Magnetic resonance signal noise suppression and image definition enhancement method

Abstract

The invention discloses a magnetic resonance signal noise suppression and image definition enhancement method, and relates to the technical field of medical image processing. According to the magnetic resonance signal noise suppression and image definition enhancement method, a preprocessed magnetic resonance image is extracted and divided into a plurality of local areas to analyze and determine the noise level of each local area, and noise suppression operation is performed on the local areas with noise influence through multi-stage nonlinear filtering processing. And carrying out image reconstruction on the signal subjected to noise suppression to obtain a preliminary magnetic resonance image, adopting a multi-stage nonlinear filtering strategy, removing impulse noise by using adaptive median filtering based on local variance, and then further smoothing the image by adopting improved bilateral filtering and keeping edge information. The contradiction between noise suppression and edge preservation of a traditional filtering method is effectively avoided, and details and edge information of the image are better reserved.

Description

Technical Field

[0001] This invention relates to the field of medical image processing technology, specifically to a method for suppressing magnetic resonance signal noise and enhancing image sharpness. Background Technology

[0002] Magnetic resonance imaging (MRI) is an important medical imaging diagnostic technique with advantages such as no radiation and high soft tissue contrast. It is widely used for examination of various parts of the human body. However, during the MRI process, it is inevitably affected by various noises, such as thermal noise, motion noise, and radio frequency interference.

[0003] The reference patent title is: A Magnetic Resonance Image Enhancement Method Based on Multi-Task Collaborative Network (Patent Publication No.: CN113096028A, Patent Publication Date: 2021-07-09), which includes the following steps: Step 1: Construct a multi-task enhancement model, coordinating contrast enhancement, edge enhancement, and signal-to-noise ratio enhancement modules, and introducing clumping structures in each module; Step 2: Train the multi-task enhancement model; Step 3: Input the magnetic resonance image to be enhanced into the trained multi-task enhancement model to obtain the enhanced magnetic resonance image; The multi-task image enhancement method realizes functional collaboration between multiple modules and efficient information feedback within modules, effectively improving the enhancement effect of magnetic resonance images in terms of contrast, detail clarity, and signal-to-noise ratio.

[0004] Based on the description in the above document, existing methods for suppressing magnetic resonance signal noise and enhancing image sharpness have certain limitations. They are not accurate in evaluating noise in local areas of image data, and traditional filtering methods, although combining grayscale and pixel distance for denoising, will lose image detail information, resulting in blurred image edges. Therefore, this invention provides a method for suppressing magnetic resonance signal noise and enhancing image sharpness. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a method for suppressing magnetic resonance signal noise and enhancing image sharpness. This method solves the limitations of existing methods for suppressing magnetic resonance signal noise and enhancing image sharpness, such as inaccurate noise assessment of local areas in image data, and the fact that traditional filtering methods, while combining grayscale and pixel distance for noise reduction, lose image detail information and cause blurred image edges.

[0006] To achieve the above objectives, the present invention provides a method for suppressing magnetic resonance signal noise and enhancing image sharpness, specifically comprising the following steps:

[0007] A1. Use magnetic resonance imaging equipment to collect raw magnetic resonance signal data of the human body to be examined, and extract the data after consultation with the cooperating hospital to achieve optimized training operation;

[0008] A2. Implement preprocessing operations on the acquired data, extract the preprocessed magnetic resonance image and divide it into multiple local regions for analysis and determination of the noise level of each local region, perform noise suppression operations on the local regions affected by noise through multi-stage nonlinear filtering, and perform image reconstruction on the noise-suppressed signal to obtain a preliminary magnetic resonance image.

[0009] A3. Post-process the reconstructed magnetic resonance images to optimize the visual effects and highlight the edges and details of the images. Output the post-processed high-definition magnetic resonance images to the display interface for image display or to the storage terminal for storage.

[0010] Preferably, the preprocessing operation for the collected data in A2 is as follows:

[0011] B1. Extract the acquired raw magnetic resonance signal data and classify the data according to the set nodes;

[0012] B2. Convert the format of the data under the corresponding node, convert the signal data into image data suitable for subsequent processing, and associate the parameter data corresponding to the image data with the image data;

[0013] B3. Remove abnormal data after conversion and use normalization method to normalize the signal strength to the range of [0, 1] to eliminate the dimensional differences between different data.

