Self-supervised based medical MRI image denoising method, device, system, equipment, medium and product

Abstract

This invention discloses a self-supervised medical MRI image denoising method, apparatus, system, device, medium, and product, relating to the field of image enhancement technology. The method first converts the medical MRI image signal to the image domain to obtain three-dimensional image data. Then, it extracts all two-dimensional images of both sides of the slice as pure noise images and all two-dimensional images of the non-side slice as images to be denoised. Next, it uses the pure noise images to add noise to the images to be denoised, generating an initial noise-enhanced image. Using this as input and the images to be denoised as output, it trains an initial denoising model to obtain an initial denoised image. Finally, through an iterative process: generating labeled images based on the current denoised image, adding noise to the pure noise images to obtain sample images, and training the iterative denoising model until a preset maximum number of iterations is reached, the final denoised image is obtained, and the three-dimensional image data is updated. This results in a significant denoising effect.

Description

Technical Field

[0001] This invention belongs to the field of image enhancement technology, specifically relating to a self-supervised medical MRI image denoising method, device, system, equipment, medium, and product. Background Technology

[0002] Magnetic resonance imaging (MRI) plays a vital role in human medicine due to its non-ionizing, radiation-free nature and high contrast for soft tissue imaging. With technological advancements, multi-nuclear MRI has also developed accordingly. In multi-nuclear MRI, commonly used nuclides include… 23 Sodium (Na) is one of the most important electrolytes in human physiology. 23 Sodium (Na+) plays a crucial role in osmotic regulation and cell physiology. Transmembrane transport is maintained through the active transport of Na+ / K+-ATPase via the sodium-potassium pump. 23 The sodium concentration gradient differs by more than 10-fold between intracellular (10-15 mm) and extracellular (140-150 mm) concentrations. Under pathological conditions, impaired cellular metabolism or membrane integrity may disrupt sodium pump function, leading to tissue-specific [sodium concentration gradients]. 23 Disruptions in sodium homeostasis and abnormal ion concentrations. These changes in tissue sodium concentration (TSC) reflect underlying metabolic dysfunction and have been observed in a variety of neurological diseases, including brain tumors, stroke, multiple sclerosis, and epilepsy. 23 Na magnetic resonance imaging (NMR) can non-invasively measure TSC, providing a valuable tool for studying these pathologies. However, compared to proton (¹H) imaging, 23 Na MRI faces greater technical challenges because its inherently low signal-to-noise ratio (SNR) stems from a lower cyclotron ratio, lower in vivo concentration, and shorter transverse relaxation time.

[0003] To further improve the quality of magnetic resonance imaging, especially to improve 23 To improve the quality of Na NMR imaging, various post-processing and reconstruction strategies have been proposed. Broadly speaking, these denoising methods fall into two categories: deep learning-based methods and traditional non-learning techniques.

[0004] Deep learning models, with their powerful nonlinear modeling capabilities and automatic feature extraction abilities, have become important tools in the field of MRI image reconstruction and denoising. In recent years, research has attempted to introduce deep learning methods into these areas. 23Na MRI image reconstruction and denoising techniques. Adlung et al. employed a U-Net-based network structure, using fully sampled data as supervision, to address highly undersampled MRI images. 23 Reconstruction of Na MRI data effectively suppresses undersampling artifacts while shortening scan time; however, this method primarily focuses on reconstruction fidelity and does not explicitly address the inherent noise problem remaining in fully sampled data, thus limiting the signal-to-noise ratio of the reconstructed image. On the other hand, Baker et al. constructed synthetic low signal-to-noise ratio data by adding Gaussian noise to high-quality¹H MRI k-space and trained a convolutional neural network on this basis for denoising. Although this method achieved good results under controlled conditions, it relies on an idealized noise model and large-scale paired "clean-noisy" datasets, often struggling to handle real-world scenarios. 23 The complex and varied noise distribution in Na MRI significantly limits the generalization ability. More fundamentally, supervised learning-based denoising methods are not suitable for... 23 Na MRI image reconstruction and denoising: On the one hand 23 Narration imaging inherently has a low signal-to-noise ratio; even fully sampled or multi-averaged data still contain significant noise, making it difficult to obtain a truly "noise-free" reference image. Furthermore, the long scan time limits the acquisition of large-scale, high-quality datasets, further restricting the training of supervised networks. Therefore, self-supervised denoising methods that do not rely on reference images have become a potential solution. However, existing self-supervised models are mostly based on simple noise assumptions. In multi-channel MRI imaging, the inter-channel coupling noise introduced by phased array coils is extremely complex, breaking the presuppositions of these models; this also leads to limitations in current self-supervised methods. 23 It has not yet been successfully applied in complex imaging scenarios such as Na MRI.

[0005] Traditional non-learning-based denoising techniques typically rely on manually crafted priors to formulate noise suppression algorithms. While they perform well in conventional MRI denoising, their denoising capabilities become insufficient when extended to multi-core MRI denoising scenarios. For example, Madelin et al. were among the first to apply compressed sensing (CS) to MRI. 23 One research group working on sodium MRI halved the acquisition time while preserving reconstruction accuracy to some extent. However, this method often resulted in residual blurring and loss of detail, limiting its applicability for quantitative sodium imaging. Lachner et al. and Gnahm et al., through... 1Incorporating anatomical priors into the CS framework for h MRI addresses this issue, improving fidelity but requiring additional multimodal data acquisition and image registration. Benkhedah et al. utilized an adaptive combined reconstruction method to mitigate correlated noise between coil channels, but this necessitates additional noisy scans and extended total scan time. More recently, Christensen et al. applied Global-Local Higher-Order Singular Value Decomposition (GL-HOSVD) and Tensor Marchenko-Pastur Principal Component Analysis (tMPPCA) to denoise various x-nucleus human data. While effective, both methods introduce subtle artifacts and varying degrees of image smoothing.

[0006] tMPPCA (Tensor Marchenko-Pastur Principal Component Analysis) is an improved denoising method based on High-Order Singular Value Decomposition (HOSVD). By introducing the Marchenko-Pastur distribution to automatically estimate the signal rank, it reduces the need for user-defined parameters and improves the objectivity and robustness of denoising. This method effectively utilizes data redundancy by recursively decomposing the tensor structure of multidimensional data, making it particularly suitable for small data blocks and high-dimensional data (such as multi-echo diffusion MRI). However, tMPPCA still has limitations in MRI denoising: although it can significantly reduce background noise, residual analysis shows that the signal distribution in the brain may be skewed (such as abnormally enhanced signal intensity in the ventricular region), and artifacts are introduced in some slices. In addition, for spatial variability noise caused by parallel imaging reconstruction in clinical data, tMPPCA needs to rely on additional noise level maps or local estimation methods; otherwise, it may not be able to fully adapt to complex noise distributions.

