A denoising enhancement method for fourier domain optical coherence tomography

Abstract

The application discloses a denoising enhancement method for Fourier domain optical coherence tomography, and belongs to the technical field of image processing, which comprises the following steps: obtaining sample data of an FD-OCT image and performing a standardization operation on the sample data; performing image grouping on the standardized image; constructing a strong reflection stripe template for each group of images according to the image grouping result, which is used for removing strong reflection stripe noise in the FD-OCT image; constructing a frequency domain notch filter template to further remove periodic noise in the image background; adopting a combination of spatial domain denoising and transform domain denoising to further remove speckle noise in the image; and performing an image enhancement operation on the denoised image. The denoising enhancement method for Fourier domain optical coherence tomography can simultaneously remove strong reflection stripe noise, periodic noise and speckle noise existing in the FD-OCT and improve the overall visual effect of the image.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, and more specifically to a denoising and enhancement method for Fourier domain optical coherence tomography. Background Technology

[0002] Fourier Domain Optical Coherence Tomography (FD-OCT) is an imaging technique based on the principle of low-coherence optical interference and centered on a Michelson interferometer. It can acquire sample depth information without damage, offering advantages such as micrometer-level high resolution, non-contact operation, and non-destructive testing. It is widely used in medical imaging and industrial non-destructive testing. Common noise types in FD-OCT include strong reflection fringe noise, thermal noise, speckle noise, and periodic noise. Among these, strong reflection fringe noise, periodic noise, and speckle noise have the greatest impact on the imaging quality of FD-OCT. Strong reflection fringe noise originates from reflections from the surfaces of optical components in the system's optical path; periodic noise originates from the inherent noise of the system's electronic components; and speckle noise arises from the superposition of noise generated by the backscattering of multiple photons.

[0003] In existing technologies, common FD-OCT denoising methods mainly include spatial domain denoising and transform domain denoising. Both types of denoising methods have their unique applicable ranges. For example, spatial domain denoising is faster, but its effect on removing periodic noise is poor, while transform domain denoising will lose edge information in the image to some extent, causing image boundary blurring. With the development of computer technology, deep learning algorithms are also widely used in image denoising. For example, Chinese patent CN113935906A discloses a method for removing strong reflection stripe noise from Fourier domain optical coherence tomography. This method first collects FD-OCT data of rough objects as label data, adds random noise to obtain corresponding sample data, and forms a training dataset. After preprocessing through wavelet transform, the data is input into a prediction model for model training. The prediction model consists of a 10-layer convolutional neural network and a concat operation. The prediction model is iteratively optimized using a loss function to obtain the optimal network parameters. The sample data containing stripe noise is input into the optimized prediction model, and the output data is subjected to inverse wavelet transform to obtain the image after stripe noise removal. However, deep learning-based denoising methods heavily rely on training samples; the quality of the sample set directly affects image quality. While the sample set is primarily derived by simulating the mathematical types of noise, this idealized noise model differs significantly from the real noise encountered in FD-OCT, making it difficult to obtain a perfect training sample set. Therefore, effectively removing strong reflective stripe noise, speckle noise, and periodic noise simultaneously is a crucial issue for the application and expansion of FD-OCT. Summary of the Invention

[0004] To address the problems existing in the prior art, this invention provides a noise reduction and enhancement method for Fourier domain optical coherence tomography (FD-OCT), aiming to simultaneously remove strong reflection stripe noise, periodic noise, and speckle noise present in FD-OCT, and improve the overall visual effect of the image. To achieve the above objectives, this invention provides the following technical solution:

[0005] A noise reduction and enhancement method for Fourier domain optical coherence tomography includes the following steps:

[0006] S1, acquire sample data of FD-OCT images and perform standardization operations on them;

[0007] S2, group the images processed in step S1;

[0008] S3, extract a group of images after the image grouping in step S2, and construct a strong reflection stripe template based on the group of images;

[0009] S4. Based on the strong reflection stripe template constructed in step S3, remove the strong reflection stripe noise from the image;

[0010] S5, Construct a frequency domain notch filter template to remove periodic noise from the background of the image processed in step S4;

[0011] S6, using a combination of spatial domain denoising and transform domain denoising, to remove speckle noise from the image processed in step S5;

[0012] S7. Perform image enhancement operation on the image processed in step S6 to improve the overall visual effect of the image.

