Multi-scale non-local low-dose CT image denoising method based on region adaptation

Abstract

The present invention provides a multi-scale non-local low-dose CT image denoising method based on regional adaptation. The method adopts an adaptive search window, an adaptive multi-scale block size, and an adaptive filter coefficient for different image regions. Specifically, in non-edge regions, an isotropic square search window of 29 pixels in length and 29 pixels in width is used as a similar point region; in regions containing edges, an isotropic square search window of 15 pixels in length and 15 pixels in width is used as a similar pixel candidate point. Then, based on pixel point classification information and edge extraction information, an anisotropic similar point region along the edge direction is obtained. Then, multi-scale weighted non-local mean denoising is used to calculate denoised pixel values ​​in the determined similar point region. In addition, to better remove stripe artifacts and speckle noise, the denoising smoothing parameter and multi-scale action coefficient are adaptively changed according to noise intensity and intuitionistic fuzzy divergence theory. Therefore, the method effectively solves the problems existing in the prior art.

Description

Technical Field

[0001] The present invention belongs to the technical field of medical image processing, and in particular relates to a multi-scale non-local low-dose CT image denoising method based on regional adaptation. Background Art

[0002] Computed tomography (CT) technology has become a widely used imaging modality in recent years. With its increasing clinical use, concerns about the potential radiation hazards of CT scanning to the human body have garnered increased attention, leading to the development of low-dose CT technology. However, low-dose CT often causes image degradation, leading to the formation of streak artifacts and noise, significantly reducing the ability to distinguish normal tissue from pathological structures.

[0003] Currently, methods that can achieve some denoising results for low-dose CT images include the non-local means (NLM) algorithm. However, because images are generally composed of smooth, edge, and textured regions, the NLM algorithm uses the same image block size and search window size for all pixels in the entire image. This limits its performance, particularly in areas containing edges, where denoising results fall short of expectations. Furthermore, due to the uneven distribution of stripe artifacts, applying the same filter parameters across the entire image can lead to oversmoothing and loss of detail, or incomplete denoising, resulting in residual noise. Therefore, a denoising method is urgently needed to address these issues. Summary of the Invention

[0004] In order to address the shortcomings and deficiencies of the existing technology, a multi-scale non-local low-dose CT image denoising method based on regional adaptation is provided, thereby solving the problem that the existing technology cannot effectively denoise the image.

[0005] To achieve the purpose of the present invention, a multi-scale non-local low-dose CT image denoising method based on regional adaptation is provided, which includes the following steps:

[0006] Step 1: Pixel classification and edge extraction: Use the classification information of local blocks to extract edges, and then distinguish between edge and non-edge areas:

[0007] Step 2: Region-based candidate pixel screening: Based on the edge extraction results, for non-edge regions, an isotropic search window with a size of 29 pixels long and 29 pixels wide is selected as the set of similar pixels; for regions containing edges, an isotropic search window with a size of 15 pixels long and 15 pixels wide is selected as the candidate pixels. Then, based on the pixel classification information and edge information from step 1, dissimilar pixels are removed, and finally a set of similar pixels along the edge direction is obtained.

[0008] Step 3: Region-based multi-scale non-local denoising: After determining the corresponding similar pixels, a multi-scale weighted non-local denoising method is used to denoise the low-dose CT image;

[0009] Step 4: Adaptive filtering parameters based on intuitionistic fuzzy divergence: The filtering parameters of each pixel are adaptively adjusted using the intuitionistic fuzzy divergence between the smooth template and the local block centered on the current pixel.

[0010] As a further improvement to the above solution, the pixel classification and edge extraction in step 1 includes the following steps:

[0011] Step 1: Select a local block of size 3 pixels long × 3 pixels wide to represent each pixel in the image, and rearrange these local blocks into a vector;

[0012] Step 2: Use the K-means clustering method to divide these local blocks into K categories, and use the classification number obtained for each local block as the classification number of the center point of the local block;

[0013] Step 3: Calculate the gradient of all classification numbers. The places where the gradient modulus is not equal to 0 are edge points. Connecting these edge points completes edge extraction. The two adjacent pixel points with different classification numbers are the edge positions of the image.