[0007] GL-HOSVD (Global-Local High-Order Singular Value Decomposition) is a hybrid denoising algorithm that combines global pre-filtering with local block processing. Pre-denoising in the global HOSVD stage guides the local HOSVD stage, reducing fringe artifacts generated by local methods under low signal-to-noise ratio conditions. This method performs well when the noise exhibits Gaussian characteristics, but its application in... 23 Na MRI denoising has significant limitations: experimental results show that the brain signal distribution may change after denoising (e.g., abnormal signals in the ventricular region), and brain contour deviations still exist in the residual images, indicating that the denoising process introduces systematic errors. Furthermore, GL-HOSVD relies on user-defined threshold parameters (such as k_global and k_local), which need to be adjusted empirically. In clinical data, noise characteristics may be complicated by preprocessing steps (such as parallel imaging or geometric correction), leading to limited parameter generalization. Meanwhile, in... 23 Artifacts were observed in Na MRI denoising, further limiting its reliability.

[0008] In summary, existing tMPPCA and GL-HOSVD techniques demonstrate excellent performance in conventional MRI denoising; however, they lag behind in multi-core MRI denoising (especially in...). 23 In MRI denoising, varying degrees of image blurring occur, resulting in limited denoising capability in areas with signal intensity, even though noise can be removed. Furthermore, existing deep learning-based MRI image denoising techniques suffer from limitations due to supervised learning, such as the need for a sharp reference image. Self-supervised learning, which eliminates the need for a sharp reference image, offers a promising direction for MRI image denoising. Therefore, how to provide a new medical MRI image denoising scheme with self-supervised learning characteristics, so as to not only denoise conventional MRI (e.g., MRI with signal intensity), but also address these limitations? 1 Effective denoising of H MRI and more complex multinucleated MRI is a topic that urgently needs to be studied by those skilled in the art. Summary of the Invention

[0009] The purpose of this invention is to provide a self-supervised medical MRI image denoising method, device, system, computer equipment, computer-readable storage medium, and computer program product to solve the problems of blurred images after denoising and limited denoising ability for signal areas in existing tMPPCA and GL-HOSVD technologies when applied to MRI denoising, as well as the problem that existing deep learning model-based MRI image denoising technologies require clear reference images due to supervised learning.

[0010] To achieve the above objectives, the present invention adopts the following technical solution: Firstly, a self-supervised medical MRI image denoising method is provided, comprising the following steps S1 to S8: S1. Receive medical MRI image signals from multiple subjects, and then execute step S2; S2. For each of the multiple subjects, the corresponding medical MRI image signal is converted to the image domain to obtain three-dimensional image data containing multiple two-dimensional images and corresponding images. Then, all two-dimensional images belonging to the two-sided slices are extracted from the three-dimensional image data as the corresponding pure noise images, and all two-dimensional images not belonging to the two-sided slices are extracted from the three-dimensional image data as the corresponding images to be denoised. Finally, step S3 is executed. S3. Summarize all the pure noise images of each subject to obtain a pure noise image set, and summarize all the images to be denoised of each subject to obtain a denoised image set, and then execute step S4; S4. Apply the pure noise images in the pure noise image set to each image to be denoised in the image set to be denoised, respectively, to obtain each initial noise-enhanced image corresponding to each image to be denoised. Then, use all the initial noise-enhanced images as model inputs and all the images to be denoised as model outputs to train the initial denoising model using a learning network model to obtain an initial denoising model. Then, import each image to be denoised into the initial denoising model to obtain each initial denoised image corresponding to each image to be denoised. Then, execute step S5. S5. Number of iterations Initialize to 1, and then obtain the result based on the initial denoised image. The first time The label image used in the next iteration, along with the pure noise images applied to the pure noise image set, is used to add noise to the initial denoised image, resulting in... The first time The next iteration uses sample images, and the initial denoising model is used as the basis for... The first time The next iteration of denoising uses a learned network model, and then step S6 is executed; S6. All the aforementioned first... The next iteration uses sample images as model input and all the aforementioned... The next iteration uses the labeled image as the model output term, for the first iteration... The denoising iteration uses a learned network model for training, resulting in the... The iteration denoising model is used, and each image to be denoised is imported into the first iteration denoising model. The iteration denoising model obtains each of the denoised images corresponding to each of the images to be denoised. After the denoising is removed in the next iteration, step S7 is then executed. S7. Determine the number of iterations. Has the preset maximum number of iterations been reached? If so, proceed to step S8; otherwise, based on the previous step... The image after the denoising iteration is obtained as follows: The labeled image used in the next iteration, and the pure noise image applied to the pure noise image set for the first iteration. After the denoising process in the 1st iteration, the image is subjected to noise addition to obtain the 2nd iteration. The sample image used in the next iteration, and the first... The iteration denoising model is used as the first The next iteration of denoising uses a learned network model, and then the number of iterations is... Increment by 1 and return to step S6; S8. The first The image after each iteration of denoising is used as the final denoised image of the corresponding image to be denoised, and the three-dimensional image data of each subject is replaced and updated.

[0011] Based on the above-mentioned invention, a novel medical MRI image denoising scheme with self-supervised learning characteristics is provided. First, medical MRI image signals from multiple subjects are received and converted to the image domain to obtain three-dimensional image data containing multiple layers of two-dimensional images. Then, all two-dimensional images of the bilateral slices are extracted from the three-dimensional image data as pure noise images, and all two-dimensional images of the non-bilateral slices are extracted as images to be denoised. Next, the pure noise images are used to add noise to the images to be denoised, generating an initial noise-enhanced image. This initial noise-enhanced image is used as input and the images to be denoised as output to train an initial denoising model, obtaining an initial denoised image. Finally, through an iterative process: generating a label image based on the current denoised image, adding noise to the pure noise image to obtain a sample image, training the iterative denoising model until a preset maximum number of iterations is reached, obtaining the final denoised image, and updating the three-dimensional image data. This approach eliminates the need for any assumptions about the noise pattern, cleverly combines 3D MRI imaging sequences, and fully utilizes the acquired noise layer images to achieve self-supervised learning, resulting in significant denoising performance. Its denoising performance surpasses existing advanced methods such as tMPPCA and GL-HOSVD, making it convenient for practical application and promotion.

