Brain magnetic resonance image denoising method, system and device, medium and program product

Abstract

The invention provides a brain magnetic resonance image denoising method, system and device, a medium and a program product, and the method comprises the steps: obtaining a brain magnetic resonance image which comprises a plurality of stacked two-dimensional image slices with the same size, respectively carrying out denoising filtering processing and edge preserving filtering processing on each two-dimensional image slice based on a first filtering channel and a second filtering channel of a non-local mean value algorithm to obtain a corresponding first channel filtering image and a second channel filtering image; performing wavelet coefficient mixing processing on a first deviation correction image and a second deviation correction image obtained by performing Rice noise deviation correction based on the first channel filtering image and the second channel filtering image to obtain corresponding de-noised image slices; and obtaining a brain magnetic resonance de-noised image based on the stacking of the de-noised image slices corresponding to the two-dimensional image slices. According to the method, the brain imaging information can be fully reserved, the magnetic resonance image noise can be efficiently and accurately removed, and the image recovery quality and the recovery efficiency can be effectively improved.

Description

Technical Field

[0001] The present invention relates to the technical field of magnetic resonance image processing, and in particular to a method, system, device, medium and program product for denoising brain magnetic resonance images. Background Art

[0002] Magnetic Resonance Imaging (MRI) is widely used in the imaging of complex structures such as the brain and spine due to its high soft tissue contrast and non-invasive characteristics. Since the noise in the imaging process will seriously affect the clarity of the image, and thus affect the accuracy of clinical diagnosis, denoising has become a key step in brain MRI image processing.

[0003] Currently, the conventional denoising techniques for magnetic resonance images are to process the mean or weighted average of local pixels of the image by methods such as mean filters, median filters, and Gaussian filters to reduce noise. However, these traditional filters only rely on local information of pixels, which easily leads to blurred image edges. While it is difficult to effectively retain fine structural features, it is also difficult to effectively denoise Rician noise, which is more obvious in the low-intensity amplitude region, relatively weak in the high-intensity region, has a positive deviation in the low-signal region, resulting in an over-amplified noise level, and exhibits complex characteristics such as non-additivity. Although the existing standard Non-local Means (NLM) algorithm is superior to traditional filter processing in retaining edges and details due to its global search strategy, its practical application in denoising brain magnetic resonance images also has certain limitations: 1) Searching for similar blocks within the entire image range leads to high computational complexity and time cost, making it difficult to meet the real-time processing requirements. 2) In a high-noise scenario, the similarity weighting method of the NLM algorithm is easily affected by noise, resulting in the weighted mean deviating from the true value. That is, when applied to high-noise brain imaging scenarios, it is difficult to efficiently and reliably remove Rician noise in magnetic resonance images while maintaining the integrity of the imaging structure. Therefore, there is an urgent need to provide an efficient and reliable denoising method that can not only improve the denoising effect of brain MRI images, but also reduce the computational complexity and enhance the robustness to Rician noise. Summary of the Invention

[0004] In order to solve the problems of the prior art, it is necessary to provide a method, system, device, medium and program product for denoising brain magnetic resonance images for the above technical problems.

[0005] In a first aspect, an embodiment of the present invention provides a method for denoising brain magnetic resonance images, the method comprising the following steps:

[0006] Obtain a brain magnetic resonance image; the brain magnetic resonance image includes a stack of multiple two-dimensional image slices of the same size;

[0007] The first filtering channel and the second filtering channel based on the non-local mean algorithm respectively perform filtering processing on each two-dimensional image slice to obtain a first-channel filtered image and a second-channel filtered image corresponding to each two-dimensional image slice. Among them, the first filtering channel is used for image denoising; the second filtering channel is used for image edge preservation;

[0008] Based on the Rice noise deviation correction respectively performed on the first-channel filtered image and the second-channel filtered image corresponding to each two-dimensional image slice, a first deviation correction image and a second deviation correction image corresponding to each two-dimensional image slice are obtained;

[0009] Based on the wavelet coefficient mixing processing performed on the first deviation correction image and the second deviation correction image corresponding to each two-dimensional image slice, a denoised image slice corresponding to each two-dimensional image slice is obtained;

[0010] Based on the stacking of the denoised image slices corresponding to each two-dimensional image slice, a brain magnetic resonance denoised image is obtained.

[0011] In a second aspect, an embodiment of the present invention provides a brain magnetic resonance image denoising system, and the system includes:

[0012] An image acquisition module for acquiring a brain magnetic resonance image; the brain magnetic resonance image includes a stack of multiple two-dimensional image slices of the same size;

[0013] A filtering processing module for respectively performing filtering processing on each two-dimensional image slice based on a first filtering channel and a second filtering channel introducing the non-local mean algorithm to obtain a first-channel filtered image and a second-channel filtered image corresponding to each two-dimensional image slice. Among them, the first filtering channel is used for image denoising; the second filtering channel is used for image edge preservation;

[0014] A deviation correction module for respectively performing Rice noise deviation correction on the first-channel filtered image and the second-channel filtered image corresponding to each two-dimensional image slice to obtain a first deviation correction image and a second deviation correction image corresponding to each two-dimensional image slice;

[0015] A wavelet mixing module for performing wavelet coefficient mixing processing on the first deviation correction image and the second deviation correction image corresponding to each two-dimensional image slice to obtain a denoised image slice corresponding to each two-dimensional image slice;

[0016] A denoised image generation module for obtaining a brain magnetic resonance denoised image based on the stacking of the denoised image slices corresponding to each two-dimensional image slice.

[0017] In a third aspect, an embodiment of the present invention further provides a computer device, including a processor and a memory. The processor is connected to the memory. The memory is used to store a computer program, and the processor is used to execute the computer program stored in the memory so that the computer device executes the steps of the above method.

[0018] In a fourth aspect, an embodiment of the present invention further provides a computer-readable storage medium. A computer program is stored in the computer-readable storage medium. When the computer program is run, the steps of the above method are implemented.

[0019] In a fifth aspect, an embodiment of the present invention further provides a computer program product, including a program executed by at least one processor of a computing device. The program includes instructions, and the execution of the instructions enables the at least one processor to execute the steps of the above method.

[0020] Compared with the prior art, the present invention not only shortens the algorithm traversal time and reduces the computational complexity through the non-local mean algorithm that dynamically adjusts the search step of the matching pixel block based on the neighborhood search window region type designed according to the brain structure characteristics, effectively improving the denoising execution efficiency, but also effectively improves the denoising performance by introducing Rician bias correction to perform high-efficiency correction on the filtered image. At the same time, by combining the wavelet coefficient mixing technology, the useful information processed by the denoising filter channel and the edge-preserving filter channel is fully utilized, greatly retaining the anatomical structure and image detail information of the image, effectively improving the image restoration quality, and thus providing higher-quality image support for clinical brain diagnosis. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 is a schematic flowchart of the brain magnetic resonance image denoising method in an embodiment of the present invention;

[0022] Figure 2 is a schematic flowchart of searching for matching pixel blocks for a two-dimensional image slice based on the non-local mean algorithm in an embodiment of the present invention;

[0023] Figure 3 is a schematic diagram of the application effect of the search step dynamically adjusted according to the window region type of the neighborhood search window in an embodiment of the present invention;

[0024] Figure 4 is a schematic flowchart of wavelet coefficient mixing processing in an embodiment of the present invention;

[0025] Figure 5 is another schematic flowchart of the brain magnetic resonance image denoising method in an embodiment of the present invention;

[0026] Figure 6It is a schematic diagram of the verification result of the denoising processing effect of the magnetic resonance image of the present invention in the embodiment of the present invention;

[0027] Figure 7 It is a schematic structural diagram of a brain magnetic resonance image denoising device in the embodiment of the present invention. Specific embodiments

[0028] In order to make the purpose, technical solutions and beneficial effects of the present application clearer and more understandable, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Obviously, the following described embodiments are part of the embodiments of the present invention and are only used to illustrate the present invention, but not to limit the scope of the present invention. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative work fall within the scope of protection of the present invention.

