A method for removing noise from MRI images based on local directional statistical characteristics

Abstract

The present invention discloses a method for removing noise from MRI images based on local directional statistical characteristics, comprising the steps of: estimating noise on a noisy MRI image and obtaining a Gaussian-like function; obtaining a gradient direction basic model group under multiple directions under multiple product kernel sizes based on the MRI image; judging a sound point: judging whether each pixel point of the MRI image is a noise point according to the gradient direction basic model group; removing the noise point: if it is a noise point, denoising and restoring it through the Gaussian-like function. The present invention can denoise MRI images based on local factors, improve the denoising effect, and the denoising structure is closer to the actual situation.

Description

Technical Field

[0001] The invention belongs to the technical field of MRI image processing, and in particular relates to a method for removing MRI image noise based on local directional statistical characteristics. Background Art

[0002] MRI denoising methods can be roughly divided into: filtering methods, such as the unbiased non-local mean method, which uses the similarity of images to avoid distortion, but has problems such as large computational complexity; statistical methods, such as the linear minimum mean square error method that uses the minimum mean square error estimation formula to perform denoising based on statistical local information, which is fast but has the problem of edge distortion; and transform domain processing methods, such as wavelet transform.

[0003] Most of the current filtering methods for MRI denoising are based on the known variance of the noise. Often, the variance of the noise is assumed at the beginning of the processing. These methods have the following limitations: (1) The parameters used in the calculation are not the actual parameters of the image, so the denoising effect is not good; (2) At the same time, the previous methods did not consider the impact of the local area of ​​the MRI image on the denoising effect, which is difficult to meet actual needs; (3) Traditional methods such as statistical sorting filtering and optimal notch filtering also have the problem of only having good effects in specific environments and are difficult to handle large blocks of noise. Therefore, the above methods can only restore the image as a whole, and cannot determine whether a certain point is a noise point or a normal point, and cannot achieve pixel-level restoration. Summary of the invention

[0004] In order to solve the above problems, the present invention proposes a MRI image noise removal method based on local directional statistical characteristics, which can denoise MRI images based on local factors, improve the denoising effect, and make the denoising structure closer to the actual situation.

[0005] To achieve the above object, the technical solution adopted by the present invention is: a method for removing MRI image noise based on local directional statistical characteristics, comprising the steps of:

[0006] Perform noise estimation on noisy MRI images and obtain a Gaussian-like function;

[0007] Based on MRI images, a gradient direction base model group under multiple directions with multiple product kernel sizes is obtained;

[0008] Determine the sound point: for each pixel of the MRI image, determine whether it is a noise point based on the gradient direction basic model group;

[0009] Noise point removal: if it is a noise point, denoising and restoration are performed using the Gaussian-like function.

[0010] Furthermore, in order to cope with the noise variation of the local image and considering that there is no valid data information around the MRI image, the noise of the MRI image with noise is estimated and a Gaussian-like function is obtained, including the steps of:

[0011] The surrounding sub-image areas of the MRI image are collected as noise sub-images, and the variance of the sub-images is calculated;

[0012] A Gaussian-like function that estimates the noise using its variance.

[0013] Furthermore, the noise estimation is performed on the noisy MRI image, comprising the steps of:

[0014] Four noise sub-images are obtained using the surrounding sub-image regions, and the central moments of the four noise sub-images are calculated;

[0015] The variance is obtained by the central moment, and the variance is c = (c1 + c2 + c3 + c4) / 4; where c1, c2, c3 and c4 are the variances of the four noise sub-images respectively;

[0016] The Gaussian-like function of the noise is estimated using the variance, and the Gaussian-like function is = 255 / (2*pi*c*c); wherein c is the variance and pi is the circumference of a circle.

[0017] Furthermore, a gradient direction base model group under multiple directions under multiple product kernel sizes is obtained based on the MRI image, including the steps of:

[0018] Establish multiple groups of gradient direction image groups with different convolution kernel sizes, each group includes multiple directions;

[0019] By performing a convolution operation between the MRI image and each convolution kernel, a basic model group of gradient directions in different directions under each group of kernel sizes is obtained.

[0020] Furthermore, a gradient direction base model group under multiple directions under multiple product kernel sizes is obtained based on the MRI image, including the steps of:

[0021] Obtain three groups of gradient direction image groups with convolution kernel sizes of 3, 5, and 7, each group is divided into 4 directions of 0°, 45°, 90°, and 135°;

[0022] By convolving the MRI image with each convolution kernel, we obtain the gradient direction basic model group fn_α(x,y) in different directions with kernel sizes of 3, 5 and 7, where n is the convolution kernel size and α is the direction.

[0023] Furthermore, the sound point is determined: for each pixel point of the MRI image, the maximum model of the gradient direction basic model group with different convolution kernel sizes is calculated. If the maximum models under different convolution kernel sizes are equal, it is determined to be an information point, otherwise it is a noise point.

[0024] Furthermore, the method also includes the steps of post-processing the denoised MRI image: if there are noise points around the image, threshold segmentation is performed using the maximum inter-class variance method to protect the restored image in the middle and remove the noise points around it.