[0012] In one possible design, when the medical MRI image signal is obtained by scanning the subject using a three-dimensional density-adapted radial sequence, for each of the plurality of subjects, the corresponding medical MRI image signal is converted to the image domain to obtain three-dimensional image data containing multiple layers of two-dimensional images, including: For each of the multiple subjects, a non-uniform fast Fourier transform is used to convert the corresponding medical MRI image signal to the image domain to obtain three-dimensional image data containing multiple two-dimensional images.

[0013] In one possible design, the pure noise images in the pure noise image set are used to add noise to each image in the image set to be denoised, respectively, to obtain initial noise-enhanced images corresponding one-to-one with each image to be denoised, including: For each image to be denoised in the set of images to be denoised, a pure noise image is randomly selected from the set of pure noise images. Then, the pure noise image is added to the corresponding image to be denoised using a layer overlay method to obtain the corresponding initial noise enhancement image.

[0014] In one possible design, the image obtained from the initial denoised image is... The first time The label images used in the next iteration include: Based on the initial denoised image, the result is obtained according to the following formula. The first time The next iteration uses a label image. :

[0015] In the formula, This represents the preset weight coefficient corresponding to the first iteration, and is a pure decimal. This represents the image to be denoised, corresponding to the initial denoised image. This represents the functional form of the initial denoising model. This refers to the initial denoised image.

[0016] In one possible design, based on the first The image after the denoising iteration is obtained as follows: The label images used in the next iteration include: Based on the first The image after the denoising iteration is obtained according to the following formula. The next iteration uses a label image. :

[0017] In the formula, Indicates the relationship with the first The weight coefficients corresponding to the next iteration. Indicates the first Functional form of the next iteration denoising model Indicates the first Image after the denoising process in the second iteration Represents a preset positive integer. This represents the function that takes the maximum value.

[0018] Secondly, a self-supervised medical MRI image denoising device is provided, including an image signal receiving unit, an image data conversion unit, an image set aggregation unit, an initial image denoising unit, an iterative initialization unit, an iterative image denoising unit, a judgment and processing unit, and an image data updating unit. The image signal receiving unit is communicatively connected to the image data conversion unit and is used to receive medical MRI image signals from multiple subjects, and then activate the image data conversion unit. The image data conversion unit is communicatively connected to the image collection and aggregation unit. It is used to convert the corresponding medical MRI image signal to the image domain for each of the multiple subjects to obtain three-dimensional image data containing multiple two-dimensional images and corresponding images. Then, it extracts all two-dimensional images belonging to the two-sided slices from the three-dimensional image data as the corresponding pure noise images, and extracts all two-dimensional images not belonging to the two-sided slices from the three-dimensional image data as the corresponding images to be denoised. Finally, it starts the image collection and aggregation unit. The image set aggregation unit is communicatively connected to the image initial denoising unit. It is used to aggregate all pure noise images of each subject to obtain a pure noise image set, and to aggregate all images to be denoised of each subject to obtain a denoised image set, and then start the image initial denoising unit. The image initial denoising unit is communicatively connected to the iterative initialization unit. It is used to apply noise processing to each image to be denoised in the set of pure noise images, respectively, to obtain each initial noise enhancement image corresponding to each image to be denoised. Then, all the initial noise enhancement images are used as model inputs and all the images to be denoised are used as model outputs to train the initial denoising model using a learning network model, to obtain an initial denoising model. Each image to be denoised is then imported into the initial denoising model to obtain each initial denoised image corresponding to each image to be denoised. Then, the iterative initialization unit is started. The iterative initialization unit is communicatively connected to the image iterative denoising unit and is used to initialize the number of iterations. Initialize to 1, and then obtain the result based on the initial denoised image. The first time The label image used in the next iteration, along with the pure noise images applied to the pure noise image set, is used to add noise to the initial denoised image, resulting in... The first time The next iteration uses sample images, and the initial denoising model is used as the basis for... The first time The next iteration of denoising uses a learned network model, and then the image iterative denoising unit is started; The image iterative denoising unit is communicatively connected to the judgment and processing unit, and is used to process all the first... The next iteration uses sample images as model input and all the aforementioned... The next iteration uses the labeled image as the model output term, for the first iteration... The denoising iteration uses a learned network model for training, resulting in the... The iteration denoising model is used, and each image to be denoised is imported into the first iteration denoising model. The iteration denoising model obtains each of the denoised images corresponding to each of the images to be denoised. After the denoising is performed in the next iteration, the judgment and processing unit is then activated. The judgment and processing unit is communicatively connected to the image data update unit and the image iterative denoising unit, respectively, and is used to determine the number of iterations. If the preset maximum number of iterations has been reached, then the image data update unit is activated; otherwise, based on the first iteration... The image after the denoising iteration is obtained as follows: The labeled image used in the next iteration, and the pure noise image applied to the pure noise image set for the first iteration. After the denoising process in the 1st iteration, the image is subjected to noise addition to obtain the 2nd iteration. The sample image used in the next iteration, and the first... The iteration denoising model is used as the first The next iteration of denoising uses a learned network model, and then the number of iterations is... Increment by 1 and restart the image iterative denoising unit; The image data update unit is communicatively connected to the image data conversion unit, and is used to update the image data of the first image data. The image after each iteration of denoising is used as the final denoised image of the corresponding image to be denoised, and the three-dimensional image data of each subject is replaced and updated.

[0019] Thirdly, the present invention provides a medical MRI image denoising system, including a magnetic resonance imaging instrument and a host computer connected in communication, wherein the magnetic resonance imaging instrument includes a scanning module and a magnetic resonance receiving coil. The scanning module is used to scan the subject by transmitting a sequence suitable for MRI imaging through a radio frequency transmitting coil, so as to excite the hydrogen atoms in the subject to produce a resonance phenomenon. The magnetic resonance receiving coil is used to receive and transmit medical MRI image signals to the host computer during the scanning of the subject. The host computer is used to execute the medical MRI image denoising method as described in the first aspect or any possible design in the first aspect.

[0020] Fourthly, the present invention provides a computer device comprising a storage module, a processing module, and a transceiver module connected in sequence for communication, wherein the storage module is used to store a computer program, the transceiver module is used to send and receive messages, and the processing module is used to read the computer program and execute the medical MRI image denoising method as described in the first aspect or any possible design in the first aspect.

[0021] Fifthly, the present invention provides a computer-readable storage medium storing instructions that, when executed on a computer, perform the medical MRI image denoising method as described in the first aspect or any possible design of the first aspect.

