups of real magnetic resonance test samples are shown. In each group, the top two images show a visual comparison between the original magnetic resonance image (Raw) and the image after deep learning denoising (Denoised), along with the corresponding SNR and noise level evaluation results. The bottom left image is a schematic diagram of the background noise region ROI (B1-B4) before and after denoising. The bottom right image is a histogram of the noise intensity distribution in the corresponding background region, where gray represents the noise distribution of the original image and blue represents the noise distribution of the image after denoising.

[0057] Figure 6The following are statistical analysis charts of the signal-to-noise ratio (SNR) and background noise of real device magnetic resonance images before and after denoising in embodiments of the present invention; wherein, (a) is a comparison chart of SNR before and after denoising calculated based on the background region, with the horizontal axis representing the Raw image and the Denoised image respectively, and the vertical axis representing the SNR estimated based on the background region, in dB; data points of the same image before and after denoising are connected by gray lines. (b) is a comparison chart of background noise levels before and after denoising, with the horizontal axis representing the Raw image and the Denoised image respectively, and the vertical axis representing the background noise level; (c) is a distribution chart of the SNR improvement of each real device image, where the SNR improvement is calculated by subtracting the Raw image SNR from the Denoised image SNR.

[0058] Figure 7 This is a schematic diagram of the ultra-low field magnetic resonance deep learning image denoising method of the present invention. Detailed Implementation

[0059] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. 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 skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0060] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0061] Example 1 A deep learning-based image denoising method for ultra-low field magnetic resonance imaging includes the following steps: Step 1: Acquire the original ultra-low field magnetic resonance image to be denoised and perform preprocessing to obtain a standardized magnetic resonance image; Specifically, the magnetic resonance image of the object to be processed is acquired using an ultra-low field magnetic resonance imaging (MRI) device to obtain the original ultra-low field magnetic resonance image. Because ultra-low field magnetic resonance systems operate in a low magnetic field environment, their acquired signal strength is low, resulting in raw magnetic resonance images that often suffer from significant noise interference, blurred tissue edges, and loss of detail. Therefore, it is necessary to preprocess the raw ultra-low field magnetic resonance images to meet the input requirements of subsequent deep learning networks.

[0062] First, the acquired raw magnetic resonance images are subjected to amplitude conversion to convert them into single-channel magnetic resonance grayscale images:

[0063] in, This indicates an image grayscale conversion operation. This represents the converted single-channel magnetic resonance grayscale image.

[0064] Furthermore, to ensure consistent input image size, the single-channel magnetic resonance grayscale image is resized:

[0065] in, This indicates an image resizing operation. This is the resized image.

[0066] Furthermore, the resized image is normalized to limit the pixel values ​​of the input image to the range of 0 to 1:

[0067] in, and These represent the minimum and maximum grayscale values ​​of the image, respectively. This represents a function that constrains the pixel range.

[0068] After the above processing, a standardized magnetic resonance image is obtained. ,in, and These represent the height and width of the input image, respectively.

[0069] (2) Step 2: Use the trained deep learning denoising network based on image domain-frequency domain-spatial domain co-modeling to achieve denoising of ultra-low field magnetic resonance images; The standardized magnetic resonance image obtained in step 1 The input is fed into a trained deep learning denoising network based on image domain-frequency domain-spatial domain co-modeling, and the denoising result is obtained through the network's forward propagation process:

[0070] in, This indicates a trained deep learning denoising network based on collaborative modeling of the image domain, frequency domain, and spatial domain; here, it is simply referred to as the IKSFormer-light network. This represents the optimal parameters obtained during network training. This represents the denoised image output by the network.

[0071] The IKSFormer-light network adopts an encoder-decoder structure, including shallow feature extraction, multi-scale encoding, bottleneck feature extraction, multi-scale decoding, and residual reconstruction, such as... Figure 1 As shown in (a).

[0072] Specifically, firstly, standardized magnetic resonance images Input a 3×3 convolutional layer for shallow feature extraction to obtain initial spatial features:

[0073] in, This represents a 3×3 convolution operation. This represents the initial spatial features obtained by applying shallow convolution to the input image.

[0074] Furthermore, the initial spatial features are input into the multi-scale coding section, and feature extraction is performed step by step through multiple coding stages of different scales.

[0075] The multi-scale coding section comprises multiple sequentially connected coding stages. Each coding stage consists of an IKS Block feature processing unit and a scale transformation operation, used to progressively expand the network's receptive field and extract image features at different scales. The IKS Block feature processing unit is a multi-domain feature processing unit that fuses local image domain information, K-space frequency information, and spatial context information; its specific structure will be described in a subsequent supplementary explanation.

[0076] First, in the first-level encoding stage, the basic spatial features are input into the first-level IKS Block feature processing unit:

[0077] in, This represents the output features of the first-level encoding stage. This indicates that the IKS Block feature processing process is performed twice consecutively in the first-level encoding stage.

[0078] Subsequently, the feature space resolution is reduced through a first 2×2 downsampling operation:

[0079] in, This indicates the first scaling operation. These are the features after scale compression.

