Multi-stage OCT image denoising method, device and storage medium based on improved clustering

Abstract

The present invention discloses a multi-stage OCT image denoising method, device and storage medium based on improved clustering, the method comprising: registering and averaging a plurality of original OCT images obtained by repeated scanning of the same position to obtain an average image; clustering using a plurality of features of the average image to obtain a weight coefficient template of a pixel; denoising the background of each original OCT image by a TV model, and then performing multi-scale decomposition to obtain a high-frequency image and a low-frequency image at each level of decomposition; performing guided filtering on the high-frequency image and directional filtering on the low-frequency image; restoring the decomposed images at each level step by step and performing weighted fusion with the previous level of low-frequency image to obtain a denoised and enhanced OCT image; registering and averaging a plurality of denoised and enhanced OCT images to obtain a final denoised OCT image. The method of the present invention can effectively reduce the image speckle noise while retaining the detail information in the image to a greater extent.

Description

Technical Field

[0001] The present invention relates to the technical field of medical image processing, and in particular to a multi-stage OCT image denoising method, device and storage medium based on improved clustering. Background Art

[0002] Optical coherence tomography (OCT) is a non-invasive optical imaging technology with the advantages of non-invasiveness, high resolution, fast imaging speed, and high detection sensitivity. It has been widely used in clinical diagnosis and scientific research. However, the noise in OCT imaging, especially the speckle noise caused by the high scattering characteristics of biological tissues, can lead to problems such as uneven imaging inside the tissue and unclear tissue edge structure, which seriously reduces the image quality and hinders clinicians from making accurate diagnosis and effective treatment of the disease. Therefore, achieving effective OCT image denoising is an important part of improving the clinical application value of OCT technology.

[0003] The difficulty of OCT image denoising lies in the need to effectively remove noise while retaining the true information of the tissue structure to the greatest extent. There are many related image denoising methods, such as bilateral filtering that considers both the spatial difference and pixel value difference of neighboring pixels, non-local mean filtering and BM3D algorithm that use global pixel structure similarity, TV algorithm that minimizes the total variation (TV) term of the image based on gradient descent (Reference: Xu, M., Tang, C., Chen, M., et al. Texture preservation and speckle reduction in optical coherence tomography using the shearlet-based total variation algorithm. Opt. Laser. Eng. 122(14), 265–283, 2019.), anisotropic diffusion filtering that controls the diffusion degree of neighboring pixels according to the size of pixel gradient, and a method that divides the image into tissue structure and noise background by clustering pixel features and performs filtering processing separately (Reference: Chen, M., Tang, C., Xu, M., et al. A clustering framework based on FCM and texture features for denoising ESPI fringe patterns with variable density. Opt. Laser. Eng. 119, 77–86, 2019.). The above methods have achieved good denoising effects, but it is still difficult to achieve a good balance between image denoising and detail preservation. In particular, when the image noise level increases, the effects of these denoising methods will be greatly reduced.

[0004] Deep learning is a new and powerful tool in the OCT image denoising task, but most deep learning methods still need to obtain denoised label images corresponding to the original OCT images for model training. The quality of the label images determines the upper limit of the quality of the final model output image. An effective method to obtain a label image dataset is to repeatedly scan the same position to obtain multiple images (or take multiple images with similar structures at adjacent positions), register them, and then average them as label images (reference: Geng MF, Meng XX, Zhu L., et al. Triplet cross-fusion learningfor unpaired image denoising in optical coherence tomography. IEEE Transactions on Medical Imaging, 41(11), 2022.). Since the noise distribution in the measurement process is random, the average fusion method can reduce image noise to a certain extent and retain more detailed information such as tissue microstructure and texture. However, the disadvantage of this method is that obtaining a denoised label image with good quality requires a large number of repeated measured images, which increases the difficulty of obtaining a dataset. The researchers (Reference: Xu M., Tang C., Hao F., et al. Texturepreservation and speckle reduction in poor optical coherence tomography usingthe convolutional neural network. Med. Image Anal., 64, 101727, 2020.) also proposed a progressive processing method that sequentially reduces image noise through the TV algorithm, enhances structural edges and textures through the diffusion model, and finally divides the image into tissue structure and noise background through pixel feature clustering and sets the background pixels to zero. However, this method still has the disadvantage that the tissue structure and tiny details are transitionally smoothed, and pixel clustering will cause obvious boundaries in the boundary area of ​​the structure.