[0025] The beneficial effects of adopting this technical solution are:

[0026] Compared with the existing MRI denoising algorithms, it is difficult for the present invention to identify whether a certain pixel is a noise point, so overall denoising is often performed; the present invention determines and removes noise points through the statistical characteristics of local directions, thereby achieving better results.

[0027] The present invention can adaptively process images with various noises, has higher robustness and adaptability, and has a faster calculation speed than previous methods. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] Figure 1 It is a flow chart of a method for removing noise from MRI images based on local directional statistical characteristics of the present invention;

[0029] Figure 2 Schematic diagram of noise estimation of MRI images in an embodiment of the present invention;

[0030] Figure 3 A schematic diagram of obtaining a basic model group of gradient directions in an embodiment of the present invention;

[0031] Figure 4 This is a flow chart of determining noise points in an embodiment of the present invention;

[0032] Figure 5 The figure is a flow chart of a method for removing noise from MRI images based on local directional statistical characteristics optimized in an embodiment of the present invention. DETAILED DESCRIPTION

[0033] In order to make the purpose, technical solutions and advantages of the present invention more clear, the present invention is further described below with reference to the accompanying drawings.

[0034] In this embodiment, see Figure 1 As shown, the present invention proposes a method for removing MRI image noise based on local directional statistical characteristics, comprising the steps of:

[0035] Perform noise estimation on noisy MRI images and obtain a Gaussian-like function;

[0036] Based on MRI images, a gradient direction base model group under multiple directions with multiple product kernel sizes is obtained;

[0037] Determine the sound point: for each pixel of the MRI image, determine whether it is a noise point based on the gradient direction basic model group;

[0038] Noise point removal: if it is a noise point, denoising and restoration are performed using the Gaussian-like function.

[0039] As an optimization scheme 1 of the above embodiment, in order to cope with the noise change of the local image, and considering that there is no valid data information around the MRI image, the noise of the MRI image with noise is estimated, and a Gaussian-like function is obtained, including the steps of:

[0040] The surrounding sub-image areas of the MRI image are collected as noise sub-images, and the variance of the sub-images is calculated;

[0041] A Gaussian-like function that estimates the noise using its variance.

[0042] Specifically, the noise estimation is performed on the noisy MRI image, including the following steps:

[0043] Four noise sub-images are obtained using the surrounding sub-image regions, and the central moments of the four noise sub-images are calculated;

[0044] like Figure 2 As shown:

[0045] Assuming that the sizes of the images are X and Y respectively, and k is the acquisition coefficient, the acquired sub-image area around the image is ∈(1~k*X, 1~k*Y), (Xk*X~X, 1~k*Y), (1~k*X, Yk*Y~Y), (Xk*X~X, Yk*Y~Y);

[0046] The variance is obtained by the central moment, and the variance is c = (c1 + c2 + c3 + c4) / 4; where c1, c2, c3 and c4 are the variances of the four noise sub-images respectively;

[0047] The Gaussian-like function of the noise is estimated using the variance, and the Gaussian-like function is = 255 / (2*pi*c*c); wherein c is the variance and pi is the circumference of a circle.

[0048] As an optimization solution 2 of the above embodiment, a gradient direction basic model group in multiple directions with multiple product kernel sizes is obtained based on the MRI image, including the steps of:

[0049] Establish multiple groups of gradient direction image groups with different convolution kernel sizes, each group includes multiple directions;

[0050] By performing a convolution operation between the MRI image and each convolution kernel, a basic model group of gradient directions in different directions under each group of kernel sizes is obtained.

[0051] Specifically, a gradient direction basic model group under multiple directions with multiple product kernel sizes is obtained based on the MRI image, including the steps of:

[0052] like Figure 3 As shown:

[0053] Obtain three groups of gradient direction image groups with convolution kernel sizes of 3, 5, and 7, each group is divided into 4 directions of 0°, 45°, 90°, and 135°;

[0054] Example:

[0055] The convolution kernel with size 3 and orientation 0° is:

[0056] [-1-1-1;000;111];

[0057] The convolution kernel with size 5 and orientation 0° is:

[0058] [-2-2-2-2-2;-1-1-1-1-1;00000;11111;22222];

[0059] The convolution kernel with size 7 and orientation 0° is:

[0060] [-3-3-3-3-3-3-3; -2-2-2-2-2-2-2; -1-1-1-1-1-1-1; 0000000; 1111111; 2222222; 3333333].

[0061] By convolving the MRI image with each convolution kernel, we obtain the gradient direction basic model group fn_α(x,y) in different directions with kernel sizes of 3, 5 and 7, where n is the convolution kernel size and α is the direction.

[0062] Such as f3_0(x,y),f3_45(x,y),f5_90(x,y),f5_45(x,y), etc.

[0063] As an optimization scheme 3 of the above embodiment, the sound point is determined: for each pixel point of the MRI image, the maximum model of the gradient direction basic model group with different convolution kernel sizes is calculated. If the maximum models under different convolution kernel sizes are equal, it is determined to be an information point, otherwise it is a noise point.