[0080] Furthermore, the scale-compressed features are input into the second-level encoding stage:

[0081] This represents the output features of the second-level encoding stage. This indicates that the IKS Block feature processing is performed twice consecutively in the second-level encoding stage. Then, a second 2×2 downsampling operation is used to reduce the feature space resolution.

[0082] in, This indicates the second scaling operation. The deep coding features after the second scaling transformation Through the multi-scale encoding process described above, the network can gradually expand its receptive field and extract deep structural information from ultra-low field magnetic resonance images.

[0083] Furthermore, the deep coding features, after two scaling transformations, are input into the bottleneck feature extraction part:

[0084] in, This represents a bottleneck feature processing process consisting of four consecutive IKS blocks. It represents the deep semantic features of the network.

[0085] Furthermore, to restore the image spatial resolution, the network employs a symmetric decoding structure. First, bottleneck features are upsampled:

[0086] in, This indicates the first upsampling operation. These are the decoding features after the first upsampling.

[0087] Subsequently, the decoded features after the first upsampling... Scale features corresponding to the encoding stage Fusion via skip connections:

[0088] in, This indicates a feature concatenation operation. This represents the first-stage decoded features after fusion via skip connections.

[0089] The first-stage decoded features after fusion are processed by IKS Blocks in two consecutive decoding stages:

[0090] This indicates that the IKS Block feature processing process is performed twice consecutively in the first-level decoding stage. This is an enhanced decoding feature after performing IKS Block feature processing twice in the first-level decoding stage.

[0091] Furthermore, for enhanced decoding features Perform a second upsampling:

[0092] in, This indicates the first upsampling operation. These are the decoding features after the second upsampling.

[0093] Subsequently, the decoded features after the second upsampling Scale features corresponding to the encoding stage Fusion via skip connections:

[0094] This represents the second-stage decoded features after fusion via skip connections.

[0095] The fused second-stage decoded features are then processed by IKS Blocks in two consecutive decoding stages:

[0096] This indicates that the IKS Block feature processing process is performed twice consecutively in the second-level decoding stage. This is an enhanced decoding feature after performing IKS Block feature processing twice in the second-level decoding stage.

[0097] Through the above decoding process, the network gradually restores the spatial resolution of the image and utilizes the shallow structural information saved in the encoding stage to improve the ability to restore tissue edges and details.

[0098] Furthermore, the enhanced decoding features are input into a 3×3 convolutional layer to obtain the network prediction residuals:

[0099] in, This represents the noise correction residual information obtained from network learning.

[0100] Subsequently, the final network output is obtained using residual reconstruction:

[0101] in, This represents the input standardized magnetic resonance image. This represents the noise correction residual information obtained from network learning. This represents the denoised image output by the network.

[0102] The above steps achieve the conversion from the original ultra-low field magnetic resonance acquired image to a high-quality denoised image.

[0103] Furthermore, the training methods for the IKS Block feature processing unit and the IKSFormer-light deep learning model are explained in detail: (1) Introduction to the structure of IKS Block Furthermore, in order to achieve the effective extraction of different dimensional feature information of ultra-low field magnetic resonance images by the IKSFormer-light network in step 2, this invention introduces IKSBlock as a basic feature processing unit in the network's encoding stage, bottleneck feature extraction stage, and decoding stage.

[0104] The IKS Block is used to jointly model the basic spatial features using multi-domain information. By simultaneously learning local structural information in the image domain, frequency distribution information in the K-space, and long-distance spatial dependencies, it improves the network's ability to express noise features and tissue structure features in ultra-low field magnetic resonance images.

[0105] like Figure 1 As shown in (b), the IKS Block includes an Image Local Unit, a K-space Frequency Unit, and a Spatial Transformer Unit.

[0106] Specifically, the basic spatial features of the input IKS Block are represented as follows:

[0107] in, , These represent the height and width of the basic spatial feature map, respectively. Indicates the number of feature channels.

[0108] The basic spatial features first enter the image domain local feature processing unit, where image domain enhancement features are extracted through convolution operations. The processing procedure is as follows:

[0109] in, This represents the process of extracting local features in the image domain. This represents the output feature after image domain feature enhancement, i.e., the image domain enhanced feature.

[0110] The image domain local feature processing process will be described in detail next. The image domain local feature processing unit adopts a combination of layer normalization, point-wise convolution, depth-wise convolution and non-linear activation function to enhance the local feature extraction capability while reducing the number of network parameters.

[0111] Specifically, the basic spatial features are first normalized layer by layer:

[0112] in, The representation layer normalization operation is used to reduce the distribution differences between different feature channels and improve the stability of network training. These are the features after layer normalization.

[0113] Subsequently, the feature channels are adjusted using 1×1 convolution:

[0114] in, This represents a 1×1 convolution operation, used to implement feature dimension mapping. This indicates the intermediate features after channel adjustment.