[0013] S8. Iterate through the group of images in sequence, repeating steps S4 to S7 to complete the denoising and enhancement operation for the group of images.

[0014] S9. Loop through all image groups in sequence, repeat steps S3 to S8, and complete the denoising and enhancement operation for the corresponding images.

[0015] Furthermore, the calculation formula for the standardized operation is as follows: Where I0(x,y) represents the input image, min(*) represents obtaining the minimum grayscale value of the image, max(*) represents obtaining the maximum grayscale value of the image, and [*] represents the rounding function. standard (x,y) represents the standardized result image.

[0016] Furthermore, the steps for constructing the strong reflective stripe template include: image overlay step, image binarization step, image morphological operation step, and image connected component labeling step.

[0017] Furthermore, the filtering criteria for the connected component results include:

[0018] Condition 1: The length of the bounding rectangle of the connected component is greater than 50% of the total length of the image;

[0019] Condition 2: The pixel area of ​​the connected component is greater than 50% of the area of ​​the bounding rectangle.

[0020] Furthermore, constructing the frequency domain notch filter template includes the following steps:

[0021] Assuming the size of the acquired FD-OCT image is M*N, generate a matrix template of size M*N with all values ​​being 1;

[0022] At position M / 2 of the matrix, cover it with two matrices of all zeros, each with A rows and (N / 2-2*A) columns. The starting points of the two matrices are (M / 2-A / 2,A) and (M / 2-A / 2,N / 2+A) respectively.

[0023] At position N / 2 of the matrix, cover it with two matrices containing all zeros, each with (M / 2-2*A) rows and A columns. The starting points of the two matrices are (A, N / 2-A / 2) and (M / 2+A, N / 2-A / 2) respectively.

[0024] The above matrix template is convolved using a Gaussian convolution kernel to obtain the blurred frequency domain notch filter template.

[0025] Furthermore, the method of combining spatial domain denoising and transform domain denoising is a combination of adaptive median filtering denoising and wavelet threshold denoising.

[0026] Furthermore, the method of combining adaptive median filtering denoising with wavelet thresholding to remove speckle noise includes the following steps:

[0027] S61, First, perform adaptive median filtering on the image processed in step S5;

[0028] S62, perform two-dimensional wavelet decomposition on the image after adaptive median filtering to obtain the decomposition coefficients c of each layer of the wavelet decomposition. i Let i = 1, 2, ..., j, and calculate c. i The standard deviation ε;

[0029] S63, determine the threshold T for wavelet threshold denoising based on the standard deviation ε;

[0030] S64, using a threshold to perform hard thresholding denoising on the wavelet decomposition coefficients of the image;

[0031] S65, using thresholding to perform soft thresholding denoising on the wavelet decomposition coefficients of the image;

[0032] S66, the result image of wavelet threshold denoising is obtained by calculation.

[0033] Furthermore, in the wavelet thresholding denoising, the wavelet basis function used for the two-dimensional wavelet decomposition of the image is the sym5 function, and the number of decomposition layers is 2; the threshold T = 3*ε; the resulting image of the wavelet thresholding denoising is a weighted average of the hard thresholding denoising result and the soft thresholding denoising result.

[0034] Furthermore, the image enhancement operation involves enhancing the denoised image using a limited contrast histogram equalization method.

[0035] A computer-readable storage medium includes a computer program that can be loaded by a processor and executed as described above for a denoising enhancement method for Fourier domain optical coherence tomography.

[0036] The beneficial effects of this invention are:

[0037] 1. This invention provides a denoising and enhancement method for Fourier domain optical coherence tomography (FD-OCT). Based on the characteristics of FD-OCT images, a corresponding denoising algorithm is designed. Strong reflection fringe templates are constructed to eliminate strong reflection fringe noise. Frequency domain notch filters are designed to remove periodic noise. Spatial domain denoising methods are combined with transform domain denoising methods to eliminate the influence of speckle noise. Finally, the FD-OCT images are enhanced to improve the overall visual effect of the images.

