Nuclear medicine PET image automatic segmentation method and system based on deep learning

Abstract

The invention relates to the technical field of nuclear medicine PET image segmentation, in particular to a nuclear medicine PET image automatic segmentation method and system based on deep learning. The method comprises the following steps: obtaining a gray scale connected domain according to gray scale distribution in a PET image; matching the gray scale connected domains of different PET images to obtain a matched connected domain of each gray scale connected domain; obtaining a position correlation index and an organization correlation index according to the position difference and the pixel point distribution of the two; the image matching degree is obtained by combining the number of pixel points of the gray level connected domain and gray level distribution; denoising by combining the noise distribution of the PET image to obtain a denoised PET image; and segmenting all the denoised PET images. According to the method, a good denoising effect can be generated for the nuclear medicine PET image, and then the nuclear medicine PET image is accurately segmented.

Description

Technical Field

[0001] The present invention relates to the technical field of nuclear medicine PET image segmentation, and in particular to a method and system for automatic nuclear medicine PET image segmentation based on deep learning. Background Art

[0002] Nuclear medicine PET images are medical images obtained using positron emission tomography (PET). PET (Positron Emission Tomography) is a molecular imaging technology that monitors and displays biomolecular metabolic processes within the human body. It provides information on tissue and organ function by detecting the distribution and metabolic activity of radioactively labeled bioactive molecules within the body. PET offers advantages in nuclear medicine, such as early detection sensitivity, unmatched by other diagnostic devices. However, due to limitations of the drugs and their principles, image accuracy is limited. Therefore, when it comes to segmenting target areas, manual segmentation is often necessary, which is time-consuming, inaccurate, and inefficient.

[0003] To address the inefficiency of manual segmentation, automatic segmentation models are often used in existing technologies. However, these models place high demands on the quality of the input training data, requiring high-quality images to learn more accurate features. Because PET images are typically low-precision and noisy, they need to be denoised using a filtering algorithm before being input into the model. However, PET images are prone to generating significant Poisson noise, and the standardized uptake values ​​(SUVs) of radiotracers in different tissues can be similar, resulting in similar noise patterns in the images. In this case, the filtering algorithm is likely to identify similar noisy image patches during filtering, resulting in poor denoising results. The algorithm may even intensify the noise or introduce new noise, ultimately reducing the performance of the automatic segmentation model and increasing the risk of overfitting, making it impossible to achieve optimal PET image segmentation results. Summary of the Invention

[0004] In order to solve the technical problem that different nuclear medicine PET images may have similar noise patterns, which may lead to poor denoising effect when filtering them using filtering algorithms, or even intensify the noise or introduce new noise, ultimately reducing the performance of the automatic segmentation model and increasing the risk of model overfitting, making it impossible to achieve good PET image segmentation results, the purpose of the present invention is to provide a nuclear medicine PET image automatic segmentation method and system based on deep learning. The technical solutions adopted are as follows:

[0005] A deep learning-based automatic segmentation method for nuclear medicine PET images, comprising:

[0006] Acquire all nuclear medicine PET images for each patient;

[0007] A nuclear medicine PET image is selected as a reference PET image; all grayscale connected domains in the reference PET image are obtained based on the grayscale distribution of all pixels in the reference PET image; a grayscale connected domain in the reference PET image is selected as a reference connected domain; a matching connected domain of the reference connected domain in each other nuclear medicine PET image is obtained based on the position distribution difference between the reference connected domain and each grayscale connected domain in the other nuclear medicine PET images; a position correlation index between the reference connected domain and each matching connected domain is obtained based on the position distribution difference and pixel distribution characteristics between the reference connected domain and each matching connected domain; a tissue correlation index between the reference connected domain and each matching connected domain is obtained based on the position correlation index and position distribution difference between a preset neighborhood of the reference connected domain and a preset neighborhood of each matching connected domain;

[0008] Obtaining a degree of image matching between the reference PET image and each other nuclear medicine PET image based on the position correlation index and the tissue correlation index between the reference connected domain and the matching connected domain in each nuclear medicine PET image, as well as the number of pixels and grayscale distribution of each grayscale connected domain;

[0009] The reference PET image is denoised according to the image matching degree and the noise distribution to obtain a denoised PET image corresponding to the reference PET image; and all denoised PET images are segmented.

[0010] Furthermore, the method for obtaining the grayscale connected domain includes:

[0011] Obtain all connected domains in the reference PET image according to the image segmentation algorithm;

[0012] Obtaining a grayscale distribution histogram of a reference PET image; taking a grayscale interval between a minimum grayscale value and a grayscale median value as a first grayscale interval; taking a grayscale interval between a grayscale median value and a grayscale maximum value as a second grayscale interval;

[0013] Finding the grayscale value corresponding to the grayscale level with the least number of pixels in the first grayscale interval as the first boundary grayscale value, and finding the grayscale value corresponding to the grayscale level with the least number of pixels in the second grayscale interval as the second boundary grayscale value;

[0014] The area where the pixels corresponding to all gray levels between the minimum gray value and the first demarcation gray value are located is defined as a high metabolic activity area; the area where the pixels corresponding to all gray levels between the first demarcation gray value and the second demarcation gray value are located is defined as a medium metabolic activity area; and the area where the pixels corresponding to all gray levels between the second demarcation gray value and the gray maximum value are located is defined as a low metabolic activity area.