[0022] In a sixth aspect, the present invention provides a computer program product, including a computer program or instructions, which, when executed by a computer, implement the medical MRI image denoising method as described in the first aspect or any possible design in the first aspect.

[0023] The beneficial effects of the above scheme are: (1) This invention creatively provides a novel medical MRI image denoising scheme with self-supervised learning characteristics. First, medical MRI image signals from multiple subjects are received and converted to the image domain to obtain three-dimensional image data containing multiple two-dimensional images. Then, all two-dimensional images of the two-sided slices are extracted from the three-dimensional image data as pure noise images, and all two-dimensional images of the non-two-sided slices are extracted as images to be denoised. Then, the pure noise images are used to add noise to the images to be denoised to generate an initial noise-enhanced image. This image is used as input and the images to be denoised as output to train the initial denoising model to obtain the initial denoised image. Finally, through an iterative process: a label image is generated based on the current denoised image, and a sample image is obtained by adding noise to the pure noise image. The iterative denoising model is trained until the preset maximum number of iterations is reached to obtain the final denoised image and update the three-dimensional image data. Thus, no assumptions about the noise pattern are required. The 3D MRI imaging sequence is cleverly combined with the acquired noise layer images to achieve self-supervised learning, and the denoising effect is significant. That is, its denoising performance is better than existing advanced methods, such as tMPPCA and GL-HOSVD, which is convenient for practical application and promotion. Attached Figure Description

[0024] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0025] Figure 1 This is a flowchart illustrating the self-supervised medical MRI image denoising method provided in the embodiments of this application.

[0026] Figure 2 The figure shows a comparison of the denoising effect of the self-supervised medical MRI image denoising method provided in the embodiments of this application with the tMPPCA method and the GL-HOSVD method.

[0027] Figure 3 Provided for the embodiments of this application Figure 2 Example diagram showing the relationship between the central label image and the local image.

[0028] Figure 4 A schematic diagram of the structure of a self-supervised medical MRI image denoising device provided in an embodiment of this application.

[0029] Figure 5 A schematic diagram of the structure of a self-supervised medical MRI image denoising system provided in an embodiment of this application.

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

[0031] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the present invention will be briefly introduced below in conjunction with the accompanying drawings and descriptions of the embodiments or the prior art. Obviously, the following description of the structure of the accompanying drawings is only some embodiments of the present invention. For those skilled in the art, other embodiments can be obtained based on these embodiments without creative effort. It should be noted that the description of these embodiments is for the purpose of helping to understand the present invention, but does not constitute a limitation of the present invention.

[0032] It should be understood that although the terms "first" and "second", etc., may be used herein to describe various objects, these objects should not be limited by these terms. These terms are only used to distinguish one object from another. For example, the first object may be referred to as the second object, and similarly, the second object may be referred to as the first object, without departing from the scope of the exemplary embodiments of the invention.

[0033] It should be understood that the term "and / or" that may appear in this document is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can mean: A exists alone, B exists alone, or A and B exist simultaneously. Another example is A, B and / or C, which can mean that any one of A, B, and C or any combination thereof exists. The term " / and" that may appear in this document describes another relationship between related objects, indicating that two relationships can exist. For example, A / and B can mean: A exists alone or A and B exist simultaneously. In addition, the character " / " that may appear in this document generally indicates that the related objects before and after it are in an "or" relationship.

[0034] Example like Figures 1-3As shown, the self-supervised medical MRI image denoising method provided in the first aspect of this embodiment can be executed, but is not limited to, by a computer device with certain computing resources and a communication connection to a magnetic resonance imaging (MRI) scanner. The MRI scanner includes, but is not limited to, a scanning module and a magnetic resonance receiving coil. The scanning module is used to scan the subject (e.g., a patient) by transmitting a sequence suitable for MRI imaging through a radio frequency transmitting coil, so as to excite hydrogen atoms in the subject to resonate. The magnetic resonance receiving coil is used to receive and transmit medical MRI image signals to the computer device during the scanning of the subject. Figure 1 As shown, the medical MRI image denoising method includes, but is not limited to, the following steps S1 to S8.

[0035] S1. Receive medical MRI image signals from multiple subjects, and then execute step S2.

[0036] In step S1, the medical MRI image signal originates from the magnetic resonance receiving coil and can be conventionally received using existing wired communication technology. Furthermore, the specific examples of the medical MRI image signal, but not limited to medical... 23 The MRI image signal of the subject, and the subject may be, but is not limited to, a patient or a patient, and the medical MRI image signals of the multiple subjects may be received asynchronously, that is, the medical MRI image signal of one subject is received first, and then the medical MRI image signal of the next subject is received.

[0037] S2. For each of the multiple subjects, the corresponding medical MRI image signal is converted to the image domain to obtain three-dimensional image data containing multiple two-dimensional images. Then, all two-dimensional images belonging to the two-sided slices are extracted from the three-dimensional image data as the corresponding pure noise images, and all two-dimensional images not belonging to the two-sided slices are extracted from the three-dimensional image data as the corresponding images to be denoised. Finally, step S3 is executed.

[0038] In step S2, the size of the three-dimensional image data can be, for example, 50×80×80, meaning it contains 50 layers of two-dimensional images, each containing 80×80 pixels. Specifically, when the medical MRI image signal is obtained by scanning the subject using a three-dimensional density-adapted radial sequence, for each of the multiple subjects, the corresponding medical MRI image signal is converted to the image domain to obtain three-dimensional image data containing multiple layers of two-dimensional images. This includes: for each of the multiple subjects, using a non-uniform fast Fourier transform to convert the corresponding medical MRI image signal to the image domain to obtain three-dimensional image data containing multiple layers of two-dimensional images. The three-dimensional density-adapted radial sequence (DA-3DPR) is an existing scanning sequence, and because this sequence uses hard pulse excitation and does not select layers, the imaging area is larger than the target area; for example, the target area is the human brain region and its size is 18 cm. 3 The imaging area size can then be set to 25×25×25cm. 3 This ensures that the human brain is located within the imaging area; the Non-Uniform Fast Fourier Transform (NUFFT) is also an existing technology and will not be described in detail here. Furthermore, since the layers located in the two side slices (such as layers 1-5 and 46-50) in the three-dimensional image data do not contain images of any objects, they can be considered as pure noise images (which are generally discarded in traditional schemes, but will be fully utilized for denoising in this embodiment); and other layers not belonging to the two side slices (such as layers 6-45 corresponding to the middle slice) will be considered as noisy images to be denoised.