[0038] 2. The present invention provides a denoising enhancement method for Fourier domain optical coherence tomography. In the process of generating a frequency domain notch filter template to remove periodic noise in the background, Gaussian blur is introduced. The introduction of Gaussian blur can effectively reduce the ringing phenomenon in the image caused by the steep edge of the filter.

[0039] 3. The present invention provides a denoising enhancement method for Fourier domain optical coherence tomography. In the process of wavelet combination threshold denoising to remove speckle noise, soft threshold denoising has a better noise removal effect, while hard threshold denoising has a higher degree of image detail preservation. The combination of the two can both preserve more details and achieve the effect of noise removal. Attached Figure Description

[0040] Figure 1 This is a flowchart of a noise reduction and enhancement method for Fourier domain optical coherence tomography provided by the present invention;

[0041] Figure 2 This is a two-dimensional image after the read-in FD-OCT has been standardized in this embodiment of the invention;

[0042] Figure 3 This is a schematic diagram of a highly reflective stripe template constructed according to an embodiment of the present invention;

[0043] Figure 4 This is the result image of removing strong reflection stripe noise in an embodiment of the present invention;

[0044] Figure 5 This is a schematic diagram of a frequency domain notch filter template constructed according to an embodiment of the present invention;

[0045] Figure 6 This is the result image after removing periodic noise in an embodiment of the present invention;

[0046] Figure 7 This is the result image after removing speckle noise in an embodiment of the present invention;

[0047] Figure 8 This is the result image after image enhancement operation in the embodiments of the present invention;

[0048] Figure 9 This is a comparison chart of the results of embodiments of the present invention, wherein, Figure 9 (a) is the original image before processing, and (b) is the result image after processing using the present invention. Detailed Implementation

[0049] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments, but the present invention is not limited to the following embodiments.

[0050] Example 1:

[0051] See attached Figures 1-9 This invention provides a noise reduction and enhancement method for Fourier domain optical coherence tomography, the flowchart of which is shown below. Figure 1 As shown, the process includes image reading, data normalization, image grouping, removal of strong reflective stripe noise, removal of periodic noise, removal of speckle noise, and image enhancement steps. The specific process is as follows:

[0052] Image reading steps: The DICOM format FD-OCT image file is read in and converted into a three-dimensional matrix composed of Num two-dimensional B-Scan images, which is then stored. In this embodiment, Num = 500, the size of the two-dimensional B-Scan image is 512*512, and it is converted into a 512*512*500 three-dimensional matrix for storage.

[0053] Data standardization steps: Extract each 2D B-Scan image from the 3D matrix sequentially, and perform standardization operations on them. The standardized B-Scan image is shown below. Figure 2 As shown. The formula for calculating standardized operations is: Where I0(x,y) represents the input image, min(*) represents obtaining the minimum grayscale value of the image, max(*) represents obtaining the maximum grayscale value of the image, and [*] represents the rounding function. standard (x,y) represents the standardized result image.

[0054] Image grouping step: All standardized FD-OCT B-Scan images are grouped into m groups, resulting in a total of n groups, where Num = m * n. To enhance the strong reflective stripe regions in subsequent superimposed images, m is greater than or equal to 10. In this embodiment, m = 100 and n = 5.

[0055] The steps for removing strong reflective stripe noise include constructing a strong reflective stripe template and removing the strong reflective stripe noise. Specifically, firstly, the first group of images after image grouping, or a specific group of images, is extracted, and a strong reflective stripe template is obtained based on this group of images; then, a logical AND operation is performed between the 2D B-Scan image and the strong reflective stripe template image to eliminate the strong reflective stripe noise; finally, bilinear interpolation is used to remove holes in the resulting image after the logical operation, restoring the image to its original size.