[0014] As a further improvement of the above scheme, in step 3, the image is subjected to multi-scale non-local denoising, and the coefficients for controlling the degree of effect of different scales are adaptively determined according to the regional information. Finally, the denoising results at different scales are weighted to obtain the final denoising result. The denoised image Each pixel value in is:

[0015]

[0016] in, For the noisy image at position , R1 and R2 are the sets of similar pixels determined in step 2. is the size of the image block, N1 and N2 are the image block sizes at different scales; and is the coefficient that controls the degree of action of different scales and satisfies ;

[0017] When using an image block of 5 pixels long and 5 pixels wide, it is located at The pixel point of the point is located at The weight coefficient of the pixel point, When using an image block of 9 pixels long and 9 pixels wide, it is located at The pixel point of the point is located at The weight coefficients of the pixel points are calculated by the following formula:

[0018]

[0019] in, represents the Gaussian weighted distance, Indicates that the center point is located at The image block of the point, is an image block The total number of pixels in is the smoothing parameter that controls the denoising.

[0020] As a further improvement of the above scheme, when calculating the intuitionistic fuzzy divergence between the smooth template and the local block in step 4, a template matrix of size 3 × 3 with all element values ​​0 is used to represent the smooth area, which is a fuzzy set and is recorded as ; in pixels The area with a size of 3 pixels long and 3 pixels wide as the center represents the local block centered on the current pixel, which is another fuzzy set and is recorded as ; Calculated by the following formula and The intuitionistic fuzzy divergence between:

[0021]

[0022]

[0023]

[0024]

[0025]

[0026] Among them, Presentation Elements Fuzzy Set The membership degree of Presentation Elements Fuzzy Set The membership degree, Presentation Elements Fuzzy Set The degree of hesitation, Presentation Elements Fuzzy Set The degree of hesitation, Representing fuzzy sets 、 The intuitionistic fuzzy divergence between Presentation Elements 、 The intuitionistic fuzzy divergence between yes Pixels in The coordinates of yes Pixels in The intuitionistic fuzzy divergence of yes The grayscale value of the pixel in yes The pixel value at the corresponding position in .

[0027] As a further improvement to the above scheme, the coefficients for controlling the degree of effect of different scales in step 3 are adjusted using regional variance and intuitionistic fuzzy divergence:

[0028]

[0029]

[0030] in: yes Pixels in The intuitionistic fuzzy divergence of .

[0031] As a further improvement of the above solution, in step 4: calculate 、 When the intuitionistic fuzzy divergence between two fuzzy sets is calculated, a membership function that uses both gradient information and local variance information is adopted; for Pixels in For example:

[0032] The membership function is

[0033]

[0034] in, It's a pixel The local normalized variance of is the coefficient;

[0035] The hesitation function is

[0036]

[0037] in, is the coefficient;

[0038] for Pixels in For example:

[0039] The membership function is

[0040]

[0041] The hesitation function is

[0042]

[0043] in, is the coefficient.

[0044] As a further improvement of the above scheme, when The larger the value is, the more the pixel is contaminated by stripe artifacts and noise. The denoising smoothing parameter required for this pixel is:

[0045]

[0046] in, Indicates that it is located Pixel denoising smoothing parameter, is the standard deviation of the noise.

[0047] The beneficial effects of the present invention are:

[0048] Compared with the prior art, the present invention provides a multi-scale non-local low-dose CT image denoising method based on regional adaptation. This method uses adaptive search windows, adaptive multi-scale block sizes, and adaptive filter coefficients for different image regions. Specifically, for non-edge regions, an isotropic square search window of 29 pixels in length and 29 pixels in width is used to obtain a set of similar points. For regions containing edges, a set of similar points along the edge direction is obtained in an isotropic region of 15 pixels in length and 15 pixels in width based on pixel classification information and edge information. Multi-scale weighted non-local mean denoising is then used to calculate denoised pixel values ​​within the determined set of similar points. Furthermore, to better remove streak artifacts and speckle noise, the denoising smoothing parameters and the action coefficients at different scales are adaptively changed based on noise intensity and intuitionistic fuzzy divergence theory.