[0039] S3. Summarize all the pure noise images of each subject to obtain a pure noise image set, and summarize all the images to be denoised of each subject to obtain a denoised image set, and then execute step S4.

[0040] In step S3, for example, if the total number of objects of the multiple subjects is 100, then about one thousand pure noise images can be extracted to form the pure noise image set, and about four thousand images to be denoised can be extracted to form the denoised image set.

[0041] S4. Applying pure noise images to the pure noise image set, perform noise addition processing on each image to be denoised in the image set to be denoised, to obtain each initial noise-enhanced image corresponding to each image to be denoised. Then, using all the initial noise-enhanced images as model inputs and all the images to be denoised as model outputs, train the initial denoising model using a learning network model to obtain an initial denoising model. Then, import each image to be denoised into the initial denoising model to obtain each initial denoised image corresponding to each image to be denoised. Then, execute step S5.

[0042] In step S4, it is assumed that a clean image (i.e., a noise-free image) is represented as follows: The image to be denoised ,in, This represents noise; simultaneously, considering the pure noise images in the pure noise image set, they can be regarded as noise. Therefore, the initial noise-enhanced image can be... The sample image (i.e., the model input) and the image to be denoised The initial denoising learning network model is imported into the labeled image (i.e., the model output) to achieve the goal of learning from a noisy image using a noise-enhanced image and thus partially removing noise. Specifically, the pure noise images in the pure noise image set are used to add noise to each image in the image set to be denoised, resulting in initial noise-enhanced images corresponding to each image. This includes, but is not limited to: for each image in the image set to be denoised, first randomly selecting a pure noise image from the pure noise image set, and then adding the pure noise image to the corresponding image to be denoised using a layer overlay method to obtain the corresponding initial noise-enhanced image. That is, since the pure noise image set is full of noise, and the distribution statistics of the noise are highly consistent with the noise in the image set to be denoised, the two can be randomly combined to make the subsequent model training more robust. In addition, the initial denoising learning network model can be conventionally built using, but is not limited to, convolutional neural networks or U-Net-based network structures, and the specific process of model training is a conventional calibration and verification modeling method, which will not be elaborated here.

[0043] S5. Number of iterations Initialize to 1, and then obtain the result based on the initial denoised image. The first time The label image used in the next iteration, along with the pure noise images applied to the pure noise image set, is used to add noise to the initial denoised image, resulting in... The first time The next iteration uses sample images, and the initial denoising model is used as the basis for... The first time The next iteration of denoising uses a learned network model, and then step S6 is executed.

[0044] In step S5, the actual noisy image in this embodiment is considered to be the image to be denoised. Instead of the initial noise-enhanced image Therefore, the network in the initial denoising model faces a generalization problem, requiring iterative denoising through subsequent steps S5-S7, which continuously change the sample image and label image until the noise level of the input sample image is approximately [value missing]. .exist The first time mentioned The label image used in the next iteration can be directly the initial denoised image. However, to improve the denoising effect, this embodiment preferably adopts a target fusion strategy based on the initial denoised image and the image to be denoised (the purpose of which is to ensure that the label image contains information from the original noisy image each time, without distortion during the iteration process, thus preventing the image from becoming blurry). That is, the label image obtained in the next iteration is based on the initial denoised image. The first time The label image used in the next iteration includes, but is not limited to, the following: based on the initial denoised image, the label image is obtained according to the following formula. The first time The next iteration uses a label image. :

[0045] In the formula, This represents the preset weight coefficient corresponding to the first iteration, and is a pure decimal. This represents the image to be denoised, corresponding to the initial denoised image. This represents the functional form of the initial denoising model. This refers to the initial denoised image. For example, the preset weight coefficients... Furthermore, the aforementioned first The label image used in the next iteration refers to the image used for the first iteration. The denoised label image of the nth iteration, the nth The sample image used in the next iteration refers to the image used for the first iteration. The sample image for the denoising in the nth iteration, the nth The learning network model used for the second iteration of denoising refers to the model used for the third iteration. The learning network model for subsequent iterations of denoising, and the specific process of the noise addition can be found in the conventional derivation of step S4 above, and will not be repeated here.

[0046] S6. All the aforementioned first... The next iteration uses sample images as model input and all the aforementioned... The next iteration uses the labeled image as the model output term, for the first iteration... The next iteration of denoising uses the learned network model for further training, resulting in the [number]th iteration. The iteration denoising model is used, and each image to be denoised is imported into the first iteration denoising model. The iteration denoising model obtains each of the denoised images corresponding to each of the images to be denoised. After the denoising is performed in the next iteration, step S7 is executed.

[0047] In step S6, the specific process of model training is also the conventional calibration and verification modeling method, which will not be described in detail here. Furthermore, the loss function used for model training in each iteration can be expressed as follows:

[0048] In the formula, Denotes the loss function, Indicates the first Sample images used in the next iteration Indicates the first The subsequent iteration of denoising uses a function form based on the learned network model. Indicates that the first Sample images used in the next iteration Enter the first The next iteration of denoising uses the output image obtained after learning the network model. Indicates the first The next iteration uses a labeled image.

[0049] S7. Determine the number of iterations. Has the preset maximum number of iterations been reached? If so, proceed to step S8; otherwise, based on the previous step... The image after the denoising iteration is obtained as follows: The labeled image used in the next iteration, and the pure noise image applied to the pure noise image set for the first iteration. After the denoising process in the 1st iteration, the image is subjected to noise addition to obtain the 2nd iteration. The sample image used in the next iteration, and the first... The iteration denoising model is used as the first The next iteration of denoising uses a learned network model, and then the number of iterations is... Increment by 1 and return to step S6.