[0056] The specific process for obtaining the strong reflective stripe template is as follows:

[0057] Image overlay: Overlay a set of 100 two-dimensional B-Scan images and perform standardization operations as described above on the overlaid image;

[0058] Image binarization: Image binarization is performed using image thresholding. The threshold can be selected using the Otsu thresholding method to obtain the binarized image result.

[0059] Image morphological operations: Perform morphological opening operations on a binary image to remove isolated noise points in the image. The structuring element for the morphological opening operation is a rectangular element with a size of 3*3.

[0060] By using connected component labeling, connected components of a binary image are filtered to obtain the bounding rectangles and areas of the connected components. The corresponding connected components are then selected based on the following criteria:

[0061] Condition 1: The length of the bounding rectangle of the connected components is greater than 50% of the overall length of the image;

[0062] Condition 2: The area of ​​the connected region is greater than 50% of the area of ​​the circumscribed rectangle;

[0063] Connected components that satisfy both conditions one and two are selected, and the binary image containing only connected components is inverted to generate a template for removing strong reflective stripes.

[0064] The specific process for removing strong reflection stripe noise is as follows:

[0065] Extract the B-Scan image from this dataset and perform a logical AND operation with the strong reflective fringe template. The calculation formula is: I out (x,y)=I in (x,y)&I mask (x,y), in the formula, I in (x,y) represents the input B-Scan image, I mask (x,y) represents the strong reflection fringe template, I out (x,y) represents the processed image;

[0066] Then, the gray values ​​of each row of pixels in the resulting image are summed, and the rows with a sum of gray values ​​of 0 are filtered out and removed, thereby removing holes in the image.

[0067] Image interpolation is performed on the removed image to restore its original dimensions. Bilinear interpolation can be used as the image interpolation method.

[0068] Removing periodic noise: Periodic noise in the B-Scan image is further removed by designing a corresponding frequency domain notch filter template. The specific process is as follows:

[0069] S51, the image result after removing strong reflection stripe noise is mapped to the frequency domain using a two-dimensional discrete Fourier transform. The calculation formula for the two-dimensional discrete Fourier transform is as follows:

[0070]

[0071] In the above formula, I in (x,y) represents the input image after removing strong reflection stripes, and F(u,v) represents the resulting image after the two-dimensional discrete Fourier transform; M represents the number of rows in the image, and N represents the number of columns in the image. Simultaneously, after the two-dimensional discrete Fourier transform, the transform result is spectrally shifted.

[0072] S52: Generate the template for the frequency domain notch filter. First, generate a matrix template F of size M*N, all values ​​being 1. mask ;

[0073] At position M / 2 of the matrix, two matrices with 5 rows and (N / 2-10) columns, each containing only zeros, are used to cover the matrix. The starting points of the two matrices are (M / 2-2, 5) and (M / 2-2, N / 2+5), respectively. In this embodiment, M*N = 512*512, so the starting points of the two matrices are (254, 5) and (245, 261), respectively.

[0074] At position N / 2 of the matrix, two matrices with all zeros (M / 2-10 rows and 5 columns) are used to cover the matrix. The starting points of the two matrices are (5, N / 2-2) and (M / 2+5, N / 2-2) respectively. In this embodiment, M*N = 512*512, so the starting points of the two matrices are (5, 254) and (261, 245) respectively.

[0075] S53: Subsequently, Gaussian blurring is applied to the notch filter template using a Gaussian convolution kernel. The Gaussian convolution kernel can be selected as 7*7 in size, with a mean of 0 and a variance σ. 2 =3, the notch filter template after blurring is as follows Figure 5 As shown.

[0076] S54: Compare the frequency domain F(u,v) image with the frequency domain notch filter template F mask Perform a logical AND operation to generate the filtered result F. out (u,v), the calculation formula is: F out (u,v)=F(u,v)&F mask (u,v).