[0005] The current various traditional OCT image denoising technologies are difficult to achieve a good balance between image denoising and detail preservation. While the noise in the image is reduced, the original tissue structure information is often damaged to varying degrees. For example, the tissue edge structure will be over-smoothed and become blurred, and some important details such as fine structures and textures will be lost. The current clustering method clusters the original OCT image and directly distinguishes the background area and tissue structure in the image based on the clustering results, that is, a pixel is either divided into the background area or the tissue structure. This method is prone to mistakenly classify some weak tissue structures in the image as background noise, which is smoothed or set to zero in the subsequent denoising process and loses detail information. In addition, rigidly dividing pixels into structure and background and processing them separately is also prone to produce obvious boundaries in the boundary area of ​​the structure. The multiple average fusion method requires a large number of repeatedly measured images (generally more than 50 images) to achieve denoising, which increases the difficulty of obtaining a data set, and the excessive number of images also increases the requirements for image registration accuracy. When there is an error in the registration result, the multiple average fusion not only fails to effectively remove the noise, but also introduces false tissue structure artifacts, reducing the reliability of the denoised image. Summary of the invention

[0006] The present invention provides a multi-stage OCT image denoising method based on improved clustering, which can effectively reduce image speckle noise while retaining detail information such as fine structure and texture in the image to a greater extent.

[0007] The technical solution of the present invention is as follows:

[0008] In a first aspect, the present invention provides a multi-stage OCT image denoising method based on improved clustering, comprising:

[0009] (1) Multiple original OCT images obtained by repeated scanning of the same location are registered and averaged to obtain an average image I mean ;

[0010] (2) Using the average image I mean Cluster multiple features to obtain the pixel weight coefficient template w mask ;

[0011] (3) Combined with weight coefficient template w mask , the background of each original OCT image is denoised by the TV model;

[0012] (4) Perform multi-scale decomposition on each original OCT image after background denoising to obtain high-frequency images and low-frequency images at each level of decomposition;

[0013] (5) Perform guided filtering on the high-frequency images at each level of decomposition, and perform directional filtering on the low-frequency images at each level of decomposition;

[0014] (6) For each original OCT image, the decomposed images at each level are restored step by step and weighted fused with the low-frequency image at the previous level to obtain a denoised and enhanced OCT image;

[0015] (7) Multiple denoised and enhanced OCT images are registered and averaged to obtain the final denoised OCT image.

[0016] Before averaging and fusing the images, the present invention first processes the background noise and tissue structure in each image separately and specifically, including reducing the noise in the background area and the tissue structure area and enhancing the edge and texture of the tissue structure. Finally, the enhanced images are averaged and fused to obtain high-quality denoised OCT images.

[0017] Step (2) includes:

[0018] (2-1) For the average image I mean Perform multi-scale decomposition to obtain the decomposed low-frequency images and high-frequency images at all levels, upsample the low-frequency images and high-frequency images at all levels to the average image I mean Same resolution; calculate average image I mean and gradient images of various levels of high-frequency images;

[0019] (2-2) Extracting the average image I mean , image features of low-frequency images and high-frequency images at all levels, and gradient images;

[0020] (2-3) Perform pixel principal component analysis on the extracted image features, retain the first n pixel principal component features according to the cumulative contribution rate and perform clustering, cluster the image pixels into two categories: structure area and background area, and calculate the weight coefficient template based on the clustering results w mask .

[0021] Furthermore, the weight coefficient template is calculated based on the clustering results. w mask include:

[0022] (i) Calculate the distance between the principal component features of the image pixels and the class center of the background area respectively D 0 , the distance from the cluster center of the structure area D 1 ;

[0023] (ii) Calculation of Δ d= D 0 - D 1 , and calculate Δ d The relative minimum value b:

[0024] (1)

[0025] (iii) Δ d Normalized to the interval [0-1]:

[0026] (2)

[0027] (iv) Perform gamma enhancement to obtain the weight coefficient template w mask :

[0028] (3)

[0029] in, k is the gamma enhancement factor.