[0050] In step S7, theoretically, after multiple iterations based on steps S5 to S7, the model network loss will tend to stabilize and automatically converge. This is because the iterative denoising model continuously denoises the image to be denoised, and when the number of iterations is sufficient, a clean image can eventually be obtained. This makes subsequent iterations involve using images each time. Go to learn images The noise reduction effect should remain unchanged; in the experiment, the maximum number of iterations can be manually set according to different tasks. 23 In the task of denoising Na MRI images, setting the maximum number of iterations to 100 yields better results. If in The first time mentioned The labeled image used in the next iteration is directly the initial denoised image, then the... The label image used in the next iteration can also be directly the one mentioned above. The image after the denoising process in the second iteration; however, if a target fusion strategy based on the initial denoised image and the image to be denoised is adopted, then a fusion strategy based on the first iteration is also required here. The target fusion strategy for the denoised image and the image to be denoised in the denoising iteration, i.e., based on the denoised image of the denoising iteration... The image after the denoising iteration is obtained as follows: The label image used in the next iteration includes, but is not limited to, images based on the first... The image after the denoising iteration is obtained according to the following formula. The next iteration uses a label image. :

[0051] In the formula, Indicates the relationship with the first The weight coefficients corresponding to the next iteration. Indicates the first Functional form of the next iteration denoising model Indicates the first Image after the denoising process in the second iteration Represents a preset positive integer. This represents the function that takes the maximum value. Therefore, the weight coefficients will gradually decrease, eventually becoming zero. That is, in the later stages of iteration, the label is equivalent to the first... The image after the second iteration of denoising is actually an annealing strategy. Hyperparameters for the iterative process can also be manually set according to different tasks, for example in 23 In the task of denoising Na MRI images, if the maximum number of iterations has been set to 100, then Setting it to 90 results in better noise reduction. Furthermore, the specific process of the noise reduction treatment can be derived from the aforementioned step S4, and will not be repeated here.

[0052] S8. The first The image after each iteration of denoising is used as the final denoised image of the corresponding image to be denoised, and the three-dimensional image data of each subject is replaced and updated.

[0053] In step S8, for example, regarding the 25th layer of a two-dimensional image in the three-dimensional image data of a certain subject, if the corresponding layer has been obtained... After the denoising process of the second iteration, the 25th layer 2D image can be replaced with a certain denoised image. After the denoising is removed in the next iteration, the three-dimensional image data of the subject is replaced and updated.

[0054] Based on the above steps S1 to S8, this embodiment also performed the following tests: using medical... 23 Na MRI images were obtained, and the noise from the actual acquisitions was added to the simulation data to obtain, as shown below. Figure 2 The comparison chart of denoising effects shown can be seen that: compared with the tMPPCA method and the GL-HOSVD method, the denoising effect of the method described in this embodiment is better and retains better signal fidelity; and the residual plot in the third row can reflect that: the difference between the final denoised image and the label image is minimal.

[0055] Therefore, based on the medical MRI image denoising method described in steps S1 to S8 above, a novel medical MRI image denoising scheme with self-supervised learning characteristics is provided. First, medical MRI image signals from multiple subjects are received and converted to the image domain to obtain three-dimensional image data containing multiple layers of two-dimensional images. Then, all two-dimensional images of the bilateral slices are extracted from the three-dimensional image data as pure noise images, and all two-dimensional images not of the bilateral slices are extracted as images to be denoised. Next, the pure noise images are used to add noise to the images to be denoised, generating an initial noise-enhanced image. This image is used as input and the images to be denoised as output to train the initial denoising model, obtaining the initial denoised image. Finally, through an iterative process: generating label images based on the current denoised image, adding noise to the pure noise images to obtain sample images, training the iterative denoising model until a preset maximum number of iterations is reached, obtaining the final denoised image, and updating the three-dimensional image data. This approach eliminates the need for any assumptions about the noise pattern and cleverly combines 3D... The MRI imaging sequence fully utilizes the acquired noisy layer images to achieve self-supervised learning, resulting in significant denoising effects. Its denoising performance is superior to existing advanced methods such as tMPPCA and GL-HOSVD, making it convenient for practical application and promotion.

[0056] like Figure 4 As shown, the second aspect of this embodiment provides a virtual device for implementing the medical MRI image denoising method described in the first aspect, including an image signal receiving unit, an image data conversion unit, an image set aggregation unit, an initial image denoising unit, an iterative initialization unit, an iterative image denoising unit, a judgment and processing unit, and an image data updating unit. The image signal receiving unit is communicatively connected to the image data conversion unit and is used to receive medical MRI image signals from multiple subjects, and then activate the image data conversion unit. The image data conversion unit is communicatively connected to the image collection and aggregation unit. It is used to convert the corresponding medical MRI image signal to the image domain for each of the multiple subjects to obtain three-dimensional image data containing multiple two-dimensional images and corresponding images. Then, it extracts all two-dimensional images belonging to the two-sided slices from the three-dimensional image data as the corresponding pure noise images, and extracts all two-dimensional images not belonging to the two-sided slices from the three-dimensional image data as the corresponding images to be denoised. Finally, it starts the image collection and aggregation unit. The image set aggregation unit is communicatively connected to the image initial denoising unit. It is used to aggregate all pure noise images of each subject to obtain a pure noise image set, and to aggregate all images to be denoised of each subject to obtain a denoised image set, and then start the image initial denoising unit. The image initial denoising unit is communicatively connected to the iterative initialization unit. It is used to apply noise processing to each image to be denoised in the set of pure noise images, respectively, to obtain each initial noise enhancement image corresponding to each image to be denoised. Then, all the initial noise enhancement images are used as model inputs and all the images to be denoised are used as model outputs to train the initial denoising model using a learning network model, to obtain an initial denoising model. Each image to be denoised is then imported into the initial denoising model to obtain each initial denoised image corresponding to each image to be denoised. Then, the iterative initialization unit is started. The iterative initialization unit is communicatively connected to the image iterative denoising unit and is used to initialize the number of iterations. Initialize to 1, and then obtain the result based on the initial denoised image. The first time The label image used in the next iteration, along with the pure noise images applied to the pure noise image set, is used to add noise to the initial denoised image, resulting in... The first time The next iteration uses sample images, and the initial denoising model is used as the basis for... The first time The next iteration of denoising uses a learned network model, and then the image iterative denoising unit is started; The image iterative denoising unit is communicatively connected to the judgment and processing unit, and is used to process all the first... The next iteration uses sample images as model input and all the aforementioned... The next iteration uses the labeled image as the model output term, for the first iteration... The denoising iteration uses a learned network model for training, resulting in the... The iteration denoising model is used, and each image to be denoised is imported into the first iteration denoising model. The iteration denoising model obtains each of the denoised images corresponding to each of the images to be denoised. After the denoising is performed in the next iteration, the judgment and processing unit is then activated. The judgment and processing unit is communicatively connected to the image data update unit and the image iterative denoising unit, respectively, and is used to determine the number of iterations. If the preset maximum number of iterations has been reached, then the image data update unit is activated; otherwise, based on the first iteration... The image after the denoising iteration is obtained as follows: The labeled image used in the next iteration, and the pure noise image applied to the pure noise image set for the first iteration. After the denoising process in the 1st iteration, the image is subjected to noise addition to obtain the 2nd iteration. The sample image used in the next iteration, and the first... The iteration denoising model is used as the first The next iteration of denoising uses a learned network model, and then the number of iterations is... Increment by 1 and restart the image iterative denoising unit; The image data update unit is communicatively connected to the image data conversion unit, and is used to update the image data of the first image data. The image after each iteration of denoising is used as the final denoised image of the corresponding image to be denoised, and the three-dimensional image data of each subject is replaced and updated.