[0115] Furthermore, local spatial structure information is extracted through 3×3 depthwise convolution:

[0116] in, This indicates a 3×3 depth separable convolution operation. This represents enhanced local spatial structure features. Compared to traditional convolution methods, this operation can effectively extract edge, texture, and local tissue structure information from ultra-low field magnetic resonance images while reducing the number of parameters and computational cost.

[0117] Subsequently, the nonlinear expressive power of the network is enhanced by using the GELU activation function:

[0118] And the feature dimension is restored by 1×1 convolution:

[0119] in, Indicates nonlinear enhancement features, Represents the output of local features in the image domain. This indicates the output mapping convolution operation.

[0120] Furthermore, to avoid feature information loss during deep network training, residual connections are used to obtain image domain enhancement features:

[0121] in, This represents the learnable residual weight parameters. This represents the output features after image domain feature enhancement.

[0122] Through the aforementioned image domain local feature processing, the IKS Block can effectively enhance the edge structure and local texture information in ultra-low field magnetic resonance images, providing a foundation for subsequent frequency domain feature modeling.

[0123] Furthermore, the image domain enhancement features are input into the K-space frequency feature processing unit. Since magnetic resonance imaging data is essentially derived from K-space sampled signals, and noise and tissue details in the image typically have different frequency distribution characteristics, this invention introduces a frequency domain feature modeling method, enabling the network to simultaneously learn the differences between effective signals and noise in the frequency space.

[0124] Specifically, the output features after image domain feature enhancement Perform a two-dimensional fast Fourier transform:

[0125] in, This represents a two-dimensional Fast Fourier Transform. This represents the frequency domain characteristics after conversion.

[0126] Since the Fourier transform result is in complex form, it can be expressed as:

[0127] in, This represents the real part of the spectrum. This represents the imaginary part of the spectrum.

[0128] Furthermore, the real and imaginary parts of the spectrum are concatenated along the channel dimension:

[0129] in, This represents the splicing feature of the real and imaginary parts in the frequency domain.

[0130] Subsequently, frequency feature mapping is achieved using 1×1 convolution and the GELU activation function:

[0131] in, This indicates the first channel mapping layer. This indicates the second channel mapping layer. This represents the spectral characteristics after frequency domain feature transformation.

[0132] Subsequently, the frequency domain information is converted back to the spatial domain using the inverse Fourier transform:

[0133] in, This represents the spatial characteristics after frequency domain enhancement. This represents the inverse fast Fourier transform.

[0134] Furthermore, residual fusion is performed between the frequency-enhanced features and the output features after image domain feature enhancement:

[0135] in, Represents the frequency domain fusion weights. This represents the features after fusing K-space frequency information.

[0136] Through the above K-space frequency feature processing, the network can effectively enhance the ability to express high-frequency details of images by utilizing the frequency distribution characteristics of magnetic resonance signals, while improving the ability to identify noise frequency features.

[0137] Furthermore, the frequency-enhanced features are input into the spatial Transformer feature processing unit to establish long-distance dependencies between different spatial regions.

[0138] Since convolution operations mainly focus on local neighborhood information, and the tissue structures in ultra-low field magnetic resonance images have strong spatial correlation, this invention uses a window multi-head self-attention mechanism to model spatial context information.

[0139] Specifically, the input features, i.e. the features after fusing K-space frequency information, are first subjected to layer normalization:

[0140] in, This represents the frequency enhancement feature after layer normalization.

[0141] Subsequently, spatial associations are calculated using a multi-head self-attention mechanism within a window:

[0142] in, This indicates a multi-head self-attention operation in the window; Indicates the attention residual weight; This represents the features enhanced by spatial attention.

[0143] Furthermore, the features enhanced with spatial attention are input into a gated deep convolutional feedforward network to enhance local spatial representation capabilities through convolution operations:

[0144] in, , Indicates the gating branch feature; This represents element-wise multiplication; Indicates output mapping convolution; Indicates the feedforward network fusion weights; This represents the features enhanced through spatial modeling.

[0145] Finally, the IKS Block output features are obtained:

[0146] in, This represents the final feature representation after co-enhancement in the image domain, K-space frequency domain, and spatial domain.

[0147] Through the aforementioned IKS Block structure, this invention achieves complementary fusion of information between different feature domains.

[0148] (2) Training method of IKSFormer-light deep learning model To obtain the trained IKSFormer-light deep learning denoising network described in step 2, this invention constructs an ultra-low field magnetic resonance image degradation-restoration paired dataset and uses supervised learning to train and optimize the network.