[0015] All connected domains in the high metabolic activity area are regarded as high metabolic grayscale connected domains; all connected domains in the medium metabolic activity area are regarded as medium metabolic grayscale connected domains, and all connected domains in the low metabolic activity area are regarded as low metabolic grayscale connected domains; the high metabolic grayscale connected domains, the medium metabolic grayscale connected domains and the low metabolic grayscale connected domains are regarded as grayscale connected domains.

[0016] Furthermore, the method for obtaining the matching connected domain includes:

[0017] In other nuclear medicine PET images, the grayscale connected domain that is closest to the corresponding position of the reference connected domain and is in the same metabolic activity area as the reference connected domain is selected as the matching connected domain of the reference connected domain in each other nuclear medicine PET image.

[0018] Furthermore, the method for obtaining the location correlation index includes:

[0019] The pixel point corresponding to the coordinate mean of each pixel point in each grayscale connected domain is taken as the representative pixel point of the grayscale connected domain;

[0020] Obtaining the corresponding pixel points of the representative pixel points of the reference connected domain in each other nuclear medicine PET image as the pixel points to be compared;

[0021] The position correlation index is obtained according to the position correlation index calculation formula, which is as follows:

[0022]

[0023] Where a represents the reference connected domain; b represents the serial number of the matching connected domain; represents the position correlation index between the reference connected domain and the bth matching connected domain; S ab P represents the number of pixels with the same grayscale value in the reference connected domain and the bth matching connected domain; a represents the number of pixels in the reference connected domain; P b Indicates the number of pixels in the bth matching connected domain; R ab It represents the distance between the pixel to be compared in the reference connected component and the representative pixel in the bth matching connected component; norm() represents the normalization function.

[0024] Furthermore, the method for obtaining the tissue-related index includes:

[0025] Taking a preset number of grayscale connected domains closest to the reference connected domain as preset neighborhoods of the reference connected domain, traversing all grayscale connected domains to obtain the preset neighborhoods of each grayscale connected domain;

[0026] Connect the representative pixel points of each grayscale connected domain in the preset neighborhood in a clockwise order to obtain the tissue area corresponding to the preset neighborhood;

[0027] The organization association index is obtained according to the organization association index calculation formula, which is as follows:

[0028]

[0029] Where μ represents the tissue correlation index between the reference connected domain and the bth matching connected domain; represents the position correlation index between the tissue region c corresponding to the preset neighborhood of the reference connected domain and the tissue region d corresponding to the preset neighborhood of the bth matching connected domain; L represents the number of edges in the tissue region; θ c,l θ represents the angle between the lth edge of the tissue region c corresponding to the preset neighborhood of the reference connected domain and the horizontal direction; d,l represents the angle between the lth side of the tissue region d corresponding to the preset neighborhood of the bth matching connected domain and the horizontal direction; K represents the preset number of grayscale connected domains closest to the reference connected domain contained in the preset neighborhood; R represents the position correlation index between the kth grayscale connected domain in the preset neighborhood of the reference connected domain and the kth grayscale connected domain in the preset neighborhood of the bth matching connected domain; a,k R represents the distance between the representative pixel of the reference connected domain and the representative pixel of the kth grayscale connected domain in the preset neighborhood of the bth matching connected domain; b,k represents the distance between the representative pixel of the bth matching connected domain and the representative pixel of the kth grayscale connected domain in the preset neighborhood of the reference connected domain; || represents the absolute value function.

[0030] Furthermore, the method for obtaining the image matching degree includes:

[0031] The image matching degree is obtained according to the image matching degree calculation formula, and the image matching degree calculation formula is as follows:

[0032]

[0033] Where, γ n represents the matching weight coefficient of the nth grayscale connected domain in the reference PET image; P n represents the number of pixels in the nth grayscale connected domain in the reference PET image; P0 represents the number of pixels in the reference PET image; H c Indicates the maximum grayscale value of the pixel in the reference PET image; H nrepresents the grayscale mean of the nth grayscale connected domain in the reference PET image; ω represents the degree of image matching between the reference PET image and each other nuclear medicine PET image; N represents the number of grayscale connected domains in the reference PET image; represents the positional correlation index between the nth grayscale connected domain in the reference PET image and the matching connected domain in each other nuclear medicine PET image; μ n represents the tissue correlation index between the nth grayscale connected domain in the reference PET image and the matching connected domain in each other nuclear medicine PET image; norm() represents the normalization function.

[0034] Further, denoising the reference PET image according to the image matching degree and the noise distribution to obtain a denoised PET image corresponding to the reference PET image includes:

[0035] calculating an image matching degree between the reference PET image and each other nuclear medicine PET image, and sorting the other nuclear medicine PET images in descending order according to the image matching degree to obtain a preset first number of matching images of the reference PET image;

[0036] The weight coefficient of the similar image block of each pixel of the reference PET image in each matching image is obtained according to the image matching degree and noise distribution. The calculation formula is as follows:

[0037]

[0038] Where f represents the weight coefficient of each pixel of the reference PET image in the similar image block of each matching image; ρ represents the similarity of the similar image blocks of each matching image, which can be obtained by the existing technology; ω represents the image matching degree between the reference PET image and each matching image; STP represents the entropy value of each matching image, which can be obtained by the existing technology; norm() represents the normalization function.