[0049] In summary, this method effectively solves the problem that the existing technology cannot effectively denoise low-dose CT images, especially the edge areas. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 It is the technical roadmap of the present invention;

[0051] Figure 2 Schematic diagram of the results of edge extraction of Shepp-Logan head model images using different edge extraction methods;

[0052] Figure 3 This is a schematic diagram of obtaining similar pixels in the edge area;

[0053] Figure 4 This is a comparative diagram of the denoising results of simulated images using various related algorithms;

[0054] Figure 5 This is a schematic diagram of the partial enlargement and comparison of the denoising results of simulated images using various related algorithms;

[0055] Figure 6 This is a schematic diagram of the partial enlargement comparison of the denoising results of real images by various related algorithms;

[0056] Figure 7 This is a schematic diagram comparing the denoising results of real images using multiple related algorithms. DETAILED DESCRIPTION

[0057] The specific embodiments of the present invention are further described in detail below with reference to the accompanying drawings:

[0058] The present invention provides a multi-scale non-local low-dose CT image denoising method based on regional adaptation, comprising the following steps:

[0059] Step 1: Pixel classification and edge extraction: Use the classification information of the local block to realize edge extraction, and then realize the distinction between edge areas and non-edge areas. The specific method is as follows:

[0060] Step 1: Select a local block of size 3 pixels × width 3 pixels to represent each pixel in the image and rearrange these local blocks into a vector;

[0061] Step 2: Use the K-means clustering method to divide these local blocks into K categories, and use the classification number obtained for each local block as the classification number of the center point of the local block;

[0062] Step 3: Calculate the gradient of all classification numbers. The places where the gradient modulus is not equal to 0 are edge points. Connecting these edge points completes edge extraction. The two adjacent pixel points with different classification numbers are the edge positions of the image.

[0063] Figure 2 The results of edge extraction on Shepp-Logan head model images using different edge extraction methods are shown. The first row, from left to right, shows the original image, the edge image extracted using the Sobel operator, the edge image extracted using the Canny operator, and the edge image extracted in step 1. The second row shows the low-dose CT image and the corresponding edge extraction results. The third row shows the image after preprocessing and denoising and the corresponding edge extraction results. It can be seen that the edges in the fourth column of images are richer and more accurate than those in the second and third columns of images. The edge extraction method used in the present invention is more robust to streak artifacts.

[0064] Step 2: Region-based similar pixel screening: Based on the edge extraction results, for non-edge regions, an isotropic search window of 29 pixels in length and 29 pixels in width is selected as the set of similar pixels; for regions containing edges, an isotropic search window of 15 pixels in length and 15 pixels in width is selected as candidate pixels. Dissimilar pixels are then removed based on the pixel classification information and edge information from step 1, ultimately obtaining a set of similar pixels along the edge direction.

[0065] However, a large number of pixels with significantly different values ​​from the center pixel may still be included in the search window, especially in the area near the edge, such as Figure 3 As shown in (a), it is a search window of a point, in which the pixel point has three grayscale values ​​with significant differences.

[0066] However, using pixel values ​​with large differences as candidate pixels can easily lead to ringing effects, blurring boundaries, and even producing false edges. To improve denoising performance, dissimilar pixels are removed based on the pixel classification information obtained in step 1. Figure 3 (b) is the pixel classification result. By removing pixels of different classifications, Figure 3 The remaining black pixels in (c) are similar pixels used for denoising the center pixel.

[0067] Step 3: Region-based multi-scale non-local denoising: After determining the corresponding similar pixels, a multi-scale weighted non-local denoising method is used to denoise the low-dose CT image;

[0068] According to experience and experiments, multi-scale image blocks are selected for weight calculation, and finally the denoising results at different scales are weighted to obtain the final denoising result. Each pixel value in is:

[0069]

[0070] in, For the noisy image at position , R1 and R2 are the sets of similar pixels determined in step 2. is the size of the image block, N1 and N2 are the image block sizes at different scales.