[0149] Specifically, high-field magnetic resonance images are first acquired as training target reference images. The reference image can be derived from high signal-to-noise ratio, high-field magnetic resonance imaging data, and preprocessed to obtain data samples of uniform size and grayscale range. Subsequently, by simulating potential degradation factors during ultra-low field magnetic resonance imaging, the degradation parameters can be adjusted according to the acquisition conditions of the target ultra-low field magnetic resonance system for each reference image. By sequentially applying intensity changes, low-frequency multiplicative bias fields, k-space high-frequency truncation, downsampling-upsampling, and Gaussian blur, a composite degradation process without random noise is constructed:

[0150] in, This is the intensity scaling factor. This is the intensity offset. Indicates reference image The first intermediate degraded image obtained after intensity scaling and intensity offset. For a low-frequency random field that varies slowly in space, the values ​​of are normalized to the range of -1 to 1. The bias field intensity coefficient, This indicates pixel-by-pixel multiplication; Indicates the first intermediate degraded image The second intermediate degraded image is obtained by applying a low-frequency multiplicative bias field on the basis of the image. This indicates that the center frequency of the k-space is preserved by a mask. This indicates the proportion of the center frequency retained in the k-space. and Re( ) represents the two-dimensional Fourier transform and the two-dimensional inverse Fourier transform, respectively. () indicates the extraction of the real part of the inverse transform result; Indicates the second intermediate degraded image The third intermediate degraded image obtained after preserving the center frequency in k-space and truncating the high-frequency components. The proportion is downsampling operation, This indicates an upsampling operation that restores the downsampling result to the original spatial dimensions; Indicates the degradation of the third intermediate image The fourth intermediate degraded image obtained after downsampling-upsampling processing. The parameter representing the radius or standard deviation is: Gaussian blur kernel, symbol " " indicates convolution; The image is a simulated low signal-to-noise ratio ultra-low field magnetic resonance image generated by the composite degradation model.

[0151] Subsequently, independent real-part Gaussian noise and imaginary-part Gaussian noise are sampled in the complex domain, and amplitude reconstruction is performed to obtain an ultra-low field approximate degraded image with Rician distribution characteristics; at the same time, small-amplitude Gaussian perturbations can be superimposed as needed:

[0152] in, and These represent the independent real Gaussian noise and imaginary Gaussian noise, respectively. This indicates an optional small Gaussian perturbation.

[0153] Furthermore, after completing the above process, a paired training dataset consisting of the degraded image and its corresponding reference image is obtained:

[0154] in, Indicates the first One reference image, This represents the corresponding low signal-to-noise ratio ultra-low field magnetic resonance simulation image generated by the composite degradation model. This indicates the number of training samples.

[0155] Specifically, during the training process, As input to the IKSFormer-light network, the corresponding As a network supervision target, by establishing a mapping relationship between the input image and the target image, the network learns the noise distribution characteristics and effective tissue structure restoration rules in the degradation process of ultra-low field magnetic resonance images.

[0156] The forward propagation process of the IKSFormer-light network is represented as follows:

[0157] in, This represents the IKSFormer-light deep learning denoising network to be trained; Indicates the network parameters to be optimized; This represents the denoised image obtained from the network prediction.

[0158] During network training, the denoised images predicted by the network are compared. Compared with reference image The differences between the images are analyzed to construct a loss function, and the backpropagation algorithm is used to continuously optimize the network parameters so that the network output gradually approaches the target image.

[0159] Since the denoising process of ultra-low field magnetic resonance images not only needs to reduce random noise, but also needs to preserve tissue structure information and restore image detail features, a single loss function is difficult to simultaneously satisfy the requirements of noise suppression and structure restoration. This invention constructs a joint loss function composed of pixel reconstruction constraints, structure preservation constraints, and high-frequency detail restoration constraints to perform multi-objective optimization on the IKSFormer-light network.

[0160] The joint loss function is expressed as:

[0161] in, This represents the pixel reconstruction loss, used to constrain the overall grayscale consistency between the network output image and the reference image; This represents structural preservation loss, used to preserve tissue structure information in magnetic resonance images; This indicates high-frequency detail recovery loss, used to enhance the ability to recover image edge and texture details; , , These represent the weighting coefficients for different loss terms.

[0162] The following is a detailed introduction. First, in order to ensure the consistency of the overall grayscale distribution between the network output image and the reference image, this invention first uses pixel reconstruction loss to constrain the network output results.

[0163] The pixel reconstruction loss is expressed as:

[0164] in, Indicates the number of training samples. This represents the L1 norm.

[0165] Furthermore, since tissue edges and internal structural information in ultra-low field magnetic resonance images are easily affected by noise, structural blurring may occur during the denoising process. Therefore, this invention introduces structural similarity loss to constrain the structural consistency between the network output image and the reference image.

[0166] The structural retention loss is expressed as:

[0167] in, This represents the structural similarity evaluation function.

[0168] Next, since edges, textures and fine tissue structures in ultra-low field magnetic resonance images are mainly contained in high-frequency information, and traditional denoising methods are prone to causing high-frequency information loss, this invention further introduces high-frequency detail recovery loss.

[0169] First, high-frequency information from the predicted and reference images is obtained through a high-frequency feature extraction operator, and a high-frequency constraint loss is constructed:

[0170] in, This represents the high-frequency information extraction operator.

[0171] Furthermore, the gradients of the network parameters are calculated based on the aforementioned joint loss function. And update the network parameters using the backpropagation algorithm:

[0172] in, Indicates the first Network parameters at the next iteration This represents the network learning rate.