[0014] Preferably, the data classification operation in B1 according to the set nodes is as follows:

[0015] b11. The nodes are set according to the data categories, forming multi-level classification nodes to distinguish the data;

[0016] b12. The detected part category is used as the primary classification node, and the image and numerical categories are used as secondary classification nodes based on the primary classification node. The data acquisition timestamp is used as the final classification node based on the secondary classification node.

[0017] b13. Match data according to the category of the category node. That is, if there are characters in the collected data that are the same as the category node, the corresponding data will be extracted and introduced into the category node. The first-level category node, the second-level category node and the last-level category node will establish edge relationships. The last-level category node, the second-level category node and the first-level category node will establish edge relationships in sequence, and the edge direction will be from the last-level category node to the second-level category node, and the second-level category node will point to the first-level category node.

[0018] Preferably, the operation in A2 of extracting and dividing the preprocessed magnetic resonance image into multiple local regions for analysis and determination of the noise level of each local region is as follows:

[0019] C1. Extract magnetic resonance images from the preprocessed data, divide the magnetic resonance images into multiple local regions at equal intervals according to a set distance, and set the pixels in the local regions sequentially from top to bottom and from left to right.

[0020] C2. Calculate and compare the mean and variance of each local region with the mean and variance of the current global magnetic resonance image, and dynamically determine the value of the adaptive weight coefficient α based on the comparison and the set adjustment rules.

[0021] C3. Then, based on the value of the adaptive weighting coefficient α, the noise estimation values ​​in the local area are calculated, and the noise levels of each area are determined by sorting the noise estimation values.

[0022] Preferably, the operation in C2 that dynamically determines the value of the adaptive weight coefficient α based on the comparison situation and according to the set adjustment rules is as follows:

[0023] c21. Based on the deviation between local and global images of magnetic resonance images in historical data, set the threshold interval [m, n] and the corresponding weight coefficient α under the threshold change;

[0024] c22. When the absolute value of the variance L between the extracted local region image and the current global magnetic resonance image is compared with the set threshold interval [m, n];

[0025] If L < m, the similarity between the local region image and the current global magnetic resonance image is low, while the corresponding weight coefficient α is the largest.

[0026] If L > n, the similarity between the local region image and the current global magnetic resonance image is high, and the corresponding weight coefficient α is the smallest.

[0027] If L∈[m, n], the similarity between the local region image and the current global magnetic resonance image changes continuously. As L increases, the corresponding weight coefficient α decreases accordingly, and the corresponding weight coefficient α is obtained based on the corresponding threshold interval [m, n].

[0028] c23. Based on the real-time extracted local region image, extract the adaptively changing weight coefficient α and complete the noise assessment of the local region.

[0029] Preferably, the operation of calculating and comparing the mean and variance of the local region in C2 with the mean and variance of the current global magnetic resonance image is as follows:

[0030] Calculate the local mean value and label it J. h J h J represents the mean of the h-th local region. hk=(K1+K2+…+K k ) / k,K k The value of the k-th pixel in the h-th local region is used as an example. Similarly, the mean value of the global magnetic resonance image is JH.

[0031] Calculate the local variance scale as X h X h Let represent the variance of the h-th local region, and the calculation formula is: Similarly, the mean value of the global magnetic resonance image is XH;

[0032] The calculated absolute difference between the mean and absolute values ​​of the current local region and the global magnetic resonance image is |JH-J h |, while the absolute difference of variance between the current local region and the global magnetic resonance image |XH-X h |, that is, when the absolute difference of the mean conforms to the normal fluctuation range, the absolute difference of the variance is determined and the corresponding adaptive weight coefficient α is matched;

[0033] The formula for calculating the noise estimate in the local area based on the adaptive weighting coefficient α in C3 is: Y = α × X h +(1-α)×XH.

[0034] Preferably, the noise suppression operation in A2, which involves multi-stage nonlinear filtering to address noise in localized areas, is as follows:

[0035] D1. Extract the local areas affected by noise and establish a neighborhood window centered on each pixel in the image;

[0036] D2. Set the estimated noise value used for comparison to S. Compare the difference between the median value and the center pixel value in the current window with the estimated noise value. If the difference exceeds the estimated noise value by r times, expand the window uniformly in all directions and recalculate the median value until the difference meets the condition of being less than the noise value by r times or expands to the set maximum window size.