[0071] and is the coefficient that controls the degree of action of different scales and satisfies .

[0072] When using an image block of 5 pixels long and 5 pixels wide, it is located at The pixel point of the point is located at The weight coefficient of the pixel point, When using an image block of 9 pixels long and 9 pixels wide, it is located at The pixel point of the point is located at The weight coefficients of the pixel points are calculated by the following formula:

[0073]

[0074] in, represents the Gaussian weighted distance, Indicates that the center point is located at The image block of the point, is an image block The total number of pixels in is the smoothing parameter that controls the denoising.

[0075] Step 4: Adaptive filtering parameters based on intuitionistic fuzzy divergence: The filtering parameters of each pixel are adaptively adjusted using the intuitionistic fuzzy divergence between the smooth template and the local block centered on the current pixel;

[0076] When calculating the intuitionistic fuzzy divergence between the smooth template and the local block, a template matrix of size 3 × 3 with all element values ​​0 is used to represent the smooth area, which is a fuzzy set and is recorded as ; in pixels The area with a size of 3 pixels long and 3 pixels wide as the center represents the local block centered on the current pixel, which is another fuzzy set and is recorded as ; Calculated by the following formula and The intuitionistic fuzzy divergence between:

[0077]

[0078]

[0079]

[0080]

[0081]

[0082] Among them, Presentation Elements Fuzzy Set The membership degree, Presentation Elements Fuzzy Set The membership degree, Presentation Elements Fuzzy Set The degree of hesitation, Presentation Elements Fuzzy Set The degree of hesitation, Representing fuzzy sets 、 The intuitionistic fuzzy divergence between Presentation Elements 、 The intuitionistic fuzzy divergence between Therefore Pixels in The coordinates of yes Pixels in The intuitionistic fuzzy divergence of yes The grayscale value of the pixel in yes The pixel value at the corresponding position in .

[0083] In addition, regional variance and intuitionistic fuzzy divergence are used to adjust the multi-scale effect coefficient in step 3, as follows:

[0084]

[0085]

[0086] in: yes Pixels in The intuitionistic fuzzy divergence of .

[0087] calculate 、 When the intuitionistic fuzzy divergence between two fuzzy sets is calculated, a membership function that uses both gradient information and local variance information is adopted; for Pixels in For example:

[0088] The membership function is

[0089]

[0090] in, It's a pixel The local normalized variance of is the coefficient;

[0091] The hesitation function is

[0092]

[0093] in, is the coefficient;

[0094] for Pixels in For example:

[0095] The membership function is

[0096]

[0097] The hesitation function is

[0098]

[0099] in, is the coefficient.

[0100] when The larger the value is, the more the pixel is contaminated by stripe artifacts and noise, and the pixel needs a larger denoising smoothing parameter:

[0101]

[0102] in, Indicates that it is located Pixel denoising smoothing parameter, is the standard deviation of the noise.

[0103] The denoising results of this method are compared with those of various algorithms, as follows:

[0104] 1. Image results:

[0105] 1.1 Simulated Image

[0106] The performance of the proposed method was evaluated by conducting experiments on Shepp-Logan head model images. Figure 4-6 shown. Figure 4 (a) is a simulated clean image. Figure 4 (b) is a low-dose CT image obtained by reconstructing the simulated noisy sinusoidal image using the FBP method (Hanning filter, with a cutoff frequency of 80% Nyquist frequency). Figure 4 (c)-(f) show the denoising results using NLM, TV, GLD-NLM and the method of the present invention, respectively; Figure 5 and Figure 6 The magnified details of two regions of interest are shown, namely Figure 4 The red rectangles in (a) mark ROI1 and ROI2.

[0107] Among them: TV model as a classic image restoration method has been successfully applied to low-dose CT images and achieved good denoising effect, but the TV denoising algorithm is prone to cause block effect and detail loss, such as Figure 5In (d), the details are clearly blurred; Figure 6 As shown in (d), there is an obvious block effect near the edge area.