[0077] S55: Filter the result F out After inverse frequency shifting (u,v), a two-dimensional discrete Fourier inverse transform is performed to obtain the time-domain result image I(x,y), which is calculated using the following formula:

[0078]

[0079] Speckle noise removal steps: An adaptive median filtering combined with wavelet thresholding is used to further remove speckle noise from the B-Scan image. The results are as follows: Figure 7 As shown, the specific process is as follows:

[0080] S61: First, perform adaptive median filtering on the B-Scan image to remove obvious speckle noise in the background area;

[0081] S62: Further perform two-dimensional wavelet decomposition on the B-Scan image, selecting the sym5 wavelet basis function. In this embodiment, the original image undergoes a two-level wavelet decomposition; extract the decomposition coefficients c of each level of the wavelet decomposition. i And calculate c i Standard deviation ε:

[0082]

[0083] In the above formula, Denotes the decomposition coefficient c i The average value, j represents the decomposition coefficient c. i The number of;

[0084] S63: Determine the threshold for wavelet threshold denoising. Set the threshold for wavelet denoising as T, and its calculation formula is: T = 3 * ε.

[0085] S64: Hard thresholding denoising is performed on the wavelet decomposition coefficients of the image using a threshold. The calculation formula is as follows:

[0086]

[0087] After modifying the wavelet coefficients, using c i_hard Image reconstruction using wavelet coefficients yields hard-threshold image results. hard ;

[0088] S65: Soft thresholding denoising is performed on the wavelet decomposition coefficients of the image using a threshold. The calculation formula is as follows:

[0089]

[0090] In the above formula, sgn(*) is a symbolic function in mathematics. After modifying the wavelet coefficients, it is used with c i_soft Image reconstruction using wavelet coefficients yields hard-threshold image results. soft ;

[0091] S66: Image I showing the result of wavelet threshold denoising calculation out The specific calculation formula is as follows:

[0092] I out =I hard +I soft

[0093] Image enhancement steps: Contrast-limited histogram equalization is used to enhance the denoised image, improving the visual effect of the B-Scan image. The enhancement result is as follows: Figure 8 As shown.

[0094] The algorithm iterates through each or part of the two-dimensional B-Scan images in the image set, repeating the steps of removing strong reflective stripe noise, removing periodic noise, removing speckle noise, and image enhancement to complete the denoising and enhancement operation of the image set.

[0095] The algorithm iterates through all image groups or partial groups in turn to perform image denoising and enhancement operations on the corresponding images.

[0096] Finally, the denoised and enhanced 3D matrix is ​​rewritten into a DICOM file, and the patient's name, patient ID, and other relevant information read from the DICOM file are transferred to the new DICOM file.

[0097] The technical effects of the present invention will be further explained in conjunction with experiments:

[0098] like Figure 9 As shown, Figure 9 It is a comparison image of an FD-OCT image, in which, Figure 9 (a) is the original image before processing, and (b) is the result image after processing using the present invention. It is clear from the original image that there are strong reflective stripes running throughout the entire image, as well as obvious periodic noise and speckle noise in the background. After processing the original image using the denoising and enhancement method of the present invention, it can be seen that in the result image, the strong reflective stripe noise and the periodic noise in the background have been removed, the background is cleaner, and the influence of speckle noise has been basically eliminated.

[0099] Furthermore, this invention also utilizes equivalent numbers of looks (ENL) to quantitatively evaluate the effectiveness of its denoising and enhancement method. Equivalent numbers of looks are an index that measures the smoothness of a uniform region. Typically, a relatively uniform region in the image background is selected to calculate the change in equivalent numbers of looks before and after processing. The calculation process is as follows:

[0100]

[0101] In the above formula, μ represents the mean of the image gray level, and σ represents the standard deviation of the image gray level. For two images, image regions at the same location are selected for evaluation of the equivalent number of views. The selected regions are as follows: Figure 9 As shown in the red box. Through calculation, Figure 9 The equivalent number of views for the area within the red box in (a) is 0.3261. Figure 9 (b) The equivalent number of views for the area within the red box is 349.1648. Therefore, the equivalent number of views for the image processed by the denoising enhancement method of the present invention is significantly improved, indicating that the present invention has a strong noise removal effect.