[0037] The median value is the middle value of the sorted pixel values ​​in the current window. If the number of pixels is odd, it is the value of the middle pixel. If the number of pixels is odd, it is the average of the two middle pixel values. The center pixel value is the value of the pixel located at the center of the current window.

[0038] D3. After completing adaptive median filtering, the image is smoothed, and bilateral filtering is used to optimize image details and preserve edges.

[0039] Preferably, the bilateral filtering in D3 performs the optimization of image details and edge preservation as follows:

[0040] d31. Based on adaptive median filtering, extract a pixel from the image and establish a square window centered on the current pixel;

[0041] d32. The spatial distance weight is wz, which is determined by the distance from any pixel in the square window to the center pixel. The grayscale similarity weight is wt, which is determined by the difference in grayscale value between any pixel and the center pixel. Multiple rays are drawn from the center pixel to the boundary of the square window, with an angle of 45° between adjacent rays. The values ​​of adjacent pixels on the extended rays are extracted and the gradient magnitude is determined to obtain the local gradient adjustment weight wu.

[0042] d33. Combine the spatial distance weight wz, gray-level similarity weight wt, and local gradient adjustment weight wu to adjust the gray value of the current center pixel.

[0043] Preferably, the operation of adjusting the gray value of the current center pixel by combining the spatial distance weight wz, the gray-level similarity weight wt, and the local gradient adjustment weight wu is as follows:

[0044] I1=I0(i,j)×wz(i,j)×wt(i,j)×wu(i,j);

[0045] Where I1 represents the adjusted grayscale value of the center pixel of the current square window, I0(i,j) represents the initial grayscale value of pixel (i,j), wz(i,j) represents the spatial distance weight of pixel (i,j), wt(i,j) represents the grayscale similarity weight of pixel (i,j), wu(i,j) represents the local gradient adjustment weight of pixel (i,j), and wz+wt+wu=1;

[0046] Furthermore, the closer a pixel is to the center pixel, the greater its influence on the center pixel, and the larger its spatial distance weight value. Pixels with similar gray values ​​to the center pixel have a larger gray similarity weight value. In edge regions with larger gradients, the local gradient adjustment weight value is smaller.

[0047] Preferably, the operation in A2 to reconstruct the noise-suppressed signal to obtain a preliminary magnetic resonance image is as follows:

[0048] E1. Construct a deep convolutional neural network model for reconstructing and enhancing the sharpness of noise-suppressed images. The model adopts an encoder-decoder structure. The encoder part consists of multiple convolutional layers and pooling layers to extract image features, and the decoder part consists of multiple deconvolutional layers and upsampling layers to reconstruct the extracted features into a high-resolution image.

[0049] E2. The constructed CNN model is trained using labeled magnetic resonance image data. During the training process, mean squared error and structural similarity index are used as loss functions. The network parameters are continuously adjusted through the backpropagation algorithm to improve the reconstructed image output by the model.

[0050] This invention provides a method for suppressing magnetic resonance signal noise and enhancing image sharpness. Compared with existing technologies, it has the following advantages:

[0051] 1. This magnetic resonance signal noise suppression and image sharpness enhancement method extracts and divides the preprocessed magnetic resonance image into multiple local regions to analyze and determine the noise level of each local region. It then performs noise suppression operations on the noise-affected local regions through multi-stage nonlinear filtering, and reconstructs the initial magnetic resonance image from the noise-suppressed signal. The method employs a multi-stage nonlinear filtering strategy, using adaptive median filtering based on local variance to remove impulse noise, and then uses an improved bilateral filtering to further smooth the image and preserve edge information. This effectively avoids the contradiction between noise suppression and edge preservation in traditional filtering methods, and better preserves image details and edge information.

[0052] 2. This magnetic resonance signal noise suppression and image sharpness enhancement method calculates and compares the mean and variance of each local region with the mean and variance of the current global magnetic resonance image. Based on the comparison and according to the set adjustment rules, the adaptive weight coefficient α is dynamically determined. This method can dynamically adjust the noise estimation parameters according to the characteristics of different regions of the image, improve the accuracy of noise estimation, and provide a more reliable basis for subsequent noise suppression.