[0057] The working process, working details and technical effects of the aforementioned device provided in the second aspect of this embodiment can be found in the medical MRI image denoising method described in the first aspect, and will not be repeated here.

[0058] like Figure 5 As shown, the third aspect of this embodiment provides a physical system for implementing the medical MRI image denoising method described in the first aspect, including a magnetic resonance imaging instrument and a host computer that are connected in communication, wherein the magnetic resonance imaging instrument includes a scanning module and a magnetic resonance receiving coil. The scanning module is used to scan the subject by transmitting a sequence suitable for MRI imaging through a radio frequency transmitting coil, so as to excite the hydrogen atoms in the subject to produce a resonance phenomenon. The magnetic resonance receiving coil is used to receive and transmit medical MRI image signals to the host computer during the scanning of the subject. The host computer is used to execute the medical MRI image denoising method as described in the first aspect.

[0059] The working process, working details and technical effects of the aforementioned system provided in the third aspect of this embodiment can be found in the medical MRI image denoising method described in the first aspect, and will not be repeated here.

[0060] like Figure 6 As shown, the fourth aspect of this embodiment provides a computer device for performing the medical MRI image denoising method as described in the first aspect. The device includes a storage module, a processing module, and a transceiver module connected in sequence. The storage module stores a computer program, the transceiver module sends and receives messages, and the processing module reads the computer program and performs the medical MRI image denoising method as described in the first aspect. Specifically, the storage module may include, but is not limited to, random-access memory (RAM), read-only memory (ROM), flash memory, first-in-first-out (FIFO) memory, and / or first-in-last-out (FILO) memory, etc.; the processing module may, but is not limited to, use a microprocessor of the STM32F105 series. Furthermore, the computer device may also include, but is not limited to, a power supply module, a display screen, and other necessary components.

[0061] The working process, working details and technical effects of the aforementioned computer device provided in the fourth aspect of this embodiment can be found in the medical MRI image denoising method described in the first aspect, and will not be repeated here.

[0062] This fifth aspect of the embodiment provides a computer-readable storage medium storing instructions comprising the medical MRI image denoising method as described in the first aspect. Specifically, the computer-readable storage medium stores instructions that, when executed on a computer, perform the medical MRI image denoising method as described in the first aspect. The computer-readable storage medium refers to a data storage medium, and may include, but is not limited to, floppy disks, optical disks, hard disks, flash memory, USB flash drives, and / or Memory Sticks. The computer may be a general-purpose computer, a special-purpose computer, a computer network, or other programmable devices.

[0063] The working process, working details and technical effects of the aforementioned computer-readable storage medium provided in the fifth aspect of this embodiment can be found in the medical MRI image denoising method described in the first aspect, and will not be repeated here.

[0064] The sixth aspect of this embodiment provides a computer program product, including a computer program or instructions, which, when executed by a computer, implement the medical MRI image denoising method as described in the first aspect. The computer may be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device.

[0065] Finally, it should be noted that the above description is merely a preferred embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A self-supervised medical MRI image denoising method, characterized in that, This includes the following steps S1 to S8: S1. Receive medical MRI image signals from multiple subjects, and then execute step S2; S2. For each of the multiple subjects, the corresponding medical MRI image signal is converted to the image domain to obtain three-dimensional image data containing multiple two-dimensional images and corresponding images. Then, all two-dimensional images belonging to the two-sided slices are extracted from the three-dimensional image data as the corresponding pure noise images, and all two-dimensional images not belonging to the two-sided slices are extracted from the three-dimensional image data as the corresponding images to be denoised. Finally, step S3 is executed. S3. Summarize all the pure noise images of each subject to obtain a pure noise image set, and summarize all the images to be denoised of each subject to obtain a denoised image set, and then execute step S4; S4. Apply the pure noise images in the pure noise image set to each image to be denoised in the image set to be denoised, respectively, to obtain each initial noise-enhanced image corresponding to each image to be denoised. Then, use all the initial noise-enhanced images as model inputs and all the images to be denoised as model outputs to train the initial denoising model using a learning network model to obtain an initial denoising model. Then, import each image to be denoised into the initial denoising model to obtain each initial denoised image corresponding to each image to be denoised. Then, execute step S5. S5. Number of iterations Initialize to 1, and then obtain the result based on the initial denoised image. The first time The label image used in the next iteration, along with the pure noise images applied to the pure noise image set, is used to add noise to the initial denoised image, resulting in... The first time The next iteration uses sample images, and the initial denoising model is used as the basis for... The first time The next iteration of denoising uses a learned network model, and then step S6 is executed; S6. All the aforementioned first... The next iteration uses sample images as model input and all the aforementioned... The next iteration uses the labeled image as the model output term, for the first iteration... The denoising iteration uses a learned network model for training, resulting in the... The iteration denoising model is used, and each image to be denoised is imported into the first iteration denoising model. The iteration denoising model obtains each of the denoised images corresponding to each of the images to be denoised. After the denoising is removed in the next iteration, step S7 is then executed. S7. Determine the number of iterations. Has the preset maximum number of iterations been reached? If so, proceed to step S8; otherwise, based on the previous step... The image after the denoising iteration is obtained as follows: The labeled image used in the next iteration, and the pure noise image applied to the pure noise image set for the first iteration. After the denoising process in the 1st iteration, the image is subjected to noise addition to obtain the 2nd iteration. The sample image used in the next iteration, and the first... The iteration denoising model is used as the first The next iteration of denoising uses a learned network model, and then the number of iterations is... Increment by 1 and return to step S6; S8. The first The image after each iteration of denoising is used as the final denoised image of the corresponding image to be denoised, and the three-dimensional image data of each subject is replaced and updated.

2. The medical MRI image denoising method according to claim 1, characterized in that, When the medical MRI image signal is obtained by scanning the subject using a three-dimensional density-adapted radial sequence, for each of the plurality of subjects, the corresponding medical MRI image signal is converted to the image domain to obtain three-dimensional image data containing multiple layers of two-dimensional images, including: For each of the multiple subjects, a non-uniform fast Fourier transform is used to convert the corresponding medical MRI image signal to the image domain to obtain three-dimensional image data containing multiple two-dimensional images.