[0030] In order to better distinguish between tissue structure and background area, the present invention comprehensively utilizes the feature information of the average fusion image and its multi-scale decomposition images at all levels to first perform pixel clustering, and then calculates the difference between the distance between the pixel feature and the center of the two classes, and finally obtains the weight coefficient template reflecting the tissue structure and background noise represented by each pixel through normalization and gamma enhancement. The advantage of this weight coefficient template is that the pixels are not rigidly divided into structure or background, but are assigned weight coefficients in the range [0,1]. The closer the weight coefficient is to 1, the more the pixel represents the tissue structure information, and vice versa, the more it represents the background noise. Based on this weight coefficient template, in subsequent processing, different degrees of denoising can be performed for tissue structure, background noise, and weak boundaries, and no obvious boundaries will be produced in the boundary area of ​​the structure, thereby achieving a good balance between denoising and detail retention.

[0031] Step (3) includes: denoising the background of each original OCT image based on the minimized TV model:

[0032] (4)

[0033] In the formula, the first term is the fidelity term; the second term is the total variation term, where I 0 ( i,j ) is the original noisy image, I ( i,j ) is the denoised image, Representing images I ( i,j ) is the sum of the absolute values ​​of the gradients in the horizontal and vertical directions, is the weight coefficient of the total variation term, C is a constant greater than 1.

[0034] The present invention uses the weight coefficient template generated as the weight coefficient of the total variation term in the TV method, assigns a larger weight coefficient to the TV term of background pixels, and assigns a smaller weight coefficient to the TV term of tissue structure pixels (equivalent to enhancing the weight of its fidelity term). The benefit brought by this is that the background noise can be removed more targetedly and the image structure detail information can be retained to the greatest extent.

[0035] In step (2) and step (4), the multi-scale decomposition is Laplacian pyramid decomposition. The high-frequency image is a Laplacian image, and the low-frequency image is a Gaussian image.

[0036] In step (5), guided filtering is performed on the high-frequency images at each level of decomposition to improve the noise at the tissue edge structure in the high-frequency images at each level of multi-scale decomposition; directional filtering is performed on the low-frequency images at each level of decomposition to reduce the noise in the tissue structure area in the low-frequency images at each level of multi-scale decomposition and enhance the edge and texture of the structure.

[0037] Guided filtering is performed on the high-frequency images at each level of decomposition, including: performing Gaussian filtering and smoothing on the high-frequency images, and using the images smoothed by Gaussian filtering as reference images to guide filtering on the high-frequency images.

[0038] Directional filtering is performed on low-frequency images at each level of decomposition, including:

[0039] (1) Using a directional edge detection operator, the gradients of each pixel in the low-frequency image along the horizontal, vertical, 45-degree and 135-degree directions are extracted, and the direction with the largest absolute value of the gradient is taken as the main direction of the pixel;

[0040] (2) Select the neighborhood around the pixel, calculate the variance of the main direction of the pixels in the neighborhood, and compare it with the preset threshold:

[0041] If the variance is less than the preset threshold, it is determined that the central pixel has an obvious main direction, and the line neighborhood is taken along the direction perpendicular to its main direction for spatial filtering;

[0042] If the variance is greater than a preset threshold, it is determined that the central pixel does not have an obvious main direction, and isotropic filtering is performed directly in the neighborhood.

[0043] The present invention uses a directional filtering method to process low-frequency images under multi-scale decomposition of images. It has three key features: first, the main direction of the pixel is determined by the direction of the maximum gradient value extracted by the directional edge detection operator; second, the pixels with obvious main directions (i.e., tissue edge structure pixels) are screened out by judging the variance threshold in the pixel neighborhood; third, the pixels with obvious main directions are subjected to line neighborhood filtering in the direction perpendicular to the main direction, and the pixels without obvious main directions are subjected to isotropic filtering. The advantage of this directional filtering method is that it can effectively improve the noise in the tissue structure area and enhance the tissue edge and texture, while avoiding directional filtering of the background noise to prevent the abnormal texture structure of the image.

[0044] The directional edge detection operator is a Sobel operator or a Prewitt operator.

[0045] Furthermore, the preset threshold is 0.83.

[0046] In step (6), starting from the highest level decomposition image, the high-frequency image after guided filtering and the low-frequency image after directional filtering are restored to the previous level step by step, and each level of restored image is weightedly fused with the original low-frequency image of the previous level until the original image resolution is restored to obtain a denoised and enhanced OCT image.