[0039] Each pixel of the reference PET image is filtered according to the weight coefficient of the similar image block of each pixel in each matching image to obtain a denoised PET image corresponding to the reference PET image.

[0040] A deep learning-based automatic segmentation system for nuclear medicine PET images, the system comprising a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the steps of the deep learning-based automatic segmentation method for nuclear medicine PET images are implemented.

[0041] A computer-readable storage medium stores a computer program, which, when executed by a processor, implements the steps of the above-mentioned method for automatic segmentation of nuclear medicine PET images based on deep learning.

[0042] A computer device 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 above-mentioned method for automatic segmentation of nuclear medicine PET images based on deep learning are implemented.

[0043] The present invention has the following beneficial effects:

[0044] The present invention obtains all nuclear medicine PET images of the patient; due to the different metabolic activity levels of cells in different areas, there are obvious differences in the grayscale values ​​of different metabolic levels in the nuclear medicine images, and the differences between the two types of areas are obvious. In order to distinguish the parts with different metabolic degrees in the nuclear medicine PET images, all grayscale connected domains in the reference PET images are obtained according to the grayscale distribution of all pixels in the reference PET images, and the metabolic levels of different grayscale connected domains are distinguished; since the grayscale features of the areas with high metabolic activity in different nuclear medicine PET images should be similar, the connected domains in different nuclear medicine PET images are matched; since the parts where the human body has lesions are similar, the lesion areas are different when they spread and metastasize. Obvious spatial correlations can lead to similar structural patterns in nuclear medicine PET images. Therefore, the positional correlation and tissue correlation between the reference connected domain and the matching connected domain in each other nuclear medicine PET image are analyzed to obtain positional correlation indices and tissue correlation indices. Since the grayscale distribution in nuclear medicine PET images can reflect the level of cellular metabolism in the human body, the degree of image matching between the reference PET image and each other nuclear medicine PET image is obtained by combining the number of pixels and grayscale distribution of each grayscale connected domain. Based on the image matching degree and noise distribution, the reference PET image is denoised to obtain a denoised PET image corresponding to the reference PET image. All denoised PET images are then segmented. The present invention can produce a good denoising effect on nuclear medicine PET images, thereby accurately segmenting nuclear medicine PET images. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] In order to more clearly illustrate the technical solutions and advantages of the embodiments of the present invention or the prior art, the following briefly introduces the drawings required for use in the embodiments or the prior art descriptions. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0046] Figure 1A flowchart of a method for automatic segmentation of nuclear medicine PET images based on deep learning provided by one embodiment of the present invention;

[0047] Figure 2 A nuclear medicine PET image of a human body is provided in one embodiment of the present invention. DETAILED DESCRIPTION

[0048] To further illustrate the technical means and effectiveness of the present invention in achieving its intended objectives, the following, in conjunction with the accompanying drawings and preferred embodiments, describes in detail the specific implementation, structure, features, and effectiveness of a deep learning-based nuclear medicine PET image automatic segmentation method and system proposed by the present invention. In the following description, references to different "one embodiment" or "another embodiment" do not necessarily refer to the same embodiment. Furthermore, specific features, structures, or characteristics of one or more embodiments may be combined in any suitable manner.

[0049] Unless defined otherwise, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention belongs.

[0050] The following describes in detail a specific scheme of a nuclear medicine PET image automatic segmentation method based on deep learning provided by the present invention with reference to the accompanying drawings.

[0051] See also Figure 1 , which shows a method and system for automatic segmentation of nuclear medicine PET images based on deep learning provided by one embodiment of the present invention, the method comprising:

[0052] Step S1: Acquire all nuclear medicine PET images of each patient.

[0053] The embodiment of the present invention is mainly used in application scenarios where abnormalities in the patient's body are discovered by segmenting nuclear medicine PET images of the patient at different stages. Therefore, the embodiment of the present invention obtains all nuclear medicine PET images of each patient.

[0054] In one embodiment of the present invention, a set of nuclear medicine PET images of each patient is obtained through channels such as medical research institutions or hospitals, all nuclear medicine PET images are grayscaled, and an image segmentation algorithm is used to extract the areas belonging to the human body from all the grayscaled nuclear medicine PET images to prevent irrelevant background areas from affecting subsequent operations.

[0055] Because nuclear medicine PET images may contain noise, image denoising is performed using a filtering algorithm before image segmentation. In one embodiment of the present invention, a non-local means filtering algorithm is used for denoising. It should be noted that in other embodiments of the present invention, filtering algorithms such as a bilateral filtering algorithm may also be used for image denoising. These algorithms are well known to those skilled in the art and are not limited or detailed herein.