[0053] 3. This magnetic resonance signal noise suppression and image sharpness enhancement method constructs a deep convolutional neural network model to reconstruct and enhance the sharpness of noise-suppressed images. The model adopts an encoder-decoder structure. The encoder part consists of multiple convolutional layers and pooling layers to extract image features, and the decoder part consists of multiple deconvolutional layers and upsampling layers to reconstruct the extracted features into a high-resolution image. It is trained with a large amount of labeled magnetic resonance image data to learn the intrinsic features and structural information of the image, which can automatically recover lost details and edges, enhance the edge information and contrast of the image, and significantly improve the image sharpness. Attached Figure Description

[0054] Figure 1 This is a flowchart illustrating the operation of the noise suppression and image sharpness enhancement method of the present invention. Detailed Implementation

[0055] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0056] Please see Figure 1 This invention provides two technical solutions:

[0057] Example 1: A method for suppressing magnetic resonance signal noise and enhancing image sharpness, specifically including the following steps:

[0058] A1. Use magnetic resonance imaging equipment to collect raw magnetic resonance signal data of the human body to be examined, and extract the data after consultation with the cooperating hospital to achieve optimized training operation;

[0059] A2. Implement preprocessing operations on the acquired data, extract the preprocessed magnetic resonance image and divide it into multiple local regions for analysis and determination of the noise level of each local region, perform noise suppression operations on the local regions affected by noise through multi-stage nonlinear filtering, and perform image reconstruction on the noise-suppressed signal to obtain a preliminary magnetic resonance image.

[0060] A3. Post-process the reconstructed magnetic resonance images to optimize the visual effects and highlight the edges and details of the images. Output the post-processed high-definition magnetic resonance images to the display interface for image display or to the storage terminal for storage.

[0061] By extracting and dividing the preprocessed magnetic resonance image into multiple local regions for analysis to determine the noise level of each local region, noise suppression is performed on the local regions affected by noise through multi-stage nonlinear filtering. The noise-suppressed signal is then used to reconstruct the preliminary magnetic resonance image. A multi-stage nonlinear filtering strategy is adopted, using adaptive median filtering based on local variance to remove impulse noise. Then, an improved bilateral filtering is used to further smooth the image and preserve edge information, effectively avoiding the contradiction between noise suppression and edge preservation in traditional filtering methods, and better preserving the details and edge information of the image.

[0062] In this embodiment of the invention, the preprocessing operation for the collected data implemented in A2 is as follows:

[0063] B1. Extract the acquired raw magnetic resonance signal data and classify the data according to the set nodes;

[0064] B2. Convert the format of the data under the corresponding node, convert the signal data into image data suitable for subsequent processing, and associate the parameter data corresponding to the image data with the image data;

[0065] B3. Remove abnormal data after conversion and use normalization method to normalize the signal strength to the range of [0, 1] to eliminate the dimensional differences between different data.

[0066] In this embodiment of the invention, the data classification operation in B1 according to the set nodes is as follows:

[0067] b11. The nodes are set according to the data categories, forming multi-level classification nodes to distinguish the data;

[0068] b12. The detected part category is used as the primary classification node, and the image and numerical categories are used as secondary classification nodes based on the primary classification node. The data acquisition timestamp is used as the final classification node based on the secondary classification node.

[0069] b13. Match data according to the category of the category node. That is, if there are characters in the collected data that are the same as the category node, the corresponding data will be extracted and introduced into the category node. The first-level category node, the second-level category node and the last-level category node will establish edge relationships. The last-level category node, the second-level category node and the first-level category node will establish edge relationships in sequence, and the edge direction will be from the last-level category node to the second-level category node, and the second-level category node will point to the first-level category node.

[0070] In this embodiment of the invention, the operation in A2 of extracting and dividing the preprocessed magnetic resonance image into multiple local regions for analysis and determination of the noise level of each local region is as follows:

[0071] C1. Extract magnetic resonance images from the preprocessed data, divide the magnetic resonance images into multiple local regions at equal intervals according to a set distance, and set the pixels in the local regions sequentially from top to bottom and from left to right.

[0072] C2. Calculate and compare the mean and variance of each local region with the mean and variance of the current global magnetic resonance image, and dynamically determine the value of the adaptive weight coefficient α based on the comparison and the set adjustment rules.