[0047] Step (6) includes:

[0048] (6-1) Starting from the highest level K-level decomposition image, restore the denoised low-frequency image and the corresponding high-frequency image to the next level K-1 level to obtain K -1 level restored image ;

[0049] (6-2) The weight coefficient template w mask Downsample to and G K-1 Same pixel resolution ;Will And the low-frequency image of the K-1th level decomposition after directional filtering G K-1 Perform weighted fusion to obtain the fused K -Level 1 low frequency image :

[0050] (5)

[0051] (6-3) and K The high-frequency image decomposed at level -1 is restored to the next level, level K-2. ;

[0052] (6-4) Repeat steps (G-2) and (G-3), and restore and fuse them step by step until the original OCT image scale is restored to obtain a denoised and enhanced OCT image.

[0053] Finally, multiple denoised and enhanced OCT images are registered and averaged to obtain the final denoised OCT image.

[0054] Multiple sheets can be 5-10 sheets.

[0055] The multi-stage OCT image denoising method of the present invention first processes the background noise and tissue structure in each image separately and specifically before averaging and fusing the images, including reducing the noise in the background area and the tissue structure area and enhancing the edge and texture of the tissue structure. Finally, the enhanced images are averaged and fused to obtain a high-quality denoised OCT image.

[0056] In order to reasonably distinguish the tissue structure and background area, the present invention first aligns a few noisy images and then averages and fuses them to obtain an average image, clusters the multiple features of the average image, and obtains the weight coefficient template of the tissue structure and background noise represented by each pixel. In the processing of background noise, the TV algorithm is combined with the above-mentioned weight coefficient template to specifically remove the background noise and retain the structural detail information in the image to the greatest extent. In the processing of tissue structure, guided filtering is performed on high-frequency images under the framework of multi-scale image decomposition to improve the noise at the edge structure, and directional filtering is performed on low-frequency images to improve the noise in the internal area of ​​the tissue and enhance the edge structure of the tissue. On this basis, the processed few images are averaged and fused to further improve the noise in the tissue structure area, and finally a high-quality denoised OCT image is obtained.

[0057] In a second aspect, the present invention also provides a computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the steps of the multi-stage OCT image denoising method based on improved clustering when executing the computer program.

[0058] In a third aspect, the present invention further provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when processed and executed, implements the steps of the multi-stage OCT image denoising method based on improved clustering.

[0059] Compared with the prior art, the present invention has the following beneficial effects:

[0060] (1) Compared with the direct average fusion method, the technical solution of the present invention only requires fewer images to obtain higher quality labeled images, and the number of OCT images required for repeated measurement can be reduced by more than 5 times;

[0061] (2) Compared with the single image enhancement method, the technical solution of the present invention can effectively reduce the image speckle noise while retaining the detailed information such as fine structure and texture in the image to a greater extent. In general, the comprehensive performance of the label image obtained by the technical solution of the present invention in image evaluation indicators such as signal-to-noise ratio, contrast, and edge preservation index is better than that of the existing solution. BRIEF DESCRIPTION OF THE DRAWINGS

[0062] Figure 1 It is a schematic diagram of the overall process of the multi-stage OCT image denoising method of the present invention.

[0063] Figure 2 It is a schematic diagram of preprocessing the OCT image to obtain the average image according to the present invention.

[0064] Figure 3 It is a schematic diagram of calculating a weight coefficient template according to clustering results of the present invention.

[0065] Figure 4 It is a schematic diagram of the process of performing directional filtering on a low-frequency image according to the present invention.

[0066] Figure 5 It is a schematic diagram of the process of restoring the image after each level of decomposition filtering in the present invention.

[0067] Figure 6 This is a comparison chart of the denoising effects of a denoising example of the present invention and other existing solutions. DETAILED DESCRIPTION

[0068] The present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be pointed out that the embodiments described below are intended to facilitate the understanding of the present invention and do not have any limiting effect on the present invention.

[0069] The overall process of the multi-stage OCT image denoising method of the present invention is as follows: Figure 1 As shown, the following steps are included:

[0070] (A) Preprocess multiple OCT images to obtain multiple registered images and a fused average image I mean .

[0071] In a specific implementation, five OCT images repeatedly measured at the same position are taken as an example, and the first image is used as a reference for image registration, and the five registered images are averagely fused. I mean like Figure 2 shown.