[0173] After multiple rounds of training iterations, the optimal network parameters are selected based on the loss function value:

[0174] The optimal parameters obtained during training Load it into the IKSFormer-light network to obtain the IKSFormer-light deep learning denoising network trained in step 2.

[0175] Example 2 IKSFormer-light Network Multi-Domain Cooperative Structure Validation To verify the effectiveness of the IKSFormer-light deep learning image denoising method based on image domain-frequency domain-spatial domain co-modeling proposed in this invention, this embodiment employs ablation experiments to analyze the impact of different feature processing modules on network performance. While maintaining consistent training data, training strategies, and network parameter settings, E1, E2, E3, and the complete IKSFormer-light model were constructed for comparison. Specifically, the E1 model includes only an image domain local feature extraction module, the E2 model includes both image and frequency domain feature processing modules, the E3 model includes both image and spatial context feature processing modules, and the complete IKSFormer-light model includes image domain, K-space frequency domain, and spatial context feature processing modules.

[0176] like Figure 2 As shown, from left to right in the figure, the denoising results are for the degraded image (Degraded), the reference image (Clean), the E1 model, the E2 model, the E3 model, and the complete IKSFormer-light model (simplified as IKSF-L in the figure). (a), (b), and (c) represent three different test samples. The top of each result group shows the complete image, and the bottom shows a magnified view of the corresponding local region. Red, green, and orange boxes and corresponding arrows indicate the local observation area. Figure 2 It is evident that different models can reduce noise in the input image, but they differ in terms of tissue edge preservation, detail recovery, and noise residue. Compared to the E1, E2, and E3 models, the output of the complete IKSFormer-light model is closer to the reference image Clean, with more continuous local tissue boundaries and better preservation of detailed structures. This indicates that the image domain-frequency domain-spatial domain collaborative modeling strategy proposed in this invention can effectively improve the denoising performance of ultra-low field magnetic resonance images.

[0177] Furthermore, to quantitatively evaluate the impact of different feature domain combinations on denoising performance, this embodiment uses four metrics—PSNR, SSIM, HFEN, and EPI—to evaluate different models. The results are as follows: Figure 3 As shown in the figure. Among them, the larger the values ​​of PSNR, SSIM, and EPI, the better the denoising and structural restoration performance, and the smaller the HFEN value, the lower the high-frequency structural error.

[0178] Depend on Figure 3 It can be seen that the complete IKSFormer-light model achieves the best results in all evaluation metrics. The PSNR, SSIM, HFEN, and EPI of the E1 model were 28.8608±2.6786 dB, 0.7747±0.0934, 0.6064±0.1333, and 0.8166±0.0619, respectively; those of the E2 model were 29.5656±2.6464 dB, 0.8308±0.0902, 0.5814±0.1321, and 0.8503±0.0594, respectively; those of the E3 model were 29.3623±2.6401 dB, 0.8297±0.0920, 0.5688±0.1330, and 0.8544±0.0604, respectively; and the complete IKS Former-light model achieved 30.1536±2.6617. dB, 0.8869±0.0764, 0.5338±0.1326 and 0.8776±0.0584.

[0179] The above results demonstrate that the IKSFormer-light model proposed in this invention can effectively improve the noise suppression capability of ultra-low field magnetic resonance images by fusing local structural information in the image domain, frequency information in the K space, and spatial context information, while also enhancing the preservation of tissue structure and the recovery of high-frequency details.

[0180] Example 3 Verification of denoising effect on 64 mT real device magnetic resonance image To verify the practical application effect of the IKSFormer-light deep learning denoising model proposed in this invention on images acquired by a real ultra-low field magnetic resonance imaging (MRI) device, this embodiment selects 64 mT MRI images acquired by a real device for denoising processing, and analyzes the denoising results. The results are as follows: Figure 4 As shown. Figure 4Each group, from left to right, displays the original image (Raw) acquired by a real device, a magnified local result of the original image, a denoised image (Denoised) output by the IKSFormer-light model, a magnified local result of the corresponding region in the denoised image, and a residual map. All images are displayed in the same spatial location and with the same display range to allow for comparison of background noise levels, local granular fluctuations, tissue edge changes, and the recovery of fine structures before and after denoising.

[0181] Example 4 Quantitative performance verification of real-device magnetic resonance imaging To quantitatively evaluate the noise suppression capability of the IKSFormer-light model proposed in this invention in real device magnetic resonance images, this embodiment employs a signal-to-noise ratio analysis method based on the region of interest (ROI) to perform statistical analysis on the real device MRI images before and after denoising. The results are as follows: Figure 5 and Figure 6 As shown.

[0182] Figure 5 This is a graph showing the signal-to-noise ratio (SNR) and background noise distribution analysis of a real device magnetic resonance image based on its region of interest (ROI) in an embodiment of the present invention. The graph sets one ROI for the signal and four background ROIs (B1, B2, B3, and B4) at the same spatial locations before and after denoising. The background noise level is estimated based on the grayscale standard deviation of these background ROIs, and the SNR of the images before and after denoising is then calculated. Simultaneously, the graph uses background pixel distribution curves to illustrate the background grayscale fluctuation range of the Raw and Denoised images.