[0073] C3. Then, based on the value of the adaptive weighting coefficient α, the noise estimation values ​​in the local area are calculated, and the noise levels of each area are determined by sorting the noise estimation values.

[0074] In this embodiment of the invention, the operation in C2 that dynamically determines the value of the adaptive weight coefficient α based on the comparison situation and according to the set adjustment rules is as follows:

[0075] c21. Based on the deviation between local and global images of magnetic resonance images in historical data, set the threshold interval [m, n] and the corresponding weight coefficient α under the threshold change;

[0076] c22. When the absolute value of the variance L between the extracted local region image and the current global magnetic resonance image is compared with the set threshold interval [m, n];

[0077] If L < m, the similarity between the local region image and the current global magnetic resonance image is low, while the corresponding weight coefficient α is the largest.

[0078] If L > n, the similarity between the local region image and the current global magnetic resonance image is high, and the corresponding weight coefficient α is the smallest.

[0079] If L∈[m, n], the similarity between the local region image and the current global magnetic resonance image changes continuously. As L increases, the corresponding weight coefficient α decreases accordingly, and the corresponding weight coefficient α is obtained based on the corresponding threshold interval [m, n].

[0080] c23. Based on the real-time extracted local region image, extract the adaptively changing weight coefficient α and complete the noise assessment of the local region.

[0081] In this embodiment of the invention, the operation of calculating and comparing the mean and variance of the local region in C2 with the mean and variance of the current global magnetic resonance image is as follows:

[0082] Calculate the local mean value and label it J. h J h J represents the mean of the h-th local region. hk =(K1+K2+…+K k ) / k,K k The value of the k-th pixel in the h-th local region is used as an example. Similarly, the mean value of the global magnetic resonance image is JH.

[0083] Calculate the local variance scale as X h X h Let represent the variance of the h-th local region, and the calculation formula is: Similarly, the mean value of the global magnetic resonance image is XH;

[0084] The calculated absolute difference between the mean and absolute values ​​of the current local region and the global magnetic resonance image is |JH-J h |, while the absolute difference of variance between the current local region and the global magnetic resonance image |XH-X h |, that is, when the absolute difference of the mean conforms to the normal fluctuation range, the absolute difference of the variance is determined and the corresponding adaptive weight coefficient α is matched;

[0085] The formula for calculating the noise estimate in the local area based on the adaptive weighting coefficient α in C3 is: Y = α × X h +(1-α)×XH.

[0086] By calculating and comparing the mean and variance of each local region with the mean and variance of the current global magnetic resonance image, and dynamically determining the value of the adaptive weight coefficient α based on the comparison and the set adjustment rules, the noise estimation parameters can be dynamically adjusted according to the characteristics of different regions of the image, thereby improving the accuracy of noise estimation and providing a more reliable basis for subsequent noise suppression.

[0087] In this embodiment of the invention, the noise suppression operation in A2 for the local region affected by noise through multi-stage nonlinear filtering is as follows:

[0088] D1. Extract the local areas affected by noise and establish a neighborhood window centered on each pixel in the image;

[0089] D2. Set the estimated noise value used for comparison to S. Compare the difference between the median value and the center pixel value in the current window with the estimated noise value. If the difference exceeds the estimated noise value by r times, expand the window uniformly in all directions and recalculate the median value until the difference meets the condition of being less than the noise value by r times or expands to the set maximum window size.

[0090] The median value is the middle value of the sorted pixel values ​​in the current window. If the number of pixels is odd, it is the value of the middle pixel. If the number of pixels is odd, it is the average of the two middle pixel values. The center pixel value is the value of the pixel located at the center of the current window.

[0091] D3. After completing adaptive median filtering, the image is smoothed, and bilateral filtering is used to optimize image details and preserve edges.

[0092] In this embodiment of the invention, the bilateral filtering in D3 performs the optimization of image details and edge preservation as follows:

[0093] d31. Based on adaptive median filtering, extract a pixel from the image and establish a square window centered on the current pixel;

[0094] d32. The spatial distance weight is wz, which is determined by the distance from any pixel in the square window to the center pixel. The grayscale similarity weight is wt, which is determined by the difference in grayscale value between any pixel and the center pixel. Multiple rays are drawn from the center pixel to the boundary of the square window, with an angle of 45° between adjacent rays. The values ​​of adjacent pixels on the extended rays are extracted and the gradient magnitude is determined to obtain the local gradient adjustment weight wu.