[0072] (B) Comprehensive use of average image I meanThe multiple features are clustered to obtain a pixel weight coefficient template, which will be used in the following steps: (1) the weight coefficient of the TV model in the background denoising process in step C; and (2) the weight coefficient when the multi-scale decomposed image is restored upward for weighted fusion in step E.

[0073] For this average image I mean Multi-scale decomposition is performed. In a specific implementation method, the multi-scale decomposition of the image takes Laplace pyramid decomposition as an example, and the image is decomposed in two levels to obtain the original average image. I mean , Gaussian image of level 1 decomposition and its corresponding Laplacian image, Gaussian image of level 2 decomposition and its corresponding Laplacian image. Upsample all images to and I mean Same pixel resolution, and calculate I mean And the gradient images of each level of Gaussian images, a total of 7 images with I mean The related feature images are expanded in the form of column vectors, that is, each pixel has 7 features. The reason for using the above features for clustering here is that the average image and its decomposition images at all levels combine the features of multiple registered images at different scales and suppress noise to a certain extent. Therefore, the use of the above features can better distinguish tissue structure and background, and can reduce the interference of noise on clustering results.

[0074] The principal component analysis is performed on the extracted features, and the first n principal component features with a cumulative contribution rate of more than 90% are retained for clustering, and the image pixels are clustered into two categories. In the specific implementation, the clustering method used is K-means clustering as an example. The class containing more structural areas in the clustering results is marked as 1, and the class containing more background areas is marked as 0. The present invention does not directly use the clustering results to rigidly divide the tissue structure and background noise, but calculates a weight coefficient template based on the difference between the principal component features of each pixel and the distance between the two class centers. w mask . Figure 3 One pixel shown p Schematic diagram of the distance between the principal component and the two class centers, the difference in distance D 0 - D 1 The larger the value, the closer the pixel is to the center of class 1 in the two classes, that is, the pixel is more likely to represent the tissue structure rather than the background noise. D 0 - D 1The smaller the value, the more likely the pixel is to represent background noise rather than tissue structure. By normalizing and gamma enhancing the distance difference, we finally get a weight coefficient template with a value range of [0,1] w mask The closer the weight coefficient is to 1, the more the pixel value represents the tissue structure information, and the closer it is to 0, the more the pixel value represents the background noise information.

[0075] The above weight coefficient template is calculated based on the clustering results w mask The specific steps are:

[0076] (1) Calculate the distance between the principal component features of the image pixels and the two cluster centers respectively D 0 (distance from cluster center 0) and D 1 (distance from cluster center 1), distance is L 2 Norm to calculate; calculate the difference in pixel distance Δ d = D 0 - D 1 ;

[0077] (2) Calculate Δ d The relative minimum value b,

[0078] (1)

[0079] (3) d Normalized to the interval [0-1];

[0080] (2)

[0081] (4) Perform gamma enhancement to obtain the final weight coefficient template:

[0082] (3)

[0083] in k is the gamma enhancement coefficient, which is set to a constant greater than 1 (such as 12). The advantage of the above weight coefficient template is that in the subsequent processing, different degrees of denoising can be performed on the tissue structure, background noise and weak boundaries in the image, and no obvious boundaries will be generated in the boundary area of ​​the structure, thereby achieving a good balance between denoising and detail preservation.

[0084] (C) For each registered image, the weight coefficient template is combined to reduce background noise through the improved TV model.

[0085] The weight coefficient template is first used to denoise the image background. The denoising process is based on the following TV minimization model:

[0086] (4)

[0087] The first term in formula (4) is the fidelity term, where I 0 ( i,j ) is the original noisy image, I ( i,j ) is the denoised image; the second term is the total variation term, Representing images I ( i,j )The sum of the absolute values ​​of the gradients in the horizontal and vertical directions; is the weight coefficient of the total variation term, C is a constant greater than 1 (for example, 15); i , j Represent the row and column indices of the image pixels, respectively. M , N Indicates the total number of rows and columns of pixels.

[0088] In this scheme, the TV item of background pixels is assigned a larger weight coefficient to better remove background noise, while the TV item of tissue structure pixels will be assigned a smaller weight coefficient, that is, the weight of the fidelity item is enhanced to retain the original structural details to a greater extent.

[0089] (D) Perform multi-scale decomposition on each OCT image after TV denoising to obtain high-frequency and low-frequency images at each level of decomposition, so that subsequent steps (E) and (F) can be processed at different image scales to more effectively reduce noise in tissue structure areas and enhance edges and textures.