[0029] In some embodiments, as Figure 1 shown, a method for denoising brain magnetic resonance images is provided, and the method includes the following steps:

[0030] S11. Obtain a brain magnetic resonance image; wherein, the brain magnetic resonance image can be understood as image format data obtained by reconstructing human body data based on a brain magnetic resonance instrument in actual applications, usually in the dcm format. The brain magnetic resonance image is in the form of a three-dimensional matrix, and its corresponding dimensions are S*H*W, where S represents the number of two-dimensional image slices, and H and W respectively represent the height and width of the two-dimensional image slices; that is, the brain magnetic resonance image includes a stack of multiple two-dimensional image slices of the same size.

[0031] S12. Perform filtering processing on each two-dimensional image slice based on the first filtering channel and the second filtering channel of the non-local mean algorithm to obtain the first-channel filtered image and the second-channel filtered image corresponding to each two-dimensional image slice. Among them, the first filtering channel is used for image denoising to quickly calculate and obtain the frame information of each two-dimensional image slice (mainly focusing on retaining the integrity of the basic contour and structure of the image); the second filtering channel is used for image edge retention to retain more detail information of the image; that is, the first-channel filtered image obtained by performing filtering processing on each two-dimensional image slice by the first filtering channel is an image mainly containing the frame information of the two-dimensional image slice; the second-channel filtered image obtained by performing filtering processing on each two-dimensional image slice by the second filtering channel is an image retaining more edge information of the two-dimensional image slice.

[0032] The search process of the non-local mean algorithm for each two-dimensional image slice in the brain magnetic resonance image is the same; as Figure 2 shown, specifically, the search process for each two-dimensional image slice is to first determine the first neighborhood search window M in the two-dimensional image slice, and calculate the reference pixel block N within the neighborhood search windowp with the first matching pixel block N q similarity weight value; after the calculation is completed, the matching pixel block N q is slid to the next non-overlapping new pixel block N q′ and its similarity weight with the reference pixel block N p is calculated; after completing the first neighborhood search window M, it is then slid to the next neighborhood search window M' at a certain step size for pixel block matching calculation within this neighborhood search window until the entire image is traversed and calculated, and the similarity weights of all pixels in the two-dimensional image slice can be obtained. Based on the weighted sum of all similarity weights, the corresponding denoised pixel value is obtained. Among them, the specific calculation method and process within each neighborhood search window M are as follows: within the selected neighborhood search window M, first determine the reference pixel block N p and its central pixel point p, and then search within the neighborhood search window for a matching pixel block N p with similar texture to the reference pixel block N q , and based on the pixel values of the central pixel points q of each matching pixel block N q and the weighted calculation of the similarity weight between them and the reference pixel block N p , the denoised pixel value of the central pixel point p of the reference pixel block N p is obtained.

[0033] By performing filtering processing on each two-dimensional image slice through the above non-local mean algorithm, the corresponding channel filtered image can be obtained. In practical applications, different algorithm parameters are selected for the first filtering channel and the second filtering channel constructed based on the non-local mean algorithm due to different requirements for the filtering functions to be realized, and specific limitations are not provided here. Although the above standard non-local mean algorithm can be used for denoising processing of brain magnetic resonance images, the search method that needs to find similar pixel blocks corresponding to the pixel points to be denoised within the entire image range results in a relatively high computational complexity of the entire algorithm, and it is difficult to meet the real-time processing requirements due to the large time cost. In order to reduce the computational complexity as much as possible while denoising, improve the efficiency of the algorithm operation and the real-time performance of brain magnetic resonance image processing, in some embodiments, considering that the image content of brain magnetic resonance images has characteristics such as clear stratification of brain structures (white matter / gray matter / cerebrospinal fluid, etc.), symmetry of brain structures, and distinct levels of brain structures (the outer right island includes scalp, skull, meninges, brain tissue, ventricles, etc.), making full use of these image characteristics to skip the calculation of some pixel points can shorten the traversal time of the algorithm and improve the algorithm efficiency. Therefore, a non-local mean algorithm for dynamically adjusting the search step size of matching pixel blocks based on the region type of the neighborhood search window is specifically designed for brain magnetic resonance image processing to effectively improve the denoising processing efficiency of brain magnetic resonance images; correspondingly, the steps of performing filtering processing on each two-dimensional image slice by the first filtering channel and the second filtering channel based on the non-local mean algorithm to obtain the first channel filtered image and the second channel filtered image corresponding to each two-dimensional image slice include:

[0034] Based on the identification of the window region types of each neighborhood search window on the two-dimensional image slice, a first region and a second region are obtained. Among them, the first region and the second region can be respectively understood as a smooth region and an edge region designed based on the above brain structure characteristics combined with the pixel change situation within the region as the classification basis. When the average pixel gradient value of the neighborhood search window is greater than or equal to the global gradient threshold, the window region type of the neighborhood search window is obtained as the first region; when the average pixel gradient value of the neighborhood search window is less than the global gradient threshold, the window region type of the neighborhood search window is obtained as the second region. In the embodiments of the present invention, the gradient threshold used for window region type identification can be set according to requirements in principle. In order to ensure the accuracy of region type identification and thus ensure the reliability of the improved non-local mean algorithm, in some embodiments, the global gradient threshold required is determined based on the pixel gradient intensity matrix of the entire two-dimensional image slice; correspondingly, the steps of performing window region type identification on each neighborhood search window on the two-dimensional image slice include:

[0035] Performing convolution calculation on the pixel grayscale value matrix of the neighborhood search window based on the Sobel operator to obtain the pixel gradient intensity matrix corresponding to the neighborhood search window; wherein, the process of obtaining the pixel gradient intensity matrix of the neighborhood search window can be understood as first performing convolution calculation on the pixel grayscale value matrix of the neighborhood search window by the Sobel operator to obtain the horizontal direction gradient and vertical direction gradient of each pixel point, and then based on the horizontal direction gradient and vertical direction gradient of each pixel point, using the following pixel gradient calculation formula to obtain the gradient intensity value of the corresponding pixel point, and further obtaining the pixel gradient intensity matrix of the neighborhood search window:

[0036]

[0037] wherein, (i,j) represents the position coordinates of the pixel point in the neighborhood search window M; s represents the size of the neighborhood search window M. G x (i,j) and G y (i,j) respectively represent the horizontal direction gradient and vertical direction gradient of the pixel point (i,j), which are obtained by performing convolution calculation on the pixel grayscale value matrix of the neighborhood search window by the Sobel operator. The specific process of calculating the horizontal direction gradient and vertical direction gradient of each pixel point based on the Sobel operator can be implemented with reference to relevant existing technologies and will not be elaborated here.