[0095] d33. Combine the spatial distance weight wz, gray-level similarity weight wt, and local gradient adjustment weight wu to adjust the gray value of the current center pixel.

[0096] In this embodiment of the invention, the operation of d33 to adjust the gray value of the current center pixel by combining the spatial distance weight wz, the gray-level similarity weight wt, and the local gradient adjustment weight wu is as follows:

[0097] I1=I0(i,j)×wz(i,j)×wt(i,j)×wu(i,j);

[0098] Where I1 represents the adjusted grayscale value of the center pixel of the current square window, I0(i,j) represents the initial grayscale value of pixel (i,j), wz(i,j) represents the spatial distance weight of pixel (i,j), wt(i,j) represents the grayscale similarity weight of pixel (i,j), wu(i,j) represents the local gradient adjustment weight of pixel (i,j), and wz+wt+wu=1;

[0099] Furthermore, the closer a pixel is to the center pixel, the greater its influence on the center pixel, and the larger its spatial distance weight value. Pixels with similar gray values ​​to the center pixel have a larger gray similarity weight value. In edge regions with larger gradients, the local gradient adjustment weight value is smaller.

[0100] In this embodiment of the invention, the operation in A2 to reconstruct the noise-suppressed signal to obtain a preliminary magnetic resonance image is as follows:

[0101] E1. Construct a deep convolutional neural network model for reconstructing and enhancing the sharpness of noise-suppressed images. The model adopts an encoder-decoder structure. The encoder part consists of multiple convolutional layers and pooling layers to extract image features, and the decoder part consists of multiple deconvolutional layers and upsampling layers to reconstruct the extracted features into a high-resolution image.

[0102] E2. The constructed CNN model is trained using labeled magnetic resonance image data. During the training process, mean squared error and structural similarity index are used as loss functions. The network parameters are continuously adjusted through the backpropagation algorithm to improve the reconstructed image output by the model.

[0103] A deep convolutional neural network model is constructed to reconstruct and enhance the sharpness of noise-suppressed images. The model adopts an encoder-decoder structure. The encoder consists of multiple convolutional and pooling layers to extract image features, while the decoder consists of multiple deconvolutional and upsampling layers to reconstruct high-resolution images from the extracted features. The model is trained on a large amount of labeled magnetic resonance image data to learn the intrinsic features and structural information of the images. It can automatically recover lost details and edges, enhance the edge information and contrast of the images, and significantly improve the sharpness of the images.

[0104] Example 2 differs from Example 1 in that this invention selects a set of clinical magnetic resonance imaging (MRI) images as the experimental subject, including images of different parts such as the head, abdomen, and joints. A 3.0T MRI scanner is used for data acquisition, and the scanning parameters are optimized according to different body parts. The acquired raw images are divided into a training set and a test set. The training set is used to train a deep convolutional neural network model, and the test set is used to evaluate the performance of the method of this invention. Specific experimental results are as follows:

[0105] Noise suppression effect: Compared with traditional linear filtering, wavelet transform, and nonlocal mean filtering methods, the adaptive noise estimation and multi-stage nonlinear filtering method proposed in this invention can more accurately estimate the noise level and effectively remove different types of noise. Experimental results show that, under the same noise level, the signal-to-noise ratio (SNR) of the image processed by the method of this invention is more than 30 dB higher than that of traditional methods, and it can better preserve the details and edge information of the image.

[0106] Image sharpness enhancement effect: After deep learning-based image reconstruction, the image sharpness and resolution are significantly improved. Compared with the original low-quality image, the reconstructed image has improved by more than 85% in structural similarity index (SSIM), and the accuracy of doctors' diagnosis of lesions has also increased by 15%. Through subjective evaluation, doctors generally believe that the image processed by the method of this invention can more clearly show the shape, size and location of lesions, providing stronger support for clinical diagnosis.

[0107] Furthermore, any content not described in detail in this specification is existing technology known to those skilled in the art.