[0056] Step S2: Select any nuclear medicine PET image as a reference PET image; obtain all grayscale connected domains in the reference PET image based on the grayscale distribution of all pixels in the reference PET image; select any grayscale connected domain in the reference PET image as a reference connected domain; obtain the matching connected domain of the reference connected domain in each other nuclear medicine PET image based on the position distribution difference between the reference connected domain and each grayscale connected domain in other nuclear medicine PET images; obtain the position correlation index between the reference connected domain and each matching connected domain based on the position distribution difference and pixel distribution characteristics between the reference connected domain and each matching connected domain; obtain the tissue correlation index between the reference connected domain and each matching connected domain based on the position correlation index and position distribution difference between the preset neighborhood of the reference connected domain and the preset neighborhood of each matching connected domain.

[0057] According to existing technology, the grayscale characteristics of the nuclear medicine PET image regions corresponding to body parts with more active cellular metabolism are more pronounced, while the grayscale values ​​of the nuclear medicine PET image regions corresponding to body parts with lower cellular metabolism are smaller. In reality, cells in tumors, inflammation, and diseased muscle regions have extremely high metabolic activity levels compared to cells in ordinary regions, while cells in non-functional regions such as fat and bone have very low metabolic activity levels compared to cells in ordinary regions. The grayscale values ​​of these two types of regions in nuclear medicine PET images are significantly different, and the boundaries between the two types of regions are clearly defined. In order to distinguish between areas with different metabolic levels in each nuclear medicine PET image, in an embodiment of the present invention, all grayscale connected domains in the reference PET image are obtained based on the grayscale distribution of all pixels in the reference PET image.

[0058] Preferably, in one embodiment of the present invention, the method for obtaining the grayscale connected component includes:

[0059] Since different parts of the body have different metabolic levels, in nuclear medicine PET images, the reference PET image is segmented according to the image segmentation algorithm to obtain all connected domains. The metabolic levels of the body parts corresponding to each connected domain are relatively consistent.

[0060] Since the grayscale difference between the image area corresponding to the part with active cell metabolism and the image area corresponding to the part with low cell metabolism is large, the difference between the histograms corresponding to different grayscale levels is large in the grayscale histogram, which can intuitively reflect the image areas with different metabolic levels. Therefore, the grayscale distribution histogram of the reference PET image is obtained.

[0061] because Figure 2 A nuclear medicine PET image of a human body is provided. In the image, it can be clearly seen that certain body parts have extremely high metabolic levels, that is, the grayscale values ​​of the image regions corresponding to such body parts are very low, while the grayscale values ​​of the image regions corresponding to fat and bone parts with very low metabolic levels are very high. The metabolic levels of normal body parts are between the above two body parts, and the grayscale values ​​of the corresponding image regions are also between the grayscale values ​​of the above two image regions. Therefore, in an embodiment of the present invention, a reference PET image is divided into a high metabolic activity region, a medium metabolic activity region, and a low metabolic activity region. The specific steps include:

[0062] The grayscale interval between the minimum grayscale value and the grayscale median is used as the first grayscale interval; the grayscale interval between the grayscale median and the grayscale maximum is used as the second grayscale interval. Because the grayscale difference between high and medium metabolic activity areas is more obvious, the number of pixels in the transition area between the two metabolic activity areas is smaller. Therefore, the grayscale value corresponding to the grayscale with the least number of pixels in the first grayscale interval is used as the first grayscale value, and the grayscale value corresponding to the grayscale with the least number of pixels in the second grayscale interval is used as the second grayscale value.

[0063] The area where the pixels corresponding to all gray levels between the minimum gray value and the first dividing gray value are located is regarded as the high metabolic activity area, the area where the pixels corresponding to all gray levels between the first dividing gray value and the second dividing gray value are located is regarded as the medium metabolic activity area; the area where the pixels corresponding to all gray levels between the second dividing gray value and the maximum gray value are located is regarded as the low metabolic activity area.

[0064] Therefore, all connected domains in the high metabolic activity area are regarded as high metabolic grayscale connected domains; all connected domains in the medium metabolic activity area are regarded as medium metabolic grayscale connected domains, and all connected domains in the low metabolic activity area are regarded as low metabolic grayscale connected domains; high metabolic grayscale connected domains, medium metabolic grayscale connected domains and low metabolic grayscale connected domains are regarded as grayscale connected domains.

[0065] The non-local means filtering algorithm requires analyzing the similarities between different nuclear medicine PET images. Since inflammation and lesions are often concentrated in certain organs and tissues, where cellular metabolism is high, the grayscale characteristics of the corresponding regions in different nuclear medicine PET images should be similar. Therefore, the connected domains in different nuclear medicine PET images are matched to obtain the matching connected domains of the reference connected domain in each other nuclear medicine PET image. Since lesions in the human body exhibit significant spatial correlations during diffusion and metastasis, resulting in similar structural patterns in nuclear medicine PET images, the positional and tissue correlations between the reference connected domain and the matching connected domains in each other nuclear medicine PET image are analyzed to obtain positional and tissue correlation indices.

[0066] Preferably, in one embodiment of the present invention, the method for obtaining the matching connected domain includes:

[0067] In other nuclear medicine PET images, the grayscale connected domain that is closest to the corresponding position of the reference connected domain and is in the same metabolic activity area as the reference connected domain is selected as the matching connected domain of the reference connected domain in each other nuclear medicine PET image.