[0038] Performing element average calculation on the pixel gradient intensity matrix of the neighborhood search window to obtain the average pixel gradient value of the neighborhood search window; that is, the average pixel gradient value is the average of the gradient intensity values of all pixel points within the corresponding neighborhood search window.

[0039] Obtaining the global gradient threshold based on the median of the gradient intensities in the pixel gradient intensity matrix of the two-dimensional image slice; wherein, the pixel gradient intensity matrix of the two-dimensional image slice can be understood as the pixel gradient intensity matrix of the entire image obtained by performing convolution calculation on the pixel grayscale value matrix of the two-dimensional image slice based on the Sobel operator. The specific calculation process can be referred to the relevant description of the pixel gradient intensity matrix of the neighborhood search window above; the corresponding median of the gradient intensities can be understood as the median of the element values in the pixel gradient intensity matrix.

[0040] Obtain the corresponding step size control parameter based on the window area type of the neighborhood search window; the step size control parameter is used to control the calculation efficiency and denoising accuracy by adjusting the search step size within the neighborhood search window. In practical applications, the larger the search step size, the more pixels are skipped, and the higher the calculation efficiency, but it may lead to a decrease in the denoising effect. On the contrary, the smaller the search step size, the higher the sampling density, the increase in the amount of calculation, but the better the denoising accuracy. Therefore, it is necessary to select an appropriate step size control parameter based on the specific window area type of the neighborhood search window to adaptively adjust the search step size. Among them, the window area type of the neighborhood search window may be the first area (smooth area) or the second area (edge area). The gradient intensity change in the first area is small, and there are many similar pixel points in the area. A larger step size control parameter value can be used to skip more similarity weight calculations to accelerate the calculation. While the gradient intensity change in the second area is large. To ensure the integrity of the image edges and details as much as possible, a smaller step size control parameter value can be used to skip relatively fewer similarity weight calculations to accelerate the calculation. In some embodiments, the step of obtaining the step size control parameter based on the window area type of the neighborhood search window includes:

[0041] In the first filtering channel, when the window area type of the neighborhood search window is the first area, the value range of the step size control parameter a1 is 30 to 50, and when the window area type of the neighborhood search window is the second area, the value range of the step size control parameter a1 is 5 to 10.

[0042] In the second filtering channel, when the window area type of the neighborhood search window is the first area, the value range of the step size control parameter a2 is 30 to 50, and when the window area type of the neighborhood search window is the second area, the value range of the step size control parameter a2 is 5 to 10, where a2 is less than a1.

[0043] It should be noted that in the embodiments of the present invention, the step control parameters of each neighborhood search window are generally determined based on the specific window area type first, and then fine-tuned within the parameter value range corresponding to the window area type according to the actual application requirements of the used filtering channels. In practical applications, the value-taking strategy of the step control parameters of different window area types within the corresponding value range is determined according to the resolution of clinical brain magnetic resonance images. When the image resolution of the input image I is higher, the step control parameter of the first area can be set closer to the upper limit of the value range of 30-50, and the step control parameter of the second area can be set closer to the lower limit of the value range of 5-10. The specific fine-tuning principle for different filtering channels is that the step control parameter of the filtering channel mainly used for denoising and focusing on retaining the image frame information is greater than the step control parameter of the filtering channel mainly used for edge retention and focusing on retaining more image detail information under the same window area type; that is, the step control parameter of the first filtering channel is greater than the step control parameter of the second filtering channel under the same window area type.

[0044] Based on the first filtering channel and the second filtering channel, filtering processing is performed on each two-dimensional image slice respectively by using a search step determined based on the average pixel gradient value and the step control parameter, to obtain a first-channel filtered image and a second-channel filtered image corresponding to each two-dimensional image slice; in order to ensure the reliability of the adaptive adjustment of the search step, in some embodiments, the step of determining the search step based on the average pixel gradient value and the step control parameter includes:

[0045] Based on rounding down the ratio of the step control parameter to the average pixel gradient value of the neighborhood search window, a candidate search step is obtained; that is, the candidate search step is expressed as floor(α / G), where α represents the step control parameter; G represents the average pixel gradient value of the neighborhood search window; floor(·) represents rounding down.

[0046] Based on the maximum value between the candidate search step and a preset minimum unit search step, the search step is obtained; that is, the search step is expressed as:

[0047] L = max(γ, floor(α / G))

[0048] where max(·) represents taking the maximum value; γ represents the preset minimum unit search step, which can be determined according to actual application requirements. In order to ensure a necessary acceleration effect, in some embodiments, the preset minimum unit search step is set to 1; L represents the search step. The larger α is, the larger L is, the more pixels are skipped, the higher the calculation efficiency is, but the denoising effect decreases. The smaller α is, the smaller L is, the higher the sampling density is, the calculation amount increases, but the denoising accuracy is better.

[0049] In order to perform efficient and comprehensive filtering processing on each two-dimensional image slice, in some embodiments, multiple neighborhood search windows are divided within each two-dimensional image slice to perform corresponding denoising estimation on each pixel point, that is, the two-dimensional image slice includes several neighborhood search windows; correspondingly, the steps of the first filtering channel and the second filtering channel based on the non-local mean algorithm to perform filtering processing on each two-dimensional image slice to obtain the first-channel filtered image and the second-channel filtered image corresponding to each two-dimensional image slice further include:

[0050] Based on the reference pixel block and the search step size in the neighborhood search window, several matching pixel blocks are obtained; among them, the reference pixel block can be understood as a pixel block centered on the pixel point to be denoised; the corresponding matching pixel block can be understood as a pixel block of the same size within the neighborhood search window that has texture similarity with the reference pixel block, and the specific number varies depending on the actual processed image and the search step size. The positional relationship between adjacent matching pixel blocks is as Figure 3 shown, that is, the distance between the central pixel points q and q' of adjacent matching pixel blocks is L.

[0051] Based on the Euclidean distance between the pixel value matrix of the reference pixel block and the pixel value matrix of the matching pixel block, the similarity weight value between the central pixel point of the reference pixel block and the central pixel point of the matching pixel block is obtained; among them, the pixel value matrix can be understood as a matrix composed of the positions and gray values of all pixel points within the corresponding pixel block. Suppose I(p) represents the gray value of the central pixel point p of N p center, I(q) represents the gray value of the central pixel point q of N q center, I(N p ) and I(N q ) respectively represent the pixel value matrices corresponding to the reference pixel block N p and the matching pixel block N q , then the similarity weight value can be expressed as:

[0052]

[0053] Among them, w(p,q) is the similarity weight value of pixel point q relative to pixel point p; ||I(N p ) - I(N q )|| 2 represents the Euclidean distance between the reference pixel block N p and the matching pixel block N q ; h represents the denoising degree control parameter, which is a smoothing parameter used to control the weighting degree; Z(p) is the normalization factor to ensure that the value range of the weight factor is in the interval (0, 1); M represents the neighborhood search window corresponding to the reference pixel block N p .