[0090] In a specific implementation, the Laplacian pyramid decomposition of an image is taken as an example, and the image is decomposed into two levels to obtain a Gaussian image G of the original image decomposed into level 1. 1 And the corresponding Laplace image L 0 , the Gaussian image G of the 2nd level decomposition 2 and the corresponding Laplacian image L 1 .

[0091] (E) Guided filtering is performed on the high-frequency images at each level of decomposition to improve the noise at the tissue edge structure in the high-frequency images at each level of multi-scale decomposition.

[0092] For each level of decomposition of the Laplacian image, a 3×3 template Gaussian filter is first used to smooth it, and the image after Gaussian smoothing is used as a reference image to guide the Laplacian image filtering. The above steps can improve the noise of the tissue edge structure in the high-frequency Laplacian image.

[0093] (F) Directional filtering is performed on the low-frequency images at each level of decomposition to reduce the noise in the tissue structure area in the low-frequency images at each level of multi-scale decomposition and enhance the edges and textures of the structure.

[0094] Directional filtering is performed on the Gaussian image at each level of decomposition. The specific process of directional filtering is as follows: Figure 4 First, a directional edge detection operator (such as Sobel operator and Prewitt operator) is used to determine the main direction of each pixel. The main directions include horizontal, vertical, 45 degrees oblique and 135 degrees oblique. The four directions are marked as 1, 2, 3 and 4 respectively. Figure 4 The dotted boxes are the Sobel detection operators corresponding to the four directions. The corresponding operators are convolved with the 3×3 neighborhood of the central pixel to obtain the gradient of the central pixel along the four directions, and the direction with the largest absolute value of the gradient is taken as the main direction of the central pixel.

[0095] The above direction detection will determine a main direction for each pixel, but in fact only the edge and texture structure of the tissue have obvious directionality and need to be filtered. Noise is non-directional and should not be filtered. Therefore, it is necessary to first determine whether the central pixel has an obvious main direction. The variance of the main direction of the pixels in the neighborhood can be used to determine. The smaller the variance, the more concentrated the main direction in the neighborhood, so the central pixel is more likely to have an obvious main direction; conversely, the central pixel is more likely to have no obvious main direction. A variance threshold of 0.83 can be set. The basis for setting it is: when the main direction is completely disordered, the probability of the four main directions appearing is equal, that is, the main direction values ​​1, 2, 3, and 4 are evenly distributed, and the corresponding variance is 1.67. The variance threshold is set to half of this value, indicating that there is a certain order in the main direction. The variance of the main direction of the neighborhood (for example, a 5×5 neighborhood) around the central pixel is compared with a preset variance threshold. If the variance is less than the preset threshold, it is determined that the central pixel has an obvious main direction, and a line neighborhood is taken along a direction perpendicular to its main direction for spatial filtering (for example, but not limited to 1×5 line neighborhood Gaussian filtering); if the variance is greater than the preset threshold, it is determined that the central pixel does not have an obvious main direction, and isotropic filtering (for example, but not limited to 5×5 neighborhood Gaussian filtering) is directly performed in the neighborhood.

[0096] The directional filtering method can bring two benefits: first, the main direction judgment can accurately extract the directionality of the pixels of the tissue edge structure, and then through directional filtering, the noise in the tissue structure area can be effectively reduced and the edges and textures can be enhanced; second, through variance threshold judgment, directional filtering of background noise can be effectively avoided to prevent the texture structure from causing abnormal images.

[0097] (G) The decomposed images at each level are restored step by step and weightedly fused with the low-frequency images at the previous level.

[0098] Figure 5 The figure shows the process of restoring the image after each level of decomposition filtering.

[0099] Starting from the highest level decomposition image, the guided filtered Laplacian image and the directional filtered Gaussian image are restored to the previous level step by step, and each level of restored image is weighted fused with the original Gaussian image until the original image resolution is restored. The weight coefficient of weighted fusion is the weight coefficient template generated by pixel clustering. w mask , downsample it to G K The same pixel resolution, expressed as .