[0183] In the first real device image sample, the signal-to-noise ratio (SNR) of the raw image before denoising was 33.58 dB, which improved to 75.63 dB after denoising. In the second real device image sample, the SNR was 33.77 dB before denoising, which improved to 67.01 dB after denoising. The grayscale fluctuations in the four background regions were significantly reduced after denoising, and the background noise distribution curve changed from a broad distribution to a more concentrated and narrow distribution, indicating that the model can effectively reduce background random noise in real device magnetic resonance images.

[0184] Furthermore, to verify the stability of the method proposed in this invention across different real-world device images, this embodiment selected 41 real-world device magnetic resonance images for statistical analysis. The results are as follows: Figure 6 As shown.

[0185] Figure 6This is a statistical analysis chart of the signal-to-noise ratio (SNR) and background noise of real device magnetic resonance images before and after denoising in an embodiment of the present invention. (a) is a comparison chart of the SNR before and after denoising calculated based on the background region. The horizontal axis represents the Raw image and the Denoised image, respectively, and the vertical axis represents the SNR estimated based on the background region, in dB. Data points of the same image before and after denoising are connected by gray lines. (b) is a comparison chart of the background noise level before and after denoising. The horizontal axis represents the Raw image and the Denoised image, respectively, and the vertical axis represents the background noise level. (c) is a distribution chart of the SNR improvement for each real device image. The SNR improvement is calculated by subtracting the Raw image SNR from the Denoised image SNR.

[0186] This embodiment selected 41 real device MRI images for paired statistical analysis. In the figure, the white rhombus represents the median, and the error bars represent statistical intervals centered on the median with lengths corresponding to the interquartile range. The results show that after denoising, the signal-to-noise ratio (SNR) based on background region estimation is significantly improved, and the background noise level is significantly reduced. The statistical test results for both paired comparisons are p < 0.0001. The overall SNR improvement in the 41 real device images is positive, with a median of 15.68 dB. The main distribution interval shown in the figure is 11.49–20.41 dB, indicating that the IKSFormer-light model has stable background noise suppression and SNR improvement effects on real device images.

[0187] Example 5 A deep learning-based image denoising system for ultra-low field magnetic resonance imaging includes: Data acquisition and preprocessing module: used to acquire the original ultra-low field magnetic resonance image to be denoised and preprocess it to obtain a standardized magnetic resonance image; Denoising module: Used to input standardized magnetic resonance images into a pre-trained deep learning denoising network based on image domain-frequency domain-spatial domain co-modeling to obtain denoised ultra-low field magnetic resonance images; The noise reduction module specifically includes: Shallow feature extraction module: used to input standardized magnetic resonance images into convolutional layers for shallow feature extraction to obtain initial spatial features; Deep extraction module: used to input the initial spatial features into the multi-scale coding part, and perform step-by-step feature extraction through multiple coding stages of different scales to obtain deep coded features; Bottleneck Feature Extraction Module: This module takes deep-encoded features as input to the bottleneck feature extraction section to obtain the bottleneck features. Decoding module: Used to decode bottleneck features through multi-scale decoding to obtain enhanced decoded features; Processing module: Used to input enhanced decoding features into convolutional layers, obtain network prediction residuals, and use residual reconstruction to obtain denoised ultra-low field magnetic resonance images.

[0188] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0189] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0190] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0191] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0192] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit its scope of protection. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that after reading the present invention, they can still make various changes, modifications or equivalent substitutions to the specific implementation of the invention, but these changes, modifications or equivalent substitutions are all within the scope of protection of the pending claims of the invention.

Claims

1. An ultra-low-field magnetic resonance deep learning image denoising method, characterized in that, Includes the following steps: Step 1: Acquire the original ultra-low field magnetic resonance image to be denoised and perform preprocessing to obtain a standardized magnetic resonance image; Step 2: Input the standardized magnetic resonance image into a pre-trained deep learning denoising network based on image domain-frequency domain-spatial domain co-modeling to obtain a denoised ultra-low field magnetic resonance image, specifically including: Step 2.1: Input the standardized magnetic resonance image into the convolutional layer for shallow feature extraction to obtain initial spatial features; Step 2.2: Input the initial spatial features into the multi-scale coding part, and perform step-by-step feature extraction through multiple coding stages of different scales to obtain deep coding features; Step 2.3: Input the deep coding features into the bottleneck feature extraction part to obtain the bottleneck features; Step 2.4: Obtain enhanced decoding features by decoding the bottleneck features at multiple scales; Step 2.5: Input the enhanced decoding features into the convolutional layer to obtain the network prediction residual, and use the residual reconstruction method to obtain the denoised ultra-low field magnetic resonance image.