[0108] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0109] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method of magnetic resonance signal noise suppression and image sharpness enhancement, characterized by: Specifically comprising the following steps: A1, using a magnetic resonance imaging device to collect the original magnetic resonance signal data of the human body to be detected, and after consultation with the cooperative hospital, the data is extracted to realize the optimization training operation; A2, realize the preprocessing operation of the collected data, extract and divide the preprocessed magnetic resonance image into multiple local areas to analyze and determine the noise level of each local area, and perform noise suppression operation on the local area with noise influence through multi-stage nonlinear filtering processing, and image reconstruction is performed on the noise suppressed signal to obtain a preliminary magnetic resonance image; A3, post-processing the reconstructed magnetic resonance image, optimizing the visual effect of the image and highlighting the edge and detail information of the image, outputting the high-definition magnetic resonance image after post-processing to the display interface for image display or the storage terminal for storage.

2. The method of magnetic resonance signal noise suppression and image sharpness enhancement of claim 1, wherein: The preprocessing operation of the collected data in A2 is: B1, extract the collected original magnetic resonance signal data, and classify the data according to the set nodes; B2, format conversion is performed on the data under the corresponding node, the signal data is converted into image data suitable for subsequent processing, and the parameter data corresponding to the image data is associated with the image data; B3, remove the abnormal data after conversion, and normalize the signal intensity to the range of [0, 1] to eliminate the dimensional difference between different data.

3. The method of magnetic resonance signal noise suppression and image sharpness enhancement of claim 2, wherein: The data classification operation according to the set nodes in B1 is: b11, the set nodes are divided according to the category of the data, and a multi-level classification node is formed to distinguish the data; b12, the category of the detected part is taken as the first classification node, the image and numerical category are taken as the secondary classification node based on the first classification node, and the time stamp of data collection is taken as the final classification node based on the secondary classification node; b13, the data is matched according to the category of the classification node, that is, the same character as the classification node exists in the collected data, the corresponding data is extracted and introduced into the classification node, and the first classification node, the secondary classification node and the final classification node establish edge association, and the final classification node, the secondary classification node and the first classification node establish edge relationship in turn, and the edge direction points from the final classification node to the secondary classification node, and the secondary classification node to the first classification node.

4. The method of magnetic resonance signal noise suppression and image sharpness enhancement of claim 1, wherein: The operation of extracting and dividing the preprocessed magnetic resonance image into multiple local areas to analyze and determine the noise level of each local area in A2 is: C1, extract the magnetic resonance image from the preprocessed data, and divide the magnetic resonance image into multiple local areas according to the set distance, and set the pixel points in the local area from top to bottom and left to right; C2, calculate and compare the mean and variance of each local area with the mean and variance of the current magnetic resonance global image, dynamically determine the value of the adaptive weight coefficient α according to the set adjustment rule combined with the comparison; C3, then calculate the noise estimation value in the local area based on the value of the adaptive weight coefficient α, and determine the noise level of each area according to the noise estimation value and sorting.

5. The method of magnetic resonance signal noise suppression and image sharpness enhancement of claim 4, wherein: The operation of dynamically determining the value of the adaptive weight coefficient a in the C2 according to the comparison of the conditions and the set adjustment rules is: c21, setting the threshold interval [m, n] and the weight coefficient a corresponding to the threshold value change according to the deviation between the local region image and the global image of the magnetic resonance image under the historical data; c22, comparing the absolute value L of the variance of the extracted local region image and the current magnetic resonance global image with the set threshold interval [m, n]; If L < m, the similarity between the local region image and the current magnetic resonance global image is low, and the corresponding weight coefficient a is maximum; If L > n, the similarity between the local region image and the current magnetic resonance global image is high, and the corresponding weight coefficient a is minimum; If L ∈ [m, n], the similarity between the local region image and the current magnetic resonance global image is constantly changing, and with the increase of L, the corresponding weight coefficient a decreases, and the corresponding weight coefficient a is obtained according to the corresponding threshold interval [m, n]; c23, extracting the adaptive changing weight coefficient a according to the real-time extracted local region image and completing the noise evaluation of the local region.