[0108] The classic NLM algorithm is superior in preserving image details, but it is not conducive to removing stripe artifacts. Figure 4 As shown in (c), many stripe artifacts still exist in the denoised image; Figure 5 (c) and Figure 6 As can be seen from the image in (c), the area in the low-dose CT image that was originally severely contaminated by streak artifacts is still full of noise in the denoised image.

[0109] The GLD-NLM algorithm has a good denoising effect on Gaussian noise, but its performance is poor for streak artifacts and speckle noise in low-dose CT images.

[0110] By comparison, the method of the present invention has a better denoising effect on low-dose CT images. Figure 5 (f) and Figure 6 As can be seen from (f) in the figure, the noise is almost removed and the image edge is relatively clear.

[0111] The image denoising effect was objectively evaluated using peak signal-to-noise ratio (PSNR) and structural similarity (SSIM). Table 1 shows the objective evaluation of denoising for Shepp-Logan head model images. It can be seen that the proposed method achieves the highest PSNR and SSIM values.

[0112] Table 1 Comparison of objective indicators of Sheep-Logan head model image denoising using different algorithms

[0113]

[0114] 1.2 Real Images

[0115] In addition, low-dose clinical real images from Mayo Clinic in the United States were used to verify the superiority of the method of the present invention. Figure 7 (a1) in the figure is an abdominal image slice with a size of 512×512.

[0116] The denoising results of the proposed method are compared with those of the NLM, GLD-NLM and TV methods. Figure 7 As shown in (a1) - (f1) in . Figure 7 (a2) - (f2) in the figure show the locally enlarged images of the white box area in (a1) (in this case, it refers to the liver tumor marked with a white box). Figure 7 (b1) in the figure is a normal dose clinical abdominal image slice. Figure 7 (c1) - (f1) in the figure show the denoising results using NLM, GLD-NLM, TV and the method of the present invention respectively.

[0117] As can be seen, the images obtained using the NLM and GLD-NLM methods are still affected by noise and artifacts. The TV method achieves better results in noise removal and detail preservation, but has a significant blocking effect in noisy areas. In contrast, the image denoised using the method of the present invention is significantly reduced in size, and the image edges are relatively clear.

[0118] In order to evaluate the quality of the denoised images, PSNR and SSIM techniques were used to quantitatively evaluate the entire clinical abdominal image and the region of interest (liver tumor). Figure 7 The other six ROI regions (indicated by red boxes) in (a1) are quantitatively described and analyzed. The three smooth regions are marked as ROI1, ROI2, and ROI3, and the three edge regions are marked as ROI4, ROI5, and ROI6.

[0119] Table 2 compares the PSNR and SSIM values ​​of various denoising methods for the entire image, the tumor region, and the region of interest. It can be seen that the method of the present invention achieves the highest PSNR value for the entire image, followed by the SSIM value. It also achieves the highest PSNR value across all ROIs and the highest SSIM value across most ROIs. Clearly, the method of the present invention strikes a good balance between noise reduction and edge preservation.

[0120] Table 2 Comparison of objective indicators of real low-dose CT image denoising using different algorithms

[0121]

[0122] The above embodiments are not limited to the technical solutions of the embodiments themselves, and the embodiments can be combined with each other to form new embodiments. The above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit them. Any modifications or equivalent replacements that do not depart from the spirit and scope of the present invention should be included within the scope of the technical solutions of the present invention.