[0068] Preferably, in one embodiment of the present invention, the method for obtaining the location correlation index includes:

[0069] The pixel point corresponding to the coordinate mean of each pixel point in each grayscale connected domain is taken as the representative pixel point of the grayscale connected domain;

[0070] Obtaining the corresponding pixel points of the representative pixel points of the reference connected domain in each other nuclear medicine PET image as the pixel points to be compared;

[0071] The location correlation index is obtained according to the location correlation index calculation formula. The location correlation index calculation formula is as follows:

[0072]

[0073] Where a represents the reference connected domain; b represents the serial number of the matching connected domain; represents the position correlation index between the reference connected domain and the bth matching connected domain; S ab P represents the number of pixels with the same grayscale value in the reference connected domain and the bth matching connected domain; a represents the number of pixels in the reference connected domain; P b Indicates the number of pixels in the bth matching connected domain; R ab It represents the distance between the pixel to be compared in the reference connected component and the representative pixel in the bth matching connected component; norm() represents the normalization function.

[0074] In the position correlation index calculation formula, the number of pixels S with the same gray value between the reference connected domain and the bth matching connected domain is ab The more there are and the smaller the total size of the reference connected domain and the bth matching connected domain is, it means that the difference in the number of pixels between the reference connected domain and the bth matching connected domain is small and the corresponding body parts are more consistent, that is, the positional association is stronger; the smaller the distance between the pixel point to be compared in the reference connected domain and the representative pixel point of the bth matching connected domain is, it means that in the same nuclear medicine PET image, the distance between the corresponding position of the reference connected domain and the position of the bth matching connected domain is closer, and the positional association is greater.

[0075] Preferably, in one embodiment of the present invention, the method for obtaining the organization-related index includes:

[0076] A preset number of grayscale connected domains closest to the reference connected domain are used as preset neighborhoods of the reference connected domain, and all grayscale connected domains are traversed to obtain the preset neighborhoods of each grayscale connected domain. In one embodiment of the present invention, the preset number is set to 4. It should be noted that in other embodiments of the present invention, the preset number can be set arbitrarily and is not limited here.

[0077] The representative pixel points of each grayscale connected domain in the preset neighborhood are connected in clockwise order to obtain the tissue area corresponding to the preset neighborhood.

[0078] The organization-related index is obtained according to the organization-related index calculation formula. The organization-related index calculation formula is as follows:

[0079]

[0080] Where μ represents the tissue correlation index between the reference connected domain and the bth matching connected domain; represents the position correlation index between the tissue region c corresponding to the preset neighborhood of the reference connected domain and the tissue region d corresponding to the preset neighborhood of the bth matching connected domain; L represents the number of edges in the tissue region; θ c,l θ represents the angle between the lth edge of the tissue region c corresponding to the preset neighborhood of the reference connected domain and the horizontal direction; d,l represents the angle between the lth side of the tissue region d corresponding to the preset neighborhood of the bth matching connected domain and the horizontal direction; K represents the preset number of grayscale connected domains closest to the reference connected domain contained in the preset neighborhood; R represents the position correlation index between the kth grayscale connected domain in the preset neighborhood of the reference connected domain and the kth grayscale connected domain in the preset neighborhood of the bth matching connected domain; a,k R represents the distance between the representative pixel of the reference connected domain and the representative pixel of the kth grayscale connected domain in the preset neighborhood of the bth matching connected domain;b,k represents the distance between the representative pixel of the bth matching connected domain and the representative pixel of the kth grayscale connected domain in the preset neighborhood of the reference connected domain; || represents the absolute value function.

[0081] In the tissue correlation index calculation formula, the larger the position correlation index between the tissue region of the reference connected domain and the tissue region of the bth matching connected domain, the more likely it is that the reference connected domain and the bth matching connected domain belong to the same body tissue part, that is, the greater the tissue correlation between the reference connected domain and the bth matching connected domain, where, Yes Correction of angle difference |θ c,l -θ d,l The smaller the average value of |, the smaller the position difference between the reference connected domain and the b matching connected domains. In this case, the organizational association between the reference connected domain and the b matching connected domains is stronger. The larger the value is, the stronger the positional association between each grayscale connected domain in the preset neighborhood of the reference connected domain and each grayscale connected domain in the preset neighborhood of the bth matching connected domain is, and the greater the organizational association between the reference connected domain and the bth matching connected domain is; the distance R between the reference connected domain and each grayscale connected domain in the preset neighborhood of the bth matching connected domain is a,k Recorded as the first distance, the distance R between the bth matching connected domain and each grayscale connected domain in the preset neighborhood of the reference connected domain is b,k The difference between the first distance and the second distance is recorded as the second distance |R a,k -R b,k The smaller | is, the closer the distribution of the reference connected domain is to the preset neighborhood of the b-th matching connected domain. In this case, the organizational association between the reference connected domain and the b-th matching connected domain is.

[0082] Step S3: Obtain the image matching degree between the reference PET image and each other nuclear medicine PET image based on the position correlation index and tissue correlation index between the reference connected domain and the matching connected domain in each nuclear medicine PET image, as well as the number of pixels and grayscale distribution of each grayscale connected domain.