[0054] Based on the similarity weight value between the central pixel point of the reference pixel block and the central pixel point of the matching pixel block, and the gray value of the central pixel point of the matching pixel block, the denoised pixel value of the central pixel point of the reference pixel block is obtained; that is, the denoised pixel value of the central pixel point is expressed as:

[0055]

[0056] In the formula, I(q) represents the gray value of the central pixel point q of the matching pixel block N q ; O(p) represents the denoised pixel value of the central pixel point p of the reference pixel block N p Based on this calculation formula, in the process of screening relevant pixels during the denoising process of the central pixel point of the neighborhood search window, the similarity weights of some pixels skipped by adjusting the search step are calculated. It can also be understood that the corresponding similarity weights are directly set to 0 and not included in the final weighted summation.

[0057] Based on the denoised pixel values of the central pixel points of the reference pixel blocks in a number of neighborhood search windows, the first-channel filtered image and the second-channel filtered image corresponding to each two-dimensional image slice are obtained; wherein, the matching pixel blocks do not overlap, and the matching pixel blocks and the reference pixel blocks are of equal size.

[0058] In the above embodiment, the pixel preselection method based on the pixel gradient intensity of the neighborhood search window for region type identification and dynamically adjusting the search step according to the corresponding region type identification result to skip the weight value calculation of some non-effective pixels can not only shorten the algorithm traversal time and effectively improve the algorithm calculation efficiency, but also effectively suppress irrelevant pixels while retaining the effective pixel values during the filtering process, thereby improving the image signal-to-noise ratio and the denoising quality of the channel filtered image.

[0059] In the embodiments of the present invention, both the first filtering channel and the second filtering channel are filtering channels designed based on the improved non-local mean algorithm given in the above embodiments. The corresponding filtering effect is directly affected by the relevant parameters in the non-local mean algorithm. The key control parameter items of the first filtering channel and the second filtering channel are the same, and the difference between the two is only the different control parameter values selected based on the difference in the function of the filtering channel. In some embodiments, the first filtering channel performs similarity weight value calculation and denoising processing on each two-dimensional image slice based on a first set of control parameters; wherein, the first set of control parameters includes the neighborhood search window size M1 in the first filtering channel, the size PS1 of the matching pixel block and the reference pixel block in the first filtering channel, and the denoising degree control parameter h1 in the first filtering channel. The second filtering channel performs similarity weight value calculation and denoising processing on each two-dimensional image slice based on a second set of control parameters; wherein, the second set of control parameters includes the neighborhood search window size M2 in the second filtering channel, the size PS2 of the matching pixel block and the reference pixel block in the second filtering channel, and the denoising degree control parameter h2 in the second filtering channel.

[0060] In order to ensure that the similarity weights of the respective filtering channels are calculated by adjusting the control parameter combinations of the two filtering channels, so as to mainly improve the denoising effect of the magnetic resonance image using the first filtering channel and mainly retain the edge details of the image using the second filtering channel, it is necessary to design a differential combination of the first set of control parameters and the second set of control parameters:

[0061] 1) Considering that the neighborhood search window size is used to balance the calculation range and calculation accuracy of the similarity weight, for the denoising effect of the first filtering channel, the neighborhood search window size M1 can be set to a larger value, such as 27*27, for image smoothing; for the edge detail retention effect of the first filtering channel, the window size M2 can be set to a smaller value, such as 15*15, for capturing more texture and edge information. In some embodiments, it is necessary to satisfy that M1 is greater than M2.

[0062] 2) Considering that for the denoising effect of the first filtering channel, the size PS1 of the matching pixel block and the reference pixel block can be set to a larger value, such as 5*5, for image smoothing; while for the edge detail retention effect of the second filtering channel, the size PS2 of the matching pixel block and the reference pixel block can be set to a smaller value, such as 3*3, for capturing more texture and edge information. In some embodiments, it is necessary to satisfy that PS1 is greater than PS2.

[0063] 3) The denoising degree control parameter is a key parameter in the above formula for calculating the pixel similarity weight. This parameter is used to control the denoising degree. Considering its characteristic of being mainly related to the mean square error of the imaging noise of the magnetic resonance device, in some embodiments, the denoising degree control parameters h1 and h2 of the first filtering channel and the second filtering channel are respectively obtained based on the products of the control factors k1 and k2 and the mean square error of the reference imaging noise of the device, and the value ranges of k1 and k2 are both 1.5 to 2.5. The difference between the two is that the first filtering channel mainly focuses on denoising, and the corresponding parameter takes a larger value, that is, k1 > k2.

[0064] The mean square error of the reference imaging noise of the device can be understood as the mean square error of the noise that may be introduced during the generation of magnetic resonance images due to the limitations of the performance of the magnetic resonance device itself; in order to ensure the reliability of the setting of the denoising degree control parameter, in some embodiments, the steps for obtaining the mean square error of the reference imaging noise of the device include:

[0065] Based on multiple interval scans of a standard water film by the magnetic resonance device, multiple water film images are obtained; among them, the standard water film can be understood as a special water film that can be used as a scanning object of the magnetic resonance device for imaging noise evaluation of the magnetic resonance device; the corresponding magnetic resonance device is also the device for obtaining brain magnetic resonance images. The number of times of scanning the standard water film by this device can be determined according to the actual application scenario, as long as it meets the requirement of obtaining the mean square error of the reference imaging noise of the device through subsequent statistical analysis.

[0066] Based on the statistical analysis of the difference between the pixel values within the region and the mean pixel value within the region of each water film image for a preset region of interest, the mean square error of the reference imaging noise of the device is obtained; among them, the preset region of interest can be understood as an image region determined based on actual needs and applicable to imaging noise evaluation; the corresponding mean square error of the reference imaging noise of the device can be calculated based on the pixel values within the region and the mean pixel value within the region of the preset region of interest on each water film image according to the mean square error calculation formula of statistics, which will not be elaborated here.

[0067] S13. Based on the Rician noise deviation correction performed on the first-channel filtered image and the second-channel filtered image corresponding to each two-dimensional image slice respectively, the first deviation correction image and the second deviation correction image corresponding to each two-dimensional image slice are obtained; among them, the first deviation correction image and the second deviation correction image can be understood as images obtained by specifically performing Rician noise deviation correction on the first-channel filtered image and the second-channel filtered image obtained by filtering in order to improve the denoising effect as much as possible, considering the inconsistent noise levels of the Rician noise on the magnetic resonance image in the high and low amplitude regions, resulting in the inability to remove the noise using conventional noise suppression methods.

[0068] In some embodiments, in order to improve the denoising performance as much as possible, the steps of performing bias correction on the channel-filtered images of each two-dimensional image slice to obtain a bias-corrected image include:

[0069] Performing bias correction processing on each pixel value in the channel-filtered image based on the mean square error of the device imaging reference noise to obtain the corresponding Rician noise bias correction value; wherein, the process of obtaining the mean square error of the device imaging reference noise is described above; that is, the Rician noise bias correction value corresponding to each pixel value of the channel-filtered image is expressed as:

[0070]

[0071] wherein, I(p) represents the pixel amplitude of pixel point p in the two-dimensional image slice, NLM(I(p)) represents the pixel amplitude of pixel point p in the channel-filtered image, that is, the output result after the two-dimensional image slice is processed by non-local means filtering; σ represents the mean square error of the device imaging reference noise; I′(p) represents the Rician noise bias correction value of pixel point p output after Rician bias correction.