[0102] The above description is merely a preferred embodiment of the present invention and does not limit the patent scope of the present invention. Any equivalent structural or procedural transformations made based on the content of the present invention's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of the present invention.

Claims

1. A method for denoising and enhancing Fourier domain optical coherence tomography, characterized in that, Includes the following steps: S1, acquire sample data of FD-OCT images and perform standardization operations on them; S2, group the images processed in step S1; S3, extract a group of images after the image grouping in step S2, and construct a strong reflection stripe template based on the group of images; S4. Based on the strong reflection stripe template constructed in step S3, remove the strong reflection stripe noise in the image; S5, Construct a frequency domain notch filter template to remove periodic noise from the background of the image processed in step S4; the construction of the frequency domain notch filter template includes the following steps: Assuming the size of the acquired FD-OCT image is Generate a size of A matrix template whose values ​​are all 1; In the matrix with row number At this location, use row number A and column number A. The two matrices are covered by a matrix of all zeros, with the starting points of the two matrices being respectively... and ; In the matrix with the number of columns At this point, the number of rows is A series of zero-filled matrices, A in number, are used to cover the matrix. The starting points of the two matrices are respectively... and ; The above matrix template is convolved using a Gaussian convolution kernel to obtain the blurred frequency domain notch filter template. S6, a combination of spatial domain denoising and transform domain denoising is used to remove speckle noise from the image processed in step S5; the combination of spatial domain denoising and transform domain denoising is a combination of adaptive median filtering denoising and wavelet thresholding denoising; the removal of speckle noise by combining adaptive median filtering denoising and wavelet thresholding denoising includes the following steps: S61, First, perform adaptive median filtering on the image processed in step S5; S62, perform two-dimensional wavelet decomposition on the image after adaptive median filtering to obtain the decomposition coefficients of each layer of the wavelet decomposition. and calculate Standard deviation ; S63, based on standard deviation Determine the threshold T for wavelet threshold denoising; S64, using a threshold to perform hard thresholding denoising on the wavelet decomposition coefficients of the image; S65, using thresholding to perform soft thresholding denoising on the wavelet decomposition coefficients of the image; S66, the result image of wavelet threshold denoising is obtained through calculation; In the wavelet thresholding denoising process, the wavelet basis function used for two-dimensional wavelet decomposition of the image is the sym5 function, and the decomposition level is 2; the threshold... The resulting image after threshold denoising is a weighted average of the hard threshold denoising result and the soft threshold denoising result. S7. Perform image enhancement operation on the image processed in step S6 to improve the overall visual effect of the image. S8. Iterate through the group of images in sequence, repeating steps S4 to S7 to complete the denoising and enhancement operation for the group of images. S9. Loop through all image groups in sequence, repeat steps S3 to S8, and complete the corresponding image denoising and enhancement operation.

2. The denoising enhancement method for Fourier domain optical coherence tomography according to claim 1, characterized in that, The calculation formula for the standardized operation is as follows: ,in, Indicates the input image. This indicates obtaining the minimum grayscale value of the image. This indicates that the maximum grayscale value of the image is obtained. This represents the floor function. The image represents the result of standardization.

3. The denoising enhancement method for Fourier domain optical coherence tomography according to claim 1, characterized in that, The steps for constructing a strong reflective stripe template include: image overlay, image binarization, image morphological operations, and image connected component labeling.

4. The denoising enhancement method for Fourier domain optical coherence tomography according to claim 3, characterized in that, The filtering criteria for the connected component results include: Condition 1: The length of the bounding rectangle of the connected component is greater than 50% of the total length of the image; Condition 2: The pixel area of ​​the connected component is greater than 50% of the area of ​​the bounding rectangle.

5. The denoising enhancement method for Fourier domain optical coherence tomography according to claim 1, characterized in that, The image enhancement operation involves enhancing the denoised image using a contrast-limited histogram equalization method.

6. A computer-readable storage medium, characterized in that, Includes a computer program that can be loaded by a processor and execute any of the noise reduction and enhancement methods for Fourier domain optical coherence tomography as described in claims 1-5.

Images