[0100] (G-1) k Taking the decomposed image of the highest level (Kth level) as an example, starting from the decomposed image of the highest level (Kth level), the denoised low-frequency image and the corresponding high-frequency image are restored to the next level (K-1th level) to obtain ;

[0101] (G-2) Weighted fusion of the K-1th level restored image and the denoised original K-1th level decomposed low-frequency image:

[0102] (5)

[0103] In the formula, G K-1 is the original image after directional filtering. K -1 level decomposition low frequency image, For the K -1 level restore image, After fusion K -1 level low frequency image, The weight coefficient template generated in step (B) w mask Downsample to and G K-1 Same pixel resolution;

[0104] (G-3) The fused image Yohara K The high-frequency image of the -1 level decomposition is restored to the next level (K-2 level) ;

[0105] (G-4) Repeat steps (G-2) and (G-3) to restore and fuse the images step by step until the original image scale is restored to obtain a denoised and enhanced OCT image.

[0106] (H) The restored images are averaged and fused to obtain the final denoised image.

[0107] Finally, the five denoised and enhanced OCT images were averaged and fused to obtain the final denoised image.

[0108] Figure 6 The figure shows a denoising example of the present invention and a comparison of the denoising effect with other existing solutions. This example uses fundus OCT images measured continuously at the same position. The present invention uses 5 measured images for denoising. The denoising methods compared are the registration average method (50 measured images), non-local means filtering (NLM, References: J. Aum, JHKim, and J. Jeong. Effective speckle noise suppression in optical coherencetomography images using nonlocal means denoising filter with double Gaussiananisotropic kernels, Appl. Opt. , 54(13), 43–50, 2015.), three-dimensional block matching filter (BM3D, Reference: B. Chong and YK Zhu. Speckle reduction in optical coherencetomography images of human finger skin by wavelet modified BM3D filter, Opt. Commun. , 291, 461–469, 2013.), and the deep learning method proposed by Xu et al. (OCTNet, Reference: M.Xu, C. Tang, F. Hao, et al. Texture preservation and speckle reduction in poor optical coherence tomography using the convolutional neural network, Med. Image Anal., 64, 101727, 2020.). From the results, the present invention can better achieve effective noise reduction while retaining structural details to the greatest extent, such as Figure 6The details of the local area marked by the arrow are shown in the enlarged figure. As can be seen from Table 1, the denoised image obtained by the present invention is optimal in terms of the three indicators of signal-to-noise ratio (SNR), contrast (CNR) and equivalent number of views (ENL). In terms of edge preservation index (EPI), although the denoised image obtained by the NLM method has the highest EPI index, most of the noise in the image is not effectively removed. The EPI index of the present invention is only slightly lower than the average effect of 50 registrations, but the present invention only needs to use 5 noisy images, which greatly reduces the requirements for measurement and registration. Therefore, in general, the present invention is better than the existing scheme.

[0109] Table 1 Comparison of image evaluation indicators of the present invention and other denoising schemes

[0110] method SNR CNR ENL EPI NLM 25.26±0.55 3.94±0.65 39.45±10.38 0.585±0.008 BM3D 34.51±2.83 6.55±0.67 42.56±16.41 0.165±0.023 OCTNet 41.41±3.01 8.56±0.64 96.89±25.91 0.143±0.021 Average (50 photos) 34.44±0.31 6.41±0.96 67.70±15.16 0.301±0.001 The present invention 42.10±1.27 8.78±0.98 119.9±20.82 0.263±0.001