[0072] S14. Obtaining the denoised image slice corresponding to each two-dimensional image slice based on the wavelet coefficient mixing process performed on the first bias-corrected image and the second bias-corrected image corresponding to each two-dimensional image slice; wherein, the wavelet coefficient mixing process can be understood as a process of decomposing the first bias-corrected image and the second bias-corrected image corresponding to two filtering channels into different frequency bands based on wavelet transform in order to extract useful information from the images filtered from the two-dimensional image slice based on different filtering control parameters, so as to retain the image anatomical structure to a great extent and improve the image restoration quality, and then extracting the useful information of the corresponding wavelet subbands, mixing the low-frequency approximation components of the two filtering channels to obtain the fused low-frequency approximation component, and then combining it with other high-frequency components in the denoising filtering channel (the first filtering channel) to perform inverse wavelet transform to obtain the corresponding denoised image slice.

[0073] In some embodiments, as Figure 4 shown, the steps of obtaining the denoised image slice corresponding to each two-dimensional image slice based on the wavelet coefficient mixing process performed on the first bias-corrected image and the second bias-corrected image corresponding to each two-dimensional image slice include:

[0074] Based on the multi-level discrete wavelet transform performed on the first bias-corrected images corresponding to each two-dimensional image slice, a first low-frequency wavelet coefficient matrix and a first high-frequency wavelet coefficient matrix are obtained; wherein, the number of levels of the multi-level discrete wavelet transform (Discrete Wavelet Transform, DWT) is usually selected from positive integers in the range of 3 to 5; considering that too high a level of classification will increase the computational cost and introduce boundary artifacts, in some embodiments, for high-resolution magnetic resonance images, the number of levels of the multi-level discrete wavelet transform is set to 4, corresponding to the multi-level discrete wavelet transform to obtain multiple frequency band components as shown in Table 1, and the components of each frequency band are in the form of two-dimensional matrices, that is, the required first low-frequency wavelet coefficient matrix (low-frequency approximation component) and the first high-frequency wavelet coefficient matrix (high-frequency component) are obtained, and the first high-frequency wavelet coefficient matrix includes a horizontal edge wavelet coefficient matrix (horizontal edge component), a vertical edge wavelet coefficient matrix (transverse edge component), and a diagonal edge wavelet coefficient matrix (45° diagonal component).

[0075] Table 1 Frequency band components obtained by 4-level discrete wavelet transform

[0076] Input Approximate component Horizontal edge component Lateral edge component 45° diagonal component First deviation correction image ω(a,k1) ω(v,k1) ω(h,k1) ω(d,k1) Second deviation correction image ω(a,k2) ω(v,k2) ω(h,k2) ω(d,k2)

[0077] Based on the multi-level discrete wavelet transform performed on the second bias-corrected images corresponding to each two-dimensional image slice, a second low-frequency wavelet coefficient matrix and a second high-frequency wavelet coefficient matrix are obtained; wherein, for the relevant descriptions of the second low-frequency wavelet coefficient matrix and the second high-frequency wavelet coefficient matrix, refer to the relevant descriptions of the first low-frequency wavelet coefficient matrix and the first high-frequency wavelet coefficient matrix above, and will not be repeated here.

[0078] Based on the first high-frequency wavelet coefficient matrix, the weighted fusion of the first low-frequency wavelet coefficient matrix and the second low-frequency wavelet coefficient matrix corresponding to each two-dimensional image slice is performed to obtain a fused low-frequency wavelet coefficient matrix; wherein, weighted fusion can be understood as a process of mixing the useful information in the first bias-corrected image and the second bias-corrected image in proportion in order to make an effective balance between edge detail retention and denoising.

[0079] Considering that the low-frequency components obtained by performing multi-level wavelet decomposition on an image can retain structural information such as the boundary positions of brain gray matter, white matter, and cerebrospinal fluid in the brain image, and the high-frequency components can retain detailed features such as brain sulci, gyri, lesion edges, and texture details in the brain image, the noise amplitude of the image after bias correction can be used as a weight factor to fuse wavelet high- and low-frequency coefficients, so as to ensure the rationality of the fusion weight setting in weighted fusion, and further obtain a better mixing effect. In some embodiments, the step of obtaining a fused low-frequency wavelet coefficient matrix by weighted fusion of the first low-frequency wavelet coefficient matrix and the second low-frequency wavelet coefficient matrix corresponding to each two-dimensional image slice based on the first high-frequency wavelet coefficient matrix includes: obtaining the noise mean square error of the first bias-corrected image based on the first high-frequency wavelet coefficient matrix; to ensure the reliability of the calculation of the noise mean square error of the first bias-corrected image, in some embodiments, the step of obtaining the noise mean square error of the first bias-corrected image based on the first high-frequency wavelet coefficient matrix includes:

[0080] Based on the root mean square of the sum of the squares of the high-frequency component wavelet coefficient matrices in the first high-frequency wavelet coefficient matrix corresponding to each two-dimensional image slice, obtain the noise mean square error matrix of the first bias-corrected image corresponding to each two-dimensional image slice; that is, the noise mean square error matrix is expressed as where ω(v,k1), ω(h,k1), and ω(d,k1) respectively represent the horizontal edge wavelet coefficient matrix, vertical edge wavelet coefficient matrix, and diagonal edge wavelet coefficient matrix in the first high-frequency wavelet coefficient matrix.

[0081] Based on the ratio of the element mean of the noise mean square error matrix of the first bias-corrected image corresponding to each two-dimensional image slice to the standard deviation mapping factor, obtain the noise mean square error of the first bias-corrected image corresponding to each two-dimensional image slice; where the standard deviation mapping factor can be understood as a proportional factor that maps the median of the noise mean square error matrix to the standard deviation. In some embodiments, the standard deviation mapping factor is 0.6745; that is, the noise mean square error of the first bias-corrected image corresponding to the two-dimensional image slice can be expressed as:

[0082]

[0083] In the formula, mean(·) represents obtaining the element mean of the noise mean square error matrix, and the corresponding statistical method is to sum each matrix element and then divide by the total number of matrix elements; σ′ represents the noise mean square error of the first bias-corrected image.

[0084] Based on the mean square error of noise of the first deviation-corrected image, a fusion weight is obtained; wherein, the fusion weight is the weight coefficient of the first low-frequency wavelet coefficient matrix set considering that when the mean square error of noise of the first deviation-corrected image is large, the noise amplitude is large, and the low-frequency components need to be penalized while the high-frequency details are enhanced. On the contrary, the denoising and smoothing effects need to be strengthened to ensure that the image after wavelet mixing is more natural and realistic. In some embodiments, the fusion weight is set as:

[0085]

[0086] In the formula, β1 and β2 respectively represent the weight coefficients of the first low-frequency wavelet coefficient matrix and the second low-frequency wavelet coefficient matrix; σ′ represents the mean square error of noise of the first deviation-corrected image.