[0083] After obtaining the positional correlation index and tissue correlation index between the grayscale connected domains of different nuclear medicine PET images, all grayscale connected domains in any two nuclear medicine PET images are analyzed to obtain the degree of image matching between the two nuclear medicine PET images. Denoising is then performed using the other nuclear medicine PET image that most closely matches the reference PET image. Because the grayscale distribution in a nuclear medicine PET image can reflect the level of cellular metabolism in the human body, the grayscale mean of each grayscale connected domain is analyzed. Therefore, in this embodiment of the present invention, the degree of image matching between the reference PET image and each other nuclear medicine PET image is obtained based on the positional correlation index and tissue correlation index between the reference connected domain and the matching connected domain in each nuclear medicine PET image, as well as the number of pixels and grayscale distribution in each grayscale connected domain.

[0084] Preferably, in one embodiment of the present invention, the method for obtaining the image matching degree includes:

[0085] The image matching degree is obtained according to the image matching degree calculation formula. The image matching degree calculation formula is as follows:

[0086]

[0087] Where, γ n represents the matching weight coefficient of the nth grayscale connected domain in the reference PET image; P n represents the number of pixels in the nth grayscale connected domain in the reference PET image; P0 represents the number of pixels in the reference PET image; H c Indicates the maximum grayscale value of the pixel in the reference PET image; H n represents the grayscale mean of the nth grayscale connected domain in the reference PET image; ω represents the degree of image matching between the reference PET image and each other nuclear medicine PET image; N represents the number of grayscale connected domains in the reference PET image; represents the positional correlation index between the nth grayscale connected domain in the reference PET image and the matching connected domain in each other nuclear medicine PET image; μ n represents the tissue correlation index between the nth grayscale connected domain in the reference PET image and the matching connected domain in each other nuclear medicine PET image; norm() represents the normalization function.

[0088] In the image matching degree calculation formula, the more pixels there are in the nth grayscale connected domain, The larger the value, the more important the nth grayscale connected domain is in the reference PET image, and the relative grayscale value of the nth grayscale connected domain in the reference PET image The smaller it is, the higher the cell metabolism level of the nth grayscale connected domain is. At this time, the nth grayscale connected domain is more important, so Normalization processing is performed to obtain the matching weight coefficient of the nth grayscale connected domain in the reference PET image. The sum of the position correlation index and tissue correlation index between the nth grayscale connected domain in the reference PET image and its matching connected domain in any nuclear medicine PET image is multiplied by the matching weight coefficient of the nth grayscale connected domain to obtain the matching degree between the nth grayscale connected domain in the reference PET image and the matching connected domain. The matching degree between each grayscale connected domain in the reference PET image and its matching connected domain is accumulated to obtain the image matching degree between the reference PET image and each other nuclear medicine PET image.

[0089] Step S4: Denoise the reference PET image based on the image matching degree and noise distribution to obtain a denoised PET image corresponding to the reference PET image; and segment all denoised PET images.

[0090] Preferably, in one embodiment of the present invention, denoising the reference PET image based on the image matching degree and the noise distribution to obtain a denoised PET image corresponding to the reference PET image includes:

[0091] The degree of image matching between the reference PET image and each other nuclear medicine PET image is calculated, and the other nuclear medicine PET images are sorted in descending order according to the degree of image matching, thereby obtaining a preset first number of matching images for the reference PET image. In one embodiment of the present invention, the preset first number is set to 10. It should be noted that the preset first number can be set arbitrarily and is not limited herein.

[0092] The weight coefficient of each matching image in the image filtering algorithm is obtained according to the image matching degree and noise distribution. The calculation formula is as follows:

[0093]

[0094] Where f represents the weight coefficient of each pixel of the reference PET image in the similar image block of each matching image; ρ represents the similarity of the similar image blocks of each matching image, which can be obtained by the existing technology; ω represents the image matching degree between the reference PET image and each matching image; STP represents the entropy value of each matching image, which can be obtained by the existing technology; norm() represents the normalization function.

[0095] In the weight coefficient calculation formula, the higher the similarity and image matching degree, the greater the filtering weight; the greater the entropy value in the matching image, the more noise in the matching image, and the smaller the filtering weight should be.

[0096] Non-local means filtering is performed on each pixel of the reference PET image based on its weight coefficient in a similar image block in each matching image, thereby obtaining a denoised PET image corresponding to the reference PET image. It should be noted that the non-local means filtering algorithm is well known to those skilled in the art and will not be described in detail here.

[0097] At this point, all nuclear medicine PET images are denoised to obtain all denoised PET images.

[0098] In one embodiment of the present invention, a U-Net model is trained using all denoised PET images, and each denoised PET image is segmented using the U-Net model to obtain accurate segmentation results.

[0099] At this point, the automatic segmentation of nuclear medicine PET images is completed.

[0100] In summary, all nuclear medicine PET images of each patient are obtained; one nuclear medicine PET image is selected as a reference PET image; according to the grayscale distribution of all pixels in the reference PET image, all grayscale connected domains in the reference PET image are obtained; one grayscale connected domain in the reference PET image is selected as a reference connected domain; according to the position distribution difference between the reference connected domain and each grayscale connected domain of other nuclear medicine PET images, the matching connected domain of the reference connected domain in each other nuclear medicine PET image is obtained; according to the position distribution difference between the reference connected domain and each matching connected domain and the pixel distribution characteristics, the position relationship between the reference connected domain and each matching connected domain is obtained. The method comprises the following steps: obtaining a tissue association index between the reference connected domain and each matching connected domain based on the position association index and position distribution difference between the preset neighborhood of the reference connected domain and the preset neighborhood of each matching connected domain; obtaining the image matching degree between the reference PET image and each other nuclear medicine PET image based on the position association index and tissue association index between the reference connected domain and the matching connected domain in each nuclear medicine PET image, as well as the number of pixels and grayscale distribution of each grayscale connected domain; denoising the reference PET image based on the image matching degree and noise distribution to obtain a denoised PET image corresponding to the reference PET image; and segmenting all denoised PET images.