[0111] The embodiments described above provide a detailed description of the technical solutions and beneficial effects of the present invention. It should be understood that the above are only specific embodiments of the present invention and are not intended to limit the present invention. Any modifications, supplements and equivalent substitutions made within the scope of the principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A multi-stage OCT image denoising method based on improved clustering, characterized in that: include: (1) Multiple original OCT images obtained by repeated scanning of the same location are registered and averaged to obtain an average image I mean ; (2) Using the average image I mean Cluster multiple features to obtain the pixel weight coefficient template w mask ; include: (2-1) For the average image I mean Perform multi-scale decomposition to obtain the decomposed low-frequency images and high-frequency images at all levels, upsample the low-frequency images and high-frequency images at all levels to the average image I mean Same resolution; calculate average image I mean and gradient images of various levels of high-frequency images; (2-2) Extracting the average image I mean , image features of low-frequency images and high-frequency images at all levels, and gradient images; (2-3) Perform pixel principal component analysis on the extracted image features, retain the first n pixel principal component features according to the cumulative contribution rate and perform clustering, cluster the image pixels into two categories: structure area and background area, and calculate the weight coefficient template based on the clustering results w mask ; Calculate weight coefficient template based on clustering results w mask include: (i) Calculate the distance between the principal component features of the image pixels and the class center of the background area respectively D 0. Distance from the class center of the structure area D 1; (ii) Calculation of Δ d = D 0 - D 1, and calculate Δ d The relative minimum value b: (1) (iii) Δ d Normalized to the interval [0-1]: (2) (iv) Perform gamma enhancement to obtain the weight coefficient template w mask : (3) in, k is the gamma enhancement coefficient, which is a constant greater than 1; (3) Combined with weight coefficient template w mask , the background of each original OCT image is denoised by the TV model; (4) Perform multi-scale decomposition on each original OCT image after background denoising to obtain high-frequency images and low-frequency images at each level of decomposition; (5) Perform guided filtering on the high-frequency images at each level of decomposition, and perform directional filtering on the low-frequency images at each level of decomposition; (6) For each original OCT image, the decomposed images at each level are restored step by step and weighted fused with the low-frequency image at the previous level to obtain a denoised and enhanced OCT image; (7) Multiple denoised and enhanced OCT images are registered and averaged to obtain the final denoised OCT image.

2. The multi-stage OCT image denoising method based on improved clustering according to claim 1, characterized in that: Step (3) includes: denoising the background of each original OCT image based on the minimized TV model: (4) The first term is the fidelity term, where I 0 ( i,j ) is the original noisy image, I ( i,j ) is the denoised image; the second term is the total variation term, Representing images I ( i,j )The sum of the absolute values ​​of the gradients in the horizontal and vertical directions; is the weight coefficient of the total variation term, C is a constant greater than 1; i , j Represent the row and column indices of the image pixels, respectively. M , N Indicates the total number of rows and columns of pixels.

3. The multi-stage OCT image denoising method based on improved clustering according to claim 1, characterized in that: In step (5), guided filtering is performed on the high-frequency images at each level of decomposition, including: Gaussian filtering and smoothing the high-frequency images, and using the images smoothed by Gaussian filtering as reference images to perform guided filtering on the high-frequency images.

4. The multi-stage OCT image denoising method based on improved clustering according to claim 1, characterized in that: In step (5), directional filtering is performed on the low-frequency images at each level of decomposition, including: (1) Using a directional edge detection operator, the gradients of each pixel in the low-frequency image along the horizontal, vertical, 45-degree and 135-degree directions are extracted, and the direction with the largest absolute value of the gradient is taken as the main direction of the pixel; (2) Select the neighborhood around the pixel, calculate the variance of the main direction of the pixels in the neighborhood, and compare it with the preset threshold: If the variance is less than the preset threshold, it is determined that the central pixel has an obvious main direction, and the line neighborhood is taken along the direction perpendicular to its main direction for spatial filtering; If the variance is greater than a preset threshold, it is determined that the central pixel does not have an obvious main direction, and isotropic filtering is performed directly in the neighborhood.

5. The multi-stage OCT image denoising method based on improved clustering according to claim 4, characterized in that: The directional edge detection operator is a Sobel operator or a Prewitt operator.

6. The multi-stage OCT image denoising method based on improved clustering according to claim 1, characterized in that: Step (6) includes: (6-1) Starting from the highest level K-level decomposition image, restore the denoised low-frequency image and the corresponding high-frequency image to the next level K-1 level to obtain K -1 level restored image ; (6-2) The weight coefficient template w mask Downsample to and G K-1 Same pixel resolution ;Will And the low-frequency image of the K-1th level decomposition after directional filtering G K-1 Perform weighted fusion to obtain the fused K -Level 1 low frequency image : (5) (6-3) and K The high-frequency image decomposed at level -1 is restored to the next level, level K-2. ; (6-4) Repeat steps (6-2) and (6-3), and restore and fuse them step by step until the original OCT image scale is restored to obtain a denoised and enhanced OCT image.

7. A computer device, characterized in that: The method comprises a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, the steps of the multi-stage OCT image denoising method based on improved clustering as described in any one of claims 1 to 6 are implemented.

8. A computer-readable storage medium, characterized in that: A computer program is stored thereon, and when the computer program is processed and executed, the steps of the multi-stage OCT image denoising method based on improved clustering according to any one of claims 1 to 6 are implemented.

Images