[0087] Based on the weighted fusion of the first low-frequency wavelet coefficient matrix and the second low-frequency wavelet coefficient matrix using the fusion weight, the fused low-frequency wavelet coefficient matrix is obtained; wherein, the fused low-frequency wavelet coefficient matrix is expressed as:

[0088] ω(a,o) = β1·ω(a,k1) + β2·ω(a,k2)

[0089] In the formula, ω(a,k1) and ω(a,k2) respectively represent the first low-frequency wavelet coefficient matrix and the second low-frequency wavelet coefficient matrix; β1 and β2 represent the weight coefficients corresponding to the first low-frequency wavelet coefficient matrix and the second low-frequency wavelet coefficient matrix in the fusion weight; ω(a,o) represents the fused low-frequency wavelet coefficient matrix. Based on its expression, it is easy to know that when using the noise amplitude of the first deviation-corrected image as the weight factor for wavelet coefficient mixing, when the noise amplitude is large, the weight coefficient of the first low-frequency wavelet coefficient matrix can be made smaller to penalize the low-frequency components and enhance the high-frequency details, ensuring that after image denoising, the edges and details of the brain image are highlighted and the situation of over-smoothing and detail loss is avoided; when the noise amplitude is small, the weight coefficient of the first low-frequency wavelet coefficient matrix can be made larger to further strengthen the denoising and smoothing effects, better suppress the noise, and ensure that the denoised brain magnetic resonance image is more natural and realistic.

[0090] Based on the multi-level inverse discrete wavelet transform performed on the fused low-frequency wavelet coefficient matrix and the first high-frequency wavelet coefficient matrix corresponding to each two-dimensional image slice, the denoised image slice corresponding to each two-dimensional image slice is obtained; among them, the multi-level inverse discrete wavelet transform can be understood as the inverse processing process of the multi-level discrete wavelet transform, and the specific processing process can refer to relevant existing technologies; in practical applications, in order to ensure the denoising effect of the image, after obtaining the fused low-frequency wavelet coefficient matrix, it is used as the low-frequency wavelet coefficient matrix input during the multi-level inverse discrete wavelet transform process, and at the same time, each high-frequency component wavelet coefficient matrix in the first high-frequency wavelet coefficient matrix is used as the corresponding high-frequency component to perform the inverse wavelet transform process to reconstruct the image, and the required denoised image slice is obtained.

[0091] S15. Based on the stacking of the denoised image slices corresponding to each two-dimensional image slice, a brain magnetic resonance denoised image is obtained; among them, the brain magnetic resonance denoised image can be understood as the dcm format data obtained by sequentially stacking the corresponding denoised image slices according to the stacking order of each two-dimensional image slice in the originally acquired brain magnetic resonance image.

[0092] As provided in the embodiment of the present invention Figure 5 shown in the method for obtaining a brain magnetic resonance image including stacking a plurality of two-dimensional image slices of the same size, filtering each two-dimensional image slice respectively by a first filtering channel for image denoising and a second filtering channel for image edge preservation designed based on an improved non-local means algorithm, obtaining the first-channel filtered image and the second-channel filtered image corresponding to each two-dimensional image slice, then obtaining the first deviation-corrected image and the second deviation-corrected image corresponding to each two-dimensional image slice based on the Rician noise deviation correction performed on the first-channel filtered image and the second-channel filtered image corresponding to each two-dimensional image slice respectively, and obtaining the denoised image slice corresponding to each two-dimensional image slice based on the wavelet coefficient mixing process performed on the first deviation-corrected image and the second deviation-corrected image corresponding to each two-dimensional image slice, and based on the stacking of the denoised image slices corresponding to each two-dimensional image slice, obtaining the brain magnetic resonance denoised image. The technical solution not only shortens the algorithm traversal time and reduces the computational complexity through the non-local means algorithm based on dynamically adjusting the matching pixel block search step according to the neighborhood search window region type designed for the brain structure characteristics, effectively improving the denoising execution efficiency, but also effectively improves the denoising performance by introducing Rician deviation correction to efficiently correct the filtered image. At the same time, by combining the wavelet coefficient mixing technology, it fully utilizes the useful information processed by the denoising filtering channel and the edge preservation filtering channel, greatly preserves the image anatomical structure and image detail information, effectively improves the image restoration quality, and thus provides higher-quality image support for clinical brain diagnosis.

[0093] To verify the effectiveness of the brain magnetic resonance image denoising method provided by the embodiments of the present invention, when the root mean square of the noise of the magnetic resonance imaging device is measured to be 7.5 or 22 respectively, under three different brain magnetic resonance image (T1 / T2 / PD) scenarios, the computational efficiency of the traditional NLM algorithm and the improved NLM algorithm proposed by the present invention (i.e., the average output time of the algorithm under the same input image) is compared and analyzed, and the computational efficiency comparison results shown in Table 2 are obtained. As can be easily seen from Table 2, the computational efficiency of the improved NLM algorithm proposed by the present invention is improved compared with the traditional algorithm under both root mean square of noise conditions, verifying the high efficiency of the denoising process of the method of the present invention.

[0094] Table 2 Comparison table of computational efficiency between the traditional NLM algorithm and the improved NLM algorithm of the present invention

[0095]

[0096]

[0097] Meanwhile, based on the comparison and analysis method of the processing results of the traditional NLM algorithm without Rician bias correction and the magnetic resonance image processing effect of the present invention, the denoising performance of the denoising method of the present invention and the improvement effect of the image restoration quality are verified, and Figure 6 the comparative diagram of the denoising effect of the brain T1 image shown in Figure 6 From left to right in the magnetic resonance brain T1 image, the images are the image without noise, the image with noise (7.5), the denoising result of the traditional NLM algorithm, and the denoising result of the method of the present invention. Among them, the inputs of both the traditional NLM algorithm and the method of the present invention are the image with noise (7.5). As can be seen from Figure 6 the image results shown, the denoising result of the method of the present invention is closer to the image without noise. Its overall output image effect is smoother than that of the traditional NLM method. The noise is greatly suppressed, and the details are more completely retained, and the image is also clearer; that is, it further verifies that the denoising effect of the method of the present invention is superior to the original image and the processing results of the traditional NLM method in terms of detail retention and image smoothness.

[0098] In some embodiments, as Figure 7 shown, a brain magnetic resonance image denoising system is provided, and the brain magnetic resonance image denoising system includes:

[0099] An image acquisition module 1 for acquiring brain magnetic resonance images; the brain magnetic resonance images include a stack of two-dimensional image slices of the same size;

[0100] A filtering processing module 2, configured to perform filtering processing on each two-dimensional image slice based on a first filtering channel and a second filtering channel introducing a non-local mean algorithm, so as to obtain a first-channel filtered image and a second-channel filtered image corresponding to each two-dimensional image slice, wherein the first filtering channel is used for image denoising; and the second filtering channel is used for image edge preservation.

[0101] A deviation correction module 3, configured to perform Rician noise deviation correction on the first-channel filtered image and the second-channel filtered image corresponding to each two-dimensional image slice respectively, so as to obtain a first deviation correction image and a second deviation correction image corresponding to each two-dimensional image slice.

[0102] A wavelet mixing module 4, configured to perform wavelet coefficient mixing processing on the first deviation correction image and the second deviation correction image corresponding to each two-dimensional image slice, so as to obtain a denoised image slice corresponding to each two-dimensional image slice.

[0103] A denoised image generation module 5, configured to stack the denoised image slices corresponding to each two-dimensional image slice to obtain a brain magnetic resonance denoised image.

[0104] It should be noted that the specific limitations on the brain magnetic resonance image denoising system can be referred to the limitations on the brain magnetic resonance image denoising method in the above text, and the corresponding technical effects can also be equivalently obtained; in addition, each module in the above brain magnetic resonance image denoising system can be implemented in whole or in part by software, hardware, and their combination. Each of the above modules can be embedded in or independent of a processor in a computer device in the form of hardware, or stored in a memory in the computer device in the form of software, so that the processor can call and execute the operations corresponding to each of the above modules.