3. The medical MRI image denoising method according to claim 1, characterized in that, The pure noise images in the pure noise image set are applied to each image in the image set to be denoised, and noise enhancement processing is performed to obtain each initial noise-enhanced image corresponding to each image to be denoised, including: For each image to be denoised in the set of images to be denoised, a pure noise image is randomly selected from the set of pure noise images. Then, the pure noise image is added to the corresponding image to be denoised using a layer overlay method to obtain the corresponding initial noise enhancement image.

4. The medical MRI image denoising method according to claim 1, characterized in that, Based on the initial denoised image, the following is obtained: The first time The label images used in the next iteration include: Based on the initial denoised image, the result is obtained according to the following formula. The first time The next iteration uses a label image. : In the formula, This represents the preset weight coefficient corresponding to the first iteration, and is a pure decimal. This represents the image to be denoised, corresponding to the initial denoised image. This represents the functional form of the initial denoising model. This refers to the initial denoised image.

5. The medical MRI image denoising method according to claim 4, characterized in that, Based on the first The image after the denoising iteration is obtained as follows: The label images used in the next iteration include: Based on the first The image after the denoising iteration is obtained according to the following formula. The next iteration uses a label image. : In the formula, Indicates the relationship with the first The weight coefficients corresponding to the next iteration. Indicates the first Functional form of the next iteration denoising model Indicates the first Image after the denoising process in the second iteration Represents a preset positive integer. This represents the function that takes the maximum value.

6. A self-supervised medical MRI image denoising device, characterized in that, It includes an image signal receiving unit, an image data conversion unit, an image set aggregation unit, an initial image denoising unit, an iterative initialization unit, an iterative image denoising unit, a judgment and processing unit, and an image data update unit; The image signal receiving unit is communicatively connected to the image data conversion unit and is used to receive medical MRI image signals from multiple subjects, and then activate the image data conversion unit. The image data conversion unit is communicatively connected to the image collection and aggregation unit. It is used to convert the corresponding medical MRI image signal to the image domain for each of the multiple subjects to obtain three-dimensional image data containing multiple two-dimensional images and corresponding images. Then, it extracts all two-dimensional images belonging to the two-sided slices from the three-dimensional image data as the corresponding pure noise images, and extracts all two-dimensional images not belonging to the two-sided slices from the three-dimensional image data as the corresponding images to be denoised. Finally, it starts the image collection and aggregation unit. The image set aggregation unit is communicatively connected to the image initial denoising unit. It is used to aggregate all pure noise images of each subject to obtain a pure noise image set, and to aggregate all images to be denoised of each subject to obtain a denoised image set, and then start the image initial denoising unit. The image initial denoising unit is communicatively connected to the iterative initialization unit. It is used to apply noise processing to each image to be denoised in the set of pure noise images, respectively, to obtain each initial noise enhancement image corresponding to each image to be denoised. Then, all the initial noise enhancement images are used as model inputs and all the images to be denoised are used as model outputs to train the initial denoising model using a learning network model, to obtain an initial denoising model. Each image to be denoised is then imported into the initial denoising model to obtain each initial denoised image corresponding to each image to be denoised. Then, the iterative initialization unit is started. The iterative initialization unit is communicatively connected to the image iterative denoising unit and is used to initialize the number of iterations. Initialize to 1, and then obtain the result based on the initial denoised image. The first time The label image used in the next iteration, along with the pure noise images applied to the pure noise image set, is used to add noise to the initial denoised image, resulting in... The first time The next iteration uses sample images, and the initial denoising model is used as the basis for... The first time The next iteration of denoising uses a learned network model, and then the image iterative denoising unit is started; The image iterative denoising unit is communicatively connected to the judgment and processing unit, and is used to process all the first... The next iteration uses sample images as model input and all the aforementioned... The next iteration uses the labeled image as the model output term, for the first iteration... The denoising iteration uses a learned network model for training, resulting in the... The iteration denoising model is used, and each image to be denoised is imported into the first iteration denoising model. The iteration denoising model obtains each of the denoised images corresponding to each of the images to be denoised. After the denoising is performed in the next iteration, the judgment and processing unit is then activated. The judgment and processing unit is communicatively connected to the image data update unit and the image iterative denoising unit, respectively, and is used to determine the number of iterations. If the preset maximum number of iterations has been reached, then the image data update unit is activated; otherwise, based on the first iteration... The image after the denoising iteration is obtained as follows: The labeled image used in the next iteration, and the pure noise image applied to the pure noise image set for the first iteration. After the denoising process in the 1st iteration, the image is subjected to noise addition to obtain the 2nd iteration. The sample image used in the next iteration, and the first... The iteration denoising model is used as the first The next iteration of denoising uses a learned network model, and then the number of iterations is... Increment by 1 and restart the image iterative denoising unit; The image data update unit is communicatively connected to the image data conversion unit, and is used to update the image data of the first image data. The image after each iteration of denoising is used as the final denoised image of the corresponding image to be denoised, and the three-dimensional image data of each subject is replaced and updated.

7. A self-supervised medical MRI image denoising system, characterized in that, It includes a magnetic resonance imaging device and a host computer that are connected by communication, wherein the magnetic resonance imaging device includes a scanning module and a magnetic resonance receiving coil; The scanning module is used to scan the subject by transmitting a sequence suitable for MRI imaging through a radio frequency transmitting coil, so as to excite the hydrogen atoms in the subject to produce a resonance phenomenon. The magnetic resonance receiving coil is used to receive and transmit medical MRI image signals to the host computer during the scanning of the subject. The host computer is used to execute the medical MRI image denoising method as described in any one of claims 1 to 5.

8. A computer device, characterized in that, The device includes a storage module, a processing module, and a transceiver module that are sequentially connected in communication. The storage module is used to store a computer program, the transceiver module is used to send and receive messages, and the processing module is used to read the computer program and execute the medical MRI image denoising method as described in any one of claims 1 to 5.

9. A computer-readable storage medium, characterized in that... The computer-readable storage medium stores instructions that, when executed on a computer, perform the medical MRI image denoising method as described in any one of claims 1 to 5.

10. A computer program product, comprising a computer program or instructions, characterized in that, When the computer program or the instructions are executed by the computer, they implement the medical MRI image denoising method as described in any one of claims 1 to 5.

Images