2. The ultra-low-field magnetic resonance deep learning image denoising method according to claim 1, characterized in that, Step 1 specifically includes: The acquired raw magnetic resonance images are subjected to amplitude conversion to convert them into single-channel magnetic resonance grayscale images: in, This indicates an image grayscale conversion operation. This represents the converted single-channel magnetic resonance grayscale image; The size of the single-channel magnetic resonance grayscale image is adjusted: in, This indicates an image resizing operation. The image after resizing; The resized image is then normalized to limit the pixel values ​​of the input image to the range of 0 to 1. in, and These represent the minimum and maximum grayscale values ​​of the image, respectively. Represents a pixel range constraint function; After the above processing, a standardized magnetic resonance image is obtained. ,in, and These represent the height and width of the input image, respectively.

3. The ultra-low field magnetic resonance deep learning image denoising method according to claim 1, characterized in that, Step 2.1 specifically involves: standardizing the magnetic resonance image. Input a 3×3 convolutional layer for shallow feature extraction to obtain the initial space. feature: in, This represents a 3×3 convolution operation. This represents the initial spatial features obtained by applying shallow convolution to the input image; The multi-scale encoding part in step 2.2 includes multiple sequentially connected encoding stages. Each encoding stage consists of an IKSBlock feature processing unit and a scale transformation operation, specifically including: In the first-level encoding stage, the basic spatial features are input into the first-level IKS Block feature processing unit: in, This represents the output features of the first-level encoding stage. This indicates that the IKS Block feature processing process is performed twice consecutively in the first-level encoding stage. Subsequently, the feature space resolution is reduced through a first 2×2 downsampling operation to obtain scale-compressed features: in, This indicates the first scaling operation. These are features after scale compression; The scale-compressed features are then input into the second-level encoding stage: in, This represents the output features of the second-level encoding stage. This indicates that the IKS Block feature processing process is performed twice consecutively in the second-level encoding stage. The feature space resolution is reduced by a second 2×2 downsampling operation to obtain deep encoded features: in, This indicates the second scaling operation. These are the deep coding features after the second scaling transformation.

4. The ultra-low field magnetic resonance deep learning image denoising method according to claim 3, characterized in that, Step 2.3 specifically involves inputting the deep coding features after two scale transformations into the bottleneck feature extraction part to obtain the bottleneck features: in, This represents a bottleneck feature processing process consisting of four consecutive IKS blocks. Represents deep semantic features of the network; Step 2.4 specifically includes: First, the bottleneck features are upsampled: in, This indicates the first upsampling operation. These are the decoding features after the first upsampling; Subsequently, the decoded features after the first upsampling... Scale features corresponding to the encoding stage Fusion via skip connections: in, This indicates a feature concatenation operation. This represents the first-stage decoded features after fusion via skip connections; The first-stage decoded features after fusion are processed by IKS Blocks in two consecutive decoding stages: in, This indicates that the IKS Block feature processing is performed twice consecutively during the first-level decoding stage. This is an enhanced decoding feature resulting from two consecutive IKS Block feature processing steps in the first-level decoding stage; Enhanced decoding features Perform a second upsampling: in, This indicates the first upsampling operation. These are the decoding features after the second upsampling; The decoded features after the second upsampling Scale features corresponding to the encoding stage Fusion via skip connections: in, This represents the second-stage decoded features after fusion via skip connections; The fused second-stage decoded features are then processed by IKS Blocks in two consecutive decoding stages: in, This indicates that the IKS Block feature processing process is performed twice consecutively in the second-level decoding stage. This is an enhanced decoding feature resulting from two consecutive IKS Block feature processing steps in the second-level decoding stage; Step 2.5 specifically includes: The enhanced decoding features are input into a 3×3 convolutional layer to obtain the network prediction residuals: in, This represents the noise correction residual information obtained from network learning; Subsequently, the final network output is obtained using residual reconstruction: in, This represents the input standardized magnetic resonance image. This represents a denoised ultra-low field magnetic resonance image.

5. A method for denoising ultra-low field magnetic resonance deep learning images according to claim 3, characterized in that, The IKS Block feature processing unit includes an image domain local feature processing unit, a K-space frequency feature processing unit, and a spatial Transformer feature processing unit. The basic spatial features are input into the local feature processing unit of the image domain, and the image domain enhancement features are extracted through convolution operation. The processing flow of the local feature processing unit of the image domain includes layer normalization, pointwise convolution, depthwise separable convolution and nonlinear activation function. The image domain enhancement features are input into the K-space frequency feature processing unit to obtain the frequency-enhanced features. The frequency-enhanced features are input into the space Transformer feature processing unit to obtain the IKS Block output features.