[0105] In some embodiments, a magnetic resonance imaging device is provided, characterized in that the magnetic resonance imaging device includes an imaging controller; the imaging controller includes a processor and a memory, the processor is coupled to the memory, the memory is used to store a computer program or instruction, and the processor is used to execute the computer program or instruction in the memory, so that the imaging controller executes the steps of the above method.

[0106] In some embodiments, a computer-readable storage medium is provided, on which a computer program is stored, and when the computer program is executed by a processor, the steps of the above method are implemented.

[0107] In summary, compared with the prior art, the brain magnetic resonance image denoising method, system, device, medium and program product provided by the embodiments of the present invention not only shortens the algorithm traversal time and reduces the computational complexity through the non-local mean algorithm that dynamically adjusts the search step of matching pixel blocks based on the neighborhood search window region type designed for brain structure characteristics, effectively improving the denoising execution efficiency, but also effectively improves the denoising performance by introducing Rician bias correction to perform efficient correction on the filtered image. At the same time, by combining the wavelet coefficient mixing technology, the useful information processed by the denoising filter channel and the edge-preserving filter channel is fully utilized, the anatomical structure and image detail information of the image are retained to a great extent, the image restoration quality is effectively improved, and thus higher-quality image support is provided for clinical brain diagnosis.

[0108] In the above embodiments, it can be implemented in whole or in part by software, hardware, firmware or any combination thereof. When implemented using software, it can be implemented in whole or in part in the form of a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, the processes or functions according to some embodiments of the present invention are generated in whole or in part. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable devices. The computer instructions can be stored in a computer-readable storage medium, or transmitted from one computer-readable storage medium to another, for example, the computer instructions can be transmitted from one website, computer, server or data center to another website, computer, server or data center by wire (such as coaxial cable, optical fiber, digital subscriber line) or wireless (such as infrared, wireless, microwave, etc.). The computer-readable storage medium can be any available medium that can be accessed by a computer or a data storage device such as a server or data center that includes one or more available media integrated. The available medium can be a magnetic medium (for example, a floppy disk, a hard disk, a magnetic tape), an optical medium (for example, a DVD), or a semiconductor medium (for example, an SSD), etc.

[0109] Those skilled in the art can understand that all or part of the processes of implementing the methods in the above embodiments can be completed by instructing relevant hardware through a computer program. The computer program can be stored in a computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above methods.

[0110] Each embodiment in this specification is described in a progressive manner. For the parts that are the same or similar in each embodiment, reference can be made to each other. Each embodiment focuses on the differences from other embodiments. In particular, for the device embodiments, since they are basically similar to the method embodiments, the description is relatively simple. For the relevant parts, reference can be made to the corresponding descriptions in the method embodiments. It should be noted that the technical features of the above embodiments can be combined arbitrarily. For the sake of brevity of description, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, it should be considered as falling within the scope described in this specification.

[0111] The above-described embodiments merely represent several preferred embodiments of the present application. The description is relatively specific and detailed, but it should not be construed as a limitation on the scope of the invention patent. It should be pointed out that for those of ordinary skill in the art in this technical field, without departing from the technical principle of the present invention, several improvements and substitutions can be made, and these improvements and substitutions should also be regarded as within the protection scope of the present application. Therefore, the protection scope of the patent of the present application shall be subject to the protection scope of the claims.

Claims

1. A method for denoising brain magnetic resonance images, characterized in that, The method includes the following steps: Obtain a brain magnetic resonance image; the brain magnetic resonance image includes a stack of multiple two-dimensional image slices of the same size; Perform filtering processing on each two-dimensional image slice respectively through a first filtering channel and a second filtering channel based on the non-local mean algorithm, to obtain a first-channel filtered image and a second-channel filtered image corresponding to each two-dimensional image slice, where the first filtering channel is used for image denoising; the second filtering channel is used for image edge preservation; Based on performing Rician noise bias correction on the first-channel filtered image and the second-channel filtered image corresponding to each two-dimensional image slice respectively, obtain a first bias-corrected image and a second bias-corrected image corresponding to each two-dimensional image slice; Based on performing wavelet coefficient mixing processing on the first bias-corrected image and the second bias-corrected image corresponding to each two-dimensional image slice, obtain a denoised image slice corresponding to each two-dimensional image slice; Based on the stacking of the denoised image slices corresponding to each two-dimensional image slice, obtain a brain magnetic resonance denoised image.

2. The brain magnetic resonance image denoising method according to claim 1, characterized in that, The step of performing filtering processing on each two-dimensional image slice respectively through the first filtering channel and the second filtering channel based on the non-local mean algorithm to obtain a first-channel filtered image and a second-channel filtered image corresponding to each two-dimensional image slice includes: Based on performing window region type identification on each neighborhood search window on the two-dimensional image slice, obtain a first region and a second region, where when the average pixel gradient value of the neighborhood search window is greater than or equal to the global gradient threshold, obtain that the window region type of the neighborhood search window is the first region; when the average pixel gradient value of the neighborhood search window is less than the global gradient threshold, obtain that the window region type of the neighborhood search window is the second region; Based on the window region type of the neighborhood search window, obtain a corresponding step size control parameter; the step size control parameter is used to control the calculation efficiency and denoising accuracy; Based on the first filtering channel and the second filtering channel, use the search step size determined based on the average pixel gradient value and the step size control parameter to perform filtering processing on each two-dimensional image slice respectively, to obtain a first-channel filtered image and a second-channel filtered image corresponding to each two-dimensional image slice.

3. The brain magnetic resonance image denoising method according to claim 2, wherein, The step of performing window region type identification on each neighborhood search window on the two-dimensional image slice includes: Based on performing convolution calculation on the pixel gray value matrix of the neighborhood search window through a Sobel operator, obtain a pixel gradient intensity matrix corresponding to the neighborhood search window; Based on performing element average calculation on the pixel gradient intensity matrix of the neighborhood search window, obtain the average pixel gradient value of the neighborhood search window; Based on the gradient intensity median in the pixel gradient intensity matrix of the two-dimensional image slice, obtain the global gradient threshold.

4. The brain magnetic resonance image denoising method according to claim 2, wherein The step of determining the search step size based on the average pixel gradient value and the step size control parameter includes: Based on performing floor operation on the ratio of the step size control parameter to the average pixel gradient value of the neighborhood search window, obtain a candidate search step size; The search step is obtained based on the maximum value between the candidate search step and the preset minimum unit search step.

5. The brain magnetic resonance image denoising method according to claim 2, characterized in that The step of obtaining the step control parameter based on the window area type of the neighborhood search window includes: In the first filtering channel, when the window area type of the neighborhood search window is the first area, the value range of the step control parameter a1 is 30 to 50, and when the window area type of the neighborhood search window is the second area, the value range of the step control parameter a1 is 5 to 10; In the second filtering channel, when the window area type of the neighborhood search window is the first area, the value range of the step control parameter a2 is 30 to 50, and when the window area type of the neighborhood search window is the second area, the value range of the step control parameter a2 is 5 to 10, where a2 is less than a1.