[0064] Specifically, Figure 4 As shown. For each pixel x, y (x∈1-X, y∈1-Y), the maximum model maxf3(x,y), maxf5(x,y), maxf7(x,y) of the gradient direction basic model group with different convolution kernel sizes can be calculated.

[0065] maxf3(x,y)=argmax(f3_0(x,y),(f3_45(x,y),(f3_90(x,y),(f3_135(x,y)),

[0066] maxf5(x,y)=argmax(f5_0(x,y),(f5_45(x,y),(f5_90(x,y),(f5_135(x,y)),

[0067] maxf7(x,y)=argmax(f7_0(x,y),(f7_45(x,y),(f7_90(x,y),(f7_135(x,y))).

[0068] like:

[0069] maxf3(1,1)=argmax(f3-0(1,1),(f3-45(1,1),(f3-90(1,1),(f3-135(1,1)). If f3-0(1,1) is the largest, then maxf3(1,1)=0, which means that the direction of the pixel (1,1) is 0 degrees when the kernel size is 3.

[0070] If f3-90(2,2) is the largest in pixel (2,2), then maxf3(2,2)=90, which means that the direction of the pixel (2,2) when the kernel size is 3 is 90 degrees.

[0071] Then traverse each pixel point. The direction of the noise point is chaotic, while the texture of the image has a certain direction. Therefore, we can judge whether it is a noise point by whether the known directions of maxf3(i,j), maxf5(i,j), and maxf7(i,j) are consistent.

[0072] As an optimization scheme 4 of the above embodiment, Figure 5 As shown, the method also includes the steps of post-processing the denoised MRI image: if there are noise points around the image, threshold segmentation is performed using the maximum inter-class variance method to protect the restored image in the middle and remove the noise points around it.

[0073] The above shows and describes the basic principles and main features of the present invention and the advantages of the present invention. It should be understood by those skilled in the art that the present invention is not limited to the above embodiments. The above embodiments and descriptions are only for explaining the principles of the present invention. Without departing from the spirit and scope of the present invention, the present invention may have various changes and improvements, which fall within the scope of the present invention to be protected. The scope of protection of the present invention is defined by the attached claims and their equivalents.

Claims

1. A method for removing noise from MRI images based on local directional statistical characteristics, characterized in that: Includes steps: Perform noise estimation on noisy MRI images and obtain a Gaussian-like function; A gradient direction base model group under multiple directions under multiple convolution kernel sizes is obtained based on the MRI image, including the following steps: Establish multiple groups of gradient direction image groups with different convolution kernel sizes, each group includes multiple directions; By performing convolution operation between the MRI image and each convolution kernel, a basic model group of gradient directions in different directions under each group of kernel sizes is obtained; Determine noise points: For each pixel point of the MRI image, calculate the maximum model of the gradient direction basic model group with different convolution kernel sizes. If the maximum models under different convolution kernel sizes are equal, it is determined to be an information point, otherwise it is a noise point. Noise point removal: if it is a noise point, denoising and restoration are performed using the Gaussian-like function.

2. The method for removing noise from MRI images based on local directional statistical characteristics according to claim 1, characterized in that: Noise estimation is performed on a noisy MRI image and a Gaussian-like function is obtained, including the following steps: The surrounding sub-image areas of the MRI image are collected as noise sub-images, and the variance of the sub-images is calculated; A Gaussian-like function that estimates the noise using its variance.

3. The method for removing noise from MRI images based on local directional statistical characteristics according to claim 2, characterized in that: Noise estimation is performed on a noisy MRI image, including the following steps: Four noise sub-images are obtained using the surrounding sub-image regions, and the central moments of the four noise sub-images are calculated; The variance is obtained by the central moment, and the variance is c = (c1 + c2 + c3 + c4) / 4; where c1, c2, c3 and c4 are the variances of the four noise sub-images respectively; The Gaussian-like function of the noise is estimated by using the variance, and the Gaussian-like function is f(x)=255 / (2*pi*c*c); wherein c is the variance and pi is the circumference of a circle.

4. The method for removing noise from MRI images based on local directional statistical characteristics according to claim 1, characterized in that: A gradient direction base model group under multiple directions under multiple convolution kernel sizes is obtained based on the MRI image, including the following steps: Obtain three groups of gradient direction image groups with convolution kernel sizes of 3, 5, and 7, each group is divided into 4 directions of 0°, 45°, 90°, and 135°; By convolving the MRI image with each convolution kernel, we obtain the gradient direction basic model group fn_α(x,y) in different directions with kernel sizes of 3, 5 and 7, where n is the convolution kernel size and α is the direction.

5. A method for removing MRI image noise based on local directional statistical characteristics according to any one of claims 1 to 4, characterized in that: The method also includes the following steps: post-processing the denoised MRI image: if there are noise points around the image, threshold segmentation is performed using the maximum inter-class variance method to protect the restored image in the middle and remove the noise points around it.

Images