[0101] One embodiment of the present invention provides a nuclear medicine PET image automatic segmentation system based on deep learning, which includes a memory, a processor and a computer program, wherein the memory is used to store the corresponding computer program, and the processor is used to run the corresponding computer program. When the computer program runs in the processor, it can implement the method described in steps S1-S4.

[0102] The third purpose of an embodiment of the present invention is to provide a computer device, including a memory, a processor, and a computer program stored in the memory and runnable on the processor. When the processor executes the computer program, the steps of the above-mentioned method for automatic segmentation of nuclear medicine PET images based on deep learning are implemented.

[0103] The fourth object of an embodiment of the present invention is to provide a computer-readable storage medium, which stores a computer program. When the computer program is executed by a processor, it implements the steps of the above-mentioned method for automatic segmentation of nuclear medicine PET images based on deep learning.

[0104] It should be noted that the order in which the embodiments of the present invention are described above is for illustrative purposes only and does not necessarily represent the superiority or inferiority of the embodiments. The processes depicted in the accompanying drawings do not necessarily require the specific order or sequential order shown to achieve the desired results. In certain embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0105] The various embodiments in this specification are described in a progressive manner, and the same or similar parts between the various embodiments can be referred to each other. Each embodiment focuses on the differences from other embodiments.

Claims

1. A method for automatic segmentation of nuclear medicine PET images based on deep learning, characterized in that: The method comprises: Acquire all nuclear medicine PET images for each patient; A nuclear medicine PET image is selected as a reference PET image; all grayscale connected domains in the reference PET image are obtained based on the grayscale distribution of all pixels in the reference PET image; a grayscale connected domain in the reference PET image is selected as a reference connected domain; a matching connected domain of the reference connected domain in each other nuclear medicine PET image is obtained based on the position distribution difference between the reference connected domain and each grayscale connected domain in the other nuclear medicine PET images; a position correlation index between the reference connected domain and each matching connected domain is obtained based on the position distribution difference and pixel distribution characteristics between the reference connected domain and each matching connected domain; a tissue correlation index between the reference connected domain and each matching connected domain is obtained based on the position correlation index and position distribution difference between a preset neighborhood of the reference connected domain and a preset neighborhood of each matching connected domain; Obtaining a degree of image matching between the reference PET image and each other nuclear medicine PET image based on the position correlation index and the tissue correlation index between the reference connected domain and the matching connected domain in each nuclear medicine PET image, as well as the number of pixels and grayscale distribution of each grayscale connected domain; The reference PET image is denoised according to the image matching degree and the noise distribution to obtain a denoised PET image corresponding to the reference PET image; and all denoised PET images are segmented.

2. The method for automatic segmentation of nuclear medicine PET images based on deep learning according to claim 1, characterized in that: The method for obtaining the grayscale connected domain includes: Obtain all connected domains in the reference PET image according to the image segmentation algorithm; Obtaining a grayscale distribution histogram of a reference PET image; taking a grayscale interval between a minimum grayscale value and a grayscale median value as a first grayscale interval; taking a grayscale interval between a grayscale median value and a grayscale maximum value as a second grayscale interval; Finding the grayscale value corresponding to the grayscale level with the least number of pixels in the first grayscale interval as the first boundary grayscale value, and finding the grayscale value corresponding to the grayscale level with the least number of pixels in the second grayscale interval as the second boundary grayscale value; The area where the pixels corresponding to all gray levels between the minimum gray value and the first demarcation gray value are located is defined as a high metabolic activity area; the area where the pixels corresponding to all gray levels between the first demarcation gray value and the second demarcation gray value are located is defined as a medium metabolic activity area; and the area where the pixels corresponding to all gray levels between the second demarcation gray value and the gray maximum value are located is defined as a low metabolic activity area. All connected domains in the high metabolic activity area are regarded as high metabolic grayscale connected domains; all connected domains in the medium metabolic activity area are regarded as medium metabolic grayscale connected domains, and all connected domains in the low metabolic activity area are regarded as low metabolic grayscale connected domains; the high metabolic grayscale connected domains, the medium metabolic grayscale connected domains and the low metabolic grayscale connected domains are regarded as grayscale connected domains.

3. The method for automatic segmentation of nuclear medicine PET images based on deep learning according to claim 1, characterized in that: The method for obtaining the matching connected domain includes: In other nuclear medicine PET images, the grayscale connected domain that is closest to the corresponding position of the reference connected domain and is in the same metabolic activity area as the reference connected domain is selected as the matching connected domain of the reference connected domain in each other nuclear medicine PET image.