6. The brain magnetic resonance image denoising method according to claim 2, wherein The two-dimensional image slice includes a plurality of neighborhood search windows; The steps of respectively performing filtering processing on each two-dimensional image slice by the first filtering channel and the second filtering channel based on the non-local mean algorithm to obtain the first-channel filtered image and the second-channel filtered image corresponding to each two-dimensional image slice further include: Based on the reference pixel block and the search step in the neighborhood search window, a plurality of matching pixel blocks are obtained; Based on the Euclidean distance between the pixel value matrix of the reference pixel block and the pixel value matrix of the matching pixel block, the similarity weight value between the central pixel point of the reference pixel block and the central pixel point of the matching pixel block is obtained; Based on the similarity weight value between the central pixel point of the reference pixel block and the central pixel point of the matching pixel block and the gray value of the central pixel point of the matching pixel block, the denoised pixel value of the central pixel point of the reference pixel block is obtained; Based on the denoised pixel values of the central pixel points of the reference pixel blocks of a plurality of neighborhood search windows, the first-channel filtered image and the second-channel filtered image corresponding to each two-dimensional image slice are obtained; wherein, the matching pixel blocks do not overlap, and the matching pixel blocks and the reference pixel blocks are of equal size.

7. The brain magnetic resonance image denoising method according to claim 6, wherein The first filtering channel performs similarity weight value calculation and denoising processing on each two-dimensional image slice based on the first set of control parameters; wherein, the first set of control parameters includes the neighborhood search window size M1 in the first filtering channel, the size PS1 of the matching pixel block and the reference pixel block in the first filtering channel, and the denoising degree control parameter h1 in the first filtering channel; The second filtering channel performs similarity weight value calculation and denoising processing on each two-dimensional image slice based on the second set of control parameters; wherein, the second set of control parameters includes the neighborhood search window size M2 in the second filtering channel, the size PS2 of the matching pixel block and the reference pixel block in the second filtering channel, and the denoising degree control parameter h2 in the second filtering channel.

8. The brain magnetic resonance image denoising method according to claim 7, characterized in that M1 is greater than M2.

9. The brain magnetic resonance image denoising method according to claim 7, characterized in that PS1 is greater than PS2.

10. The brain magnetic resonance image denoising method according to claim 7, wherein, The denoising degree control parameters h1 and h2 are respectively obtained based on the products of the control factors k1 and k2 and the mean square error of the device imaging reference noise, where the value ranges of k1 and k2 are 1.5 to 2.5, and k1 > k2.

11. The brain magnetic resonance image denoising method according to claim 10, wherein The steps for obtaining the mean square error of the device imaging reference noise include: Based on multiple interval scans of a standard water film by a magnetic resonance device, multiple water film images are obtained; Based on the statistical analysis of the difference between the pixel values within the region and the mean pixel value within the region for each water film image in a preset region of interest, the mean square error of the device imaging reference noise is obtained.

12. The brain magnetic resonance image denoising method according to claim 1, wherein, The step of obtaining the denoised image slice corresponding to each two-dimensional image slice based on the wavelet coefficient hybrid processing of the first bias correction image and the second bias correction image corresponding to each two-dimensional image slice includes: Based on the multi-level discrete wavelet transform of the first bias correction image corresponding to each two-dimensional image slice, a first low-frequency wavelet coefficient matrix and a first high-frequency wavelet coefficient matrix are obtained; Based on the multi-level discrete wavelet transform of the second bias correction image corresponding to each two-dimensional image slice, a second low-frequency wavelet coefficient matrix and a second high-frequency wavelet coefficient matrix are obtained; Based on the first high-frequency wavelet coefficient matrix, the weighted fusion of the first low-frequency wavelet coefficient matrix and the second low-frequency wavelet coefficient matrix corresponding to each two-dimensional image slice is performed to obtain a fused low-frequency wavelet coefficient matrix; Based on the multi-level inverse discrete wavelet transform of the fused low-frequency wavelet coefficient matrix and the first high-frequency wavelet coefficient matrix corresponding to each two-dimensional image slice, the denoised image slice corresponding to each two-dimensional image slice is obtained.

13. The brain magnetic resonance image denoising method according to claim 12, characterized in that, The number of levels of the multi-level discrete wavelet transform is 4; the first high-frequency wavelet coefficient matrix includes a horizontal edge wavelet coefficient matrix, a vertical edge wavelet coefficient matrix, and a diagonal edge wavelet coefficient matrix.

14. The brain magnetic resonance image denoising method according to claim 13, characterized in that, The step of obtaining the fused low-frequency wavelet coefficient matrix by performing the weighted fusion of the first low-frequency wavelet coefficient matrix and the second low-frequency wavelet coefficient matrix corresponding to each two-dimensional image slice based on the first high-frequency wavelet coefficient matrix includes: Based on the first high-frequency wavelet coefficient matrix, the mean square error of the noise of the first bias correction image is obtained; Based on the mean square error of the noise of the first bias correction image, a fusion weight is obtained; Based on the fusion weight, the weighted fusion of the first low-frequency wavelet coefficient matrix and the second low-frequency wavelet coefficient matrix is performed to obtain the fused low-frequency wavelet coefficient matrix.

15. The brain magnetic resonance image denoising method according to claim 14, wherein The step of obtaining the mean square error of the noise of the first bias correction image based on the first high-frequency wavelet coefficient matrix includes: Based on the root mean square of the sum of the squares of the wavelet coefficient matrices of each high-frequency component in the first high-frequency wavelet coefficient matrix corresponding to each two-dimensional image slice, a noise mean square error matrix of the first bias correction image corresponding to each two-dimensional image slice is obtained; Based on the ratio of the element mean value of the noise mean square error matrix of the first bias correction image corresponding to each two-dimensional image slice to the standard deviation mapping factor, the mean square error of the noise of the first bias correction image corresponding to each two-dimensional image slice is obtained; the standard deviation mapping factor is 0.6745.

16. A brain magnetic resonance image denoising system, characterized in that, The system includes: An image acquisition module for acquiring brain magnetic resonance images; the brain magnetic resonance images include a stack of multiple two-dimensional image slices of the same size; A filtering processing module for performing filtering processing on each two-dimensional image slice based on a first filtering channel and a second filtering channel introducing a non-local mean algorithm to obtain a first-channel filtered image and a second-channel filtered image corresponding to each two-dimensional image slice, wherein the first filtering channel is used for image denoising; the second filtering channel is used for image edge preservation; A bias correction module for performing Rician noise bias correction on the first-channel filtered image and the second-channel filtered image corresponding to each two-dimensional image slice respectively to obtain a first bias correction image and a second bias correction image corresponding to each two-dimensional image slice; A wavelet mixing module for performing wavelet coefficient mixing processing on the first bias correction image and the second bias correction image corresponding to each two-dimensional image slice to obtain a denoised image slice corresponding to each two-dimensional image slice; A denoised image generation module for obtaining a brain magnetic resonance denoised image based on the stacking of the denoised image slices corresponding to each two-dimensional image slice.

17. A computer device, characterized in that: Comprising a processor and a memory, the processor is connected to the memory, the memory is used for storing a computer program, and the processor is used for executing the computer program stored in the memory so that the computer device executes the method according to any one of claims 1 to 15.

18. A computer-readable storage medium, characterized in that: A computer program is stored in the computer-readable storage medium, and when the computer program is run, it executes the method according to any one of claims 1 to 15.

19. A computer program product, characterized in that: Comprising a program executed by at least one processor of a computing device, the program includes instructions, and the execution of the instructions enables the at least one processor to execute the method according to any one of claims 1 to 15.

Images