6. The method of magnetic resonance signal noise suppression and image sharpness enhancement of claim 4, wherein: The operation of calculating and comparing the mean and variance of the local region with the mean and variance of the current magnetic resonance global image in the C2 is: The local region mean value is marked as J h , J h represents the hth local region mean value, J hk = (K1+K2+…+K k ) / k, K k refers to the value of the kth pixel point in the hth local region, and the mean value of the magnetic resonance global image is JH; The local region variance is calculated as X h , X h represents the variance of the hth local region, and the calculation formula is: The mean value of the magnetic resonance global image is XH, and the calculation formula is: The mean absolute difference between the current local region and the magnetic resonance global image is calculated as |JH-J h |, and the variance absolute difference between the current local region and the magnetic resonance global image is |XH-X h |, that is, when the absolute difference of the mean value meets the normal fluctuation amplitude, the variance absolute difference condition is determined and the corresponding adaptive weight coefficient a is matched. The formula for calculating the noise estimate in the local area based on the adaptive weighting coefficient α in C3 is: Y = α × X h +(1-α)×XH.

7. The method of magnetic resonance signal noise suppression and image sharpness enhancement of claim 1, wherein: The operation of noise suppression of the local region with noise influence in the A2 through multi-stage nonlinear filtering processing is: D1, extracting the local region with noise influence, establishing a neighborhood window with each pixel point in the image as the center; D2, setting the estimated noise value S for comparison, comparing the difference between the median value in the current window and the center pixel value with the estimated noise value, if the difference exceeds r times of the estimated noise value, uniformly expanding the window in each direction and recalculating the median value until the difference meets the condition of less than r times of the noise value or expands to the set maximum window size; The median value is the middle value of the sorted pixel point values in the current window, if the number of pixel points is odd, it is the middle pixel point value, if the number of pixel points is odd, it is the average of the middle two pixel point values, and the center pixel value is the pixel value of the center position in the current window; D3, after completing the adaptive median filtering, the image is smoothed, and the optimization of image details and the reservation of edges are completed through bilateral filtering.

8. The method of magnetic resonance signal noise suppression and image sharpness enhancement of claim 7, wherein: The operation of bilateral filtering to complete the optimization of image details and the reservation of edges in the D3 is: d31, based on the adaptive median filtering, extracting a pixel point in the image and establishing a square window with the current pixel point as the center; d32, determining the spatial distance weight wz according to the distance of any pixel point in the square window to the center pixel point, determining the gray similarity weight wt according to the difference between the gray value of any pixel point and the center pixel point, taking the center pixel point as the center to draw multiple rays extending to the boundary of the square window, and the angle between adjacent rays is 45°, and the gradient amplitude of the adjacent pixel point values on the extended rays is determined to obtain the adjustment weight wu of the local gradient; The d33 adjusts the gray value of the current center pixel point by combining the spatial distance weight wz, the gray similarity weight wt and the adjustment weight wu of the local gradient.

9. The magnetic resonance signal noise suppression and image sharpness enhancement method of claim 8, wherein: The operation of adjusting the gray value of the current center pixel point by combining the spatial distance weight wz, the gray similarity weight wt and the adjustment weight wu of the local gradient in the d33 is: I1=I0(i,j)×wz(i,j)×wt(i,j)×wu(i,j); Wherein I1 represents the adjusted gray value of the center pixel point of the current square window, I0(i,j) represents the initial gray value of the pixel point (i,j), wz(i,j) represents the spatial distance weight of the pixel point (i,j), wt(i,j) represents the gray similarity weight of the pixel point (i,j), wu(i,j) represents the adjustment weight of the local gradient of the pixel point (i,j), and wz+wt+wu=1; And the closer the pixel point to the center pixel point, the greater the influence on the center pixel, the greater the spatial distance weight value, the greater the gray similarity weight value of the pixel point with the center pixel gray value, and the smaller the adjustment weight value of the local gradient in the edge region with greater gradient.

10. The method of magnetic resonance signal noise suppression and image sharpness enhancement of claim 1, wherein: The operation of obtaining a preliminary magnetic resonance image by image reconstruction on the noise-suppressed signal in the A2 is: E1, a deep convolutional neural network model is constructed for reconstructing and clarity enhancement on the image after noise suppression, the model adopts an encoder-decoder structure, the encoder part is composed of multiple convolutional layers and pooling layers for extracting image features, and the decoder part is composed of multiple deconvolutional layers and up-sampling layers for reconstructing the extracted features into a high-resolution image; E2, the constructed CNN model is trained using labeled magnetic resonance image data, in the training process, mean square error and structural similarity index are used as loss functions, and network parameters are continuously adjusted through a back propagation algorithm so that the reconstructed image output by the model.

Images