6. The ultra-low field magnetic resonance deep learning image denoising method according to claim 5, characterized in that, The step of inputting basic spatial features into the image domain local feature processing unit and extracting image domain enhancement features through convolution operations specifically includes: First, the basic spatial features are subjected to layer normalization: in, Presentation layer normalization operation, These are the features after layer normalization; The feature channels are adjusted using 1×1 convolution: in, This represents a 1×1 convolution operation, used to implement feature dimension mapping. This indicates the intermediate features after channel adjustment; Local spatial structure information is extracted using 3×3 depthwise convolution. in, This indicates a 3×3 depth separable convolution operation. Indicates local spatial structure enhancement features; Enhancing the nonlinear expressive power of the network through the GELU activation function: And the feature dimension is restored by 1×1 convolution: in, Indicates nonlinear enhancement features, Represents the output of local features in the image domain. This indicates an output mapping convolution operation; Image domain enhancement features are obtained using residual connectivity: in, This represents the learnable residual weight parameters. This represents the output features after image domain feature enhancement.

7. The ultra-low field magnetic resonance deep learning image denoising method according to claim 6, characterized in that, The step of inputting the image domain enhanced features into the K-space frequency feature processing unit to obtain the frequency-enhanced features specifically includes: Output features after image domain feature enhancement Perform a two-dimensional fast Fourier transform: in, This represents a two-dimensional Fast Fourier Transform. This represents the frequency domain characteristics after conversion; Since the Fourier transform result is in complex form, it can be further expressed as: in, Represents the real part of the spectrum. Indicates the imaginary part of the spectrum; Concatenate the real and imaginary parts of the spectrum along the channel dimension: in, This represents the splicing feature of the real and imaginary parts in the frequency domain; Frequency feature mapping is achieved using 1×1 convolution and the GELU activation function. in, This indicates the first channel mapping layer. This indicates the second channel mapping layer. This represents the spectral characteristics after frequency domain feature transformation; The frequency domain information is converted back to the spatial domain using the inverse Fourier transform: in, This represents the spatial characteristics after frequency domain enhancement. This represents the inverse fast Fourier transform; Residual fusion is performed between the frequency-enhanced features and the output features after image domain feature enhancement: in, Represents the frequency domain fusion weights. This represents the features after fusing K-space frequency information.

8. A method for denoising ultra-low field magnetic resonance deep learning images according to claim 7, characterized in that, The step of inputting the frequency-enhanced features into the space Transformer feature processing unit to obtain the IKS Block output features specifically includes: The features after fusing K spatial frequency information are then subjected to layer normalization: in, This represents the frequency enhancement features after layer normalization. Calculate spatial relationships using a multi-head self-attention mechanism within a window: in, This indicates a multi-head self-attention operation in the window; Indicates the attention residual weight; This represents the features after spatial attention enhancement; The spatially attention-enhanced features are input into a gated deep convolutional feedforward network, where convolution operations enhance the local spatial representation capabilities. in, , Indicates the gating branch feature; This represents element-wise multiplication; Indicates output mapping convolution; Indicates the feedforward network fusion weights; This represents the features enhanced through spatial modeling. Finally, the IKS Block output features are obtained: in, This represents the final feature representation after co-enhancement in the image domain, K-space frequency domain, and spatial domain.

9. The ultra-low field magnetic resonance deep learning image denoising method according to claim 1, characterized in that, The training method for the deep learning denoising network based on image domain-frequency domain-spatial domain co-modeling includes the following steps: High-field magnetic resonance images are acquired as training target reference images and preprocessed. Composite degradation processing is performed on the preprocessed training target reference image to obtain a composite degradation image without random noise; Noise simulation is performed based on the aforementioned composite degradation image without random noise to obtain an ultra-low field magnetic resonance simulation degradation image; A paired training dataset is constructed based on the ultra-low field magnetic resonance simulated degradation images and the corresponding training target reference images; Based on the paired training dataset, a deep learning denoising network based on image domain-frequency domain-spatial domain co-modeling is forward propagated to obtain a predicted denoised image. The joint loss is calculated based on the predicted denoised image and the training target reference image, and the network parameters are updated accordingly. The optimal network parameters are determined based on the iterative training results, resulting in a trained deep learning denoising network based on collaborative modeling of the image domain, frequency domain, and spatial domain.

10. A deep learning image denoising system for ultra-low field magnetic resonance imaging, characterized in that, include: Data acquisition and preprocessing module: used to acquire the original ultra-low field magnetic resonance image to be denoised and preprocess it to obtain a standardized magnetic resonance image; Denoising module: Used to input standardized magnetic resonance images into a pre-trained deep learning denoising network based on image domain-frequency domain-spatial domain co-modeling to obtain denoised ultra-low field magnetic resonance images; The noise reduction module specifically includes: Shallow feature extraction module: used to input standardized magnetic resonance images into convolutional layers for shallow feature extraction to obtain initial spatial features; Deep extraction module: used to input the initial spatial features into the multi-scale coding part, and perform step-by-step feature extraction through multiple coding stages of different scales to obtain deep coded features; Bottleneck Feature Extraction Module: This module takes deep-encoded features as input to the bottleneck feature extraction section to obtain the bottleneck features. Decoding module: Used to decode bottleneck features through multi-scale decoding to obtain enhanced decoded features; Processing module: Used to input enhanced decoding features into convolutional layers, obtain network prediction residuals, and use residual reconstruction to obtain denoised ultra-low field magnetic resonance images.