Claims

1. A multi-scale non-local low-dose CT image denoising method based on regional adaptation, characterized by: The steps include: Step 1: Pixel classification and edge extraction: Use the classification information of local blocks to extract edges, and then distinguish between edge and non-edge areas: Step 2: Region-based similar pixel screening: Based on the edge extraction results, for non-edge regions, an isotropic search window of 29 pixels in length and 29 pixels in width is selected as the set of similar pixels; for regions containing edges, an isotropic search window of 15 pixels in length and 15 pixels in width is selected as candidate pixels. Dissimilar pixels are then removed based on the pixel classification information and edge information from step 1, ultimately obtaining a set of similar pixels along the edge direction. Step 3: Region-based multi-scale non-local denoising: After determining the corresponding similar pixels, a multi-scale weighted non-local denoising method is used to denoise the low-dose CT image; Perform multi-scale non-local denoising on the image, and adaptively determine the coefficients that control the degree of effect of different scales based on regional information. Finally, perform weighted operation on the denoising results at different scales to obtain the final denoising result. The denoised image Each pixel value in is: ； in, For the noisy image at position , R1 and R2 are the sets of similar pixels determined in step 2. is the size of the image block, N1 and N2 are the image block sizes at different scales; and is the coefficient that controls the degree of action of different scales and satisfies ; When using an image block of 5 pixels long and 5 pixels wide, it is located at The pixel point of the point is located at The weight coefficient of the pixel point, When using an image block of 9 pixels long and 9 pixels wide, it is located at The pixel point of the point is located at The weight coefficients of the pixel points are calculated by the following formula: ； in, represents the Gaussian weighted distance, Indicates that the center point is located at The image block of the point, is an image block The total number of pixels in is the smoothing parameter that controls denoising; The coefficients controlling the degree of effect of different scales are adjusted using regional variance and intuitionistic fuzzy divergence: ； ； In which: pixel The area with a size of 3 pixels long and 3 pixels wide as the center represents the local block centered on the current pixel, which is a fuzzy set and is recorded as ; yes Pixels in The intuitionistic fuzzy divergence of Step 4: Adaptive filtering parameters based on intuitionistic fuzzy divergence: The filtering parameters of each pixel are adaptively adjusted using the intuitionistic fuzzy divergence between the smooth template and the local block centered on the current pixel.

2. The method for regional adaptive multi-scale non-local low-dose CT image denoising according to claim 1, characterized in that: The pixel classification and edge extraction of step 1 includes the following steps: Step 1: Select a local block of size 3 pixels long × 3 pixels wide to represent each pixel in the image, and rearrange these local blocks into a vector; Step 2: Use the K-means clustering method to divide these local blocks into K categories, and use the classification number obtained for each local block as the classification number of the center point of the local block; Step 3: Calculate the gradient of all classification numbers. The places where the gradient modulus is not equal to 0 are edge points. Connecting these edge points completes edge extraction. The two adjacent pixel points with different classification numbers are the edge positions of the image.

3. The method for regional adaptive multi-scale non-local low-dose CT image denoising according to claim 1, characterized in that: When calculating the intuitionistic fuzzy divergence between the smooth template and the local block in step 4, a template matrix of size 3 × 3 with all element values ​​0 is used to represent the smooth area, which is a fuzzy set and is recorded as ; Calculated by the following formula and The intuitionistic fuzzy divergence between: ； ； ； ； ； Among them, Presentation Elements Fuzzy Set The membership degree, Presentation Elements Fuzzy Set The membership degree, Presentation Elements Fuzzy Set The degree of hesitation, Presentation Elements Fuzzy Set The degree of hesitation, Representing fuzzy sets 、 The intuitionistic fuzzy divergence between Presentation Elements 、 The intuitionistic fuzzy divergence between yes Pixels in The coordinates of yes Pixels in The intuitionistic fuzzy divergence of yes The grayscale value of the pixel in yes The pixel value at the corresponding position in .

4. The method for regional adaptive multi-scale non-local low-dose CT image denoising according to claim 3, characterized in that: In the step 4: calculation 、 When the intuitionistic fuzzy divergence between two fuzzy sets is calculated, a membership function that uses both gradient information and local variance information is adopted; for Pixels in For example: The membership function is ； in, It's a pixel The local normalized variance of is the coefficient; The hesitation function is ； in, is the coefficient; for Pixels in For example: The membership function is ； The hesitation function is ； in, is the coefficient.

5. The method for regional adaptive multi-scale non-local low-dose CT image denoising according to claim 4, characterized in that: when The larger the value is, the more the pixel is contaminated by stripe artifacts and noise. The denoising smoothing parameter required for this pixel is: ； in, Indicates that it is located Pixel denoising smoothing parameter, is the standard deviation of the noise.

Images