4. The method for automatic segmentation of nuclear medicine PET images based on deep learning according to claim 1, characterized in that: The method for obtaining the position correlation index includes: The pixel point corresponding to the coordinate mean of each pixel point in each grayscale connected domain is taken as the representative pixel point of the grayscale connected domain; Obtaining the corresponding pixel points of the representative pixel points of the reference connected domain in each other nuclear medicine PET image as the pixel points to be compared; The position correlation index is obtained according to the position correlation index calculation formula, which is as follows: Where a represents the reference connected domain; b represents the serial number of the matching connected domain; represents the position correlation index between the reference connected domain and the bth matching connected domain; S ab P represents the number of pixels with the same grayscale value in the reference connected domain and the bth matching connected domain; a represents the number of pixels in the reference connected domain; P b Indicates the number of pixels in the bth matching connected domain; R ab It represents the distance between the pixel to be compared in the reference connected component and the representative pixel in the bth matching connected component; norm() represents the normalization function.

5. The method for automatic segmentation of nuclear medicine PET images based on deep learning according to claim 1, characterized in that: The method for obtaining the tissue-related indicator includes: Taking a preset number of grayscale connected domains closest to the reference connected domain as preset neighborhoods of the reference connected domain, traversing all grayscale connected domains to obtain the preset neighborhoods of each grayscale connected domain; Connect the representative pixel points of each grayscale connected domain in the preset neighborhood in a clockwise order to obtain the tissue area corresponding to the preset neighborhood; The organization association index is obtained according to the organization association index calculation formula, which is as follows: Where μ represents the tissue correlation index between the reference connected domain and the bth matching connected domain; represents the position correlation index between the tissue region c corresponding to the preset neighborhood of the reference connected domain and the tissue region d corresponding to the preset neighborhood of the bth matching connected domain; L represents the number of edges in the tissue region; θ c，l θ represents the angle between the lth edge of the tissue region c corresponding to the preset neighborhood of the reference connected domain and the horizontal direction; d，l represents the angle between the lth side of the tissue region d corresponding to the preset neighborhood of the bth matching connected domain and the horizontal direction; K represents the preset number of grayscale connected domains closest to the reference connected domain contained in the preset neighborhood; R represents the position correlation index between the kth grayscale connected domain in the preset neighborhood of the reference connected domain and the kth grayscale connected domain in the preset neighborhood of the bth matching connected domain; a,k R represents the distance between the representative pixel of the reference connected domain and the representative pixel of the kth grayscale connected domain in the preset neighborhood of the bth matching connected domain; b，k represents the distance between the representative pixel of the bth matching connected domain and the representative pixel of the kth grayscale connected domain in the preset neighborhood of the reference connected domain; || represents the absolute value function.

6. The method for automatic segmentation of nuclear medicine PET images based on deep learning according to claim 1, characterized in that: The method for obtaining the image matching degree includes: The image matching degree is obtained according to the image matching degree calculation formula, and the image matching degree calculation formula is as follows: Where, γ n represents the matching weight coefficient of the nth grayscale connected domain in the reference PET image; P n represents the number of pixels in the nth grayscale connected domain in the reference PET image; P0 represents the number of pixels in the reference PET image; H c Indicates the maximum grayscale value of the pixel in the reference PET image; H n represents the grayscale mean of the nth grayscale connected domain in the reference PET image; ω represents the degree of image matching between the reference PET image and each other nuclear medicine PET image; N represents the number of grayscale connected domains in the reference PET image; represents the positional correlation index between the nth grayscale connected domain in the reference PET image and the matching connected domain in each other nuclear medicine PET image; μ n represents the tissue correlation index between the nth grayscale connected domain in the reference PET image and the matching connected domain in each other nuclear medicine PET image; norm() represents the normalization function.

7. The method for automatic segmentation of nuclear medicine PET images based on deep learning according to claim 1, characterized in that: Denoising the reference PET image according to the image matching degree and the noise distribution to obtain a denoised PET image corresponding to the reference PET image, including: calculating an image matching degree between the reference PET image and each other nuclear medicine PET image, and sorting the other nuclear medicine PET images in descending order according to the image matching degree to obtain a preset first number of matching images of the reference PET image; The weight coefficient of the similar image block of each pixel of the reference PET image in each matching image is obtained according to the image matching degree and noise distribution. The calculation formula is as follows: Where f represents the weight coefficient of each pixel of the reference PET image in the similar image block of each matching image; ρ represents the similarity of the similar image blocks of each matching image, which can be obtained by the existing technology; ω represents the image matching degree between the reference PET image and each matching image; STP represents the entropy value of each matching image, which can be obtained by the existing technology; norm() represents the normalization function. Each pixel of the reference PET image is filtered according to the weight coefficient of the similar image block of each pixel in each matching image to obtain a denoised PET image corresponding to the reference PET image.

8. A deep learning-based automatic segmentation system for nuclear medicine PET images, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the computer program, the steps of the deep learning-based automatic segmentation method for nuclear medicine PET images are implemented as described in any one of claims 1 to 7.

9. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the steps of the deep learning-based automatic segmentation method for nuclear medicine PET images are implemented.

10. A computer device comprising 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 automatic segmentation method of nuclear medicine PET images based on deep learning as described in any one of claims 1 to 7 are implemented.

Images