An Image and Label Asymmetric Normalization Method Applied to Segmentation Networks

Abstract

The present invention provides an image and label asymmetric normalization method applied to a segmentation network. The asymmetric normalization method is defined, and the key parameters of the asymmetric normalization method are optimized with the goal of reducing the average Euclidean distance of the entire sample group. Therefore, after the original image and the original label image are subjected to asymmetric normalization processing using the asymmetric normalization method, the average distance of the composed sample group is reduced, that is, the similarity between samples is higher. Thus, when training the segmentation network, the convergence speed of the segmentation network can be made faster, and the accuracy of the final segmentation network can be improved. The image and label asymmetric normalization method applied to the segmentation network of the present invention provides a reliable and better image preprocessing method for the segmentation of various medical images.

Description

Technical Field

[0001] The present invention belongs to the field of image processing, and particularly relates to an image and label asymmetric normalization method applied to a segmentation network. Background Art

[0002] As an important technical means for medical image analysis, image segmentation technology plays an important role in clinical auxiliary diagnosis. With the development of deep learning technology, it has also become very important to use images and labels to train an automatic segmentation network, and then achieve rapid automatic segmentation of clinical medical images. Currently, the original image and the label image are usually directly input into the segmentation network for training, or the original image and the label image are simply normalized to [0,1] and then input into the segmentation network for training. Such preprocessing methods make the training of the segmentation network converge slowly and affect the accuracy of the final segmentation network to a certain extent. Therefore, a new image normalization method applied to the segmentation network is needed, which can make the training of the segmentation network converge faster and the accuracy of the final segmentation network better. Summary of the Invention

[0003] To solve the above problems, an image and label asymmetric normalization method applied to a segmentation network is provided. The present invention adopts the following technical solutions:

[0004] The present invention provides an image and label asymmetric normalization method applied to a segmentation network, which is characterized in that it includes step S1 of obtaining a plurality of original images and a plurality of original label images corresponding to the plurality of original images; step S2 of using a symmetric normalization method to process the plurality of original images and the plurality of original label images respectively to obtain a plurality of normalized images and a plurality of label images, and forming a sample group by the plurality of images and the plurality of labels; step S3 of calculating the average distance of the sample group; step S4 of defining an asymmetric normalization method; step S5 of using a particle swarm optimization method to optimize the range parameters of the asymmetric normalization method with the goal of reducing the average distance; step S6 of using the asymmetric normalization method to process the original image and the original label image, wherein the asymmetric normalization method is defined as:

[0005]

[0006]

[0007] where a MN is the pixel matrix of the original image, b MN is the pixel matrix of the original label image, γ a , γ b are constants, and α and β are range parameters.

[0008] The image and label asymmetric normalization method applied to the segmentation network provided by the present invention may also have the following technical features, wherein the symmetric normalization method is a normalization algorithm.

[0009] The sizes of both the image and the label image are M×N.

[0010] The normalized image and label image are represented by the following formula:

[0011]

[0012] In the formula, Ai is the pixel matrix of the image, Bi is the pixel matrix of the label image, a is the pixel of the image, and b is the pixel of the label image.

[0013] The image and label asymmetric normalization method applied to the segmentation network provided by the present invention may also have the following technical features, wherein the average distance is the average Euclidean distance, which is calculated according to the following formula:

[0014]

[0015]

[0016]

[0017] In the formula, D i is the pixel-level Euclidean distance matrix, d is the element in the pixel-level Euclidean distance matrix, DS ij is the pixel-level Euclidean distance between the image and the label image, and L is the number of samples.

[0018] The image and label asymmetric normalization method applied to the segmentation network provided by the present invention may also have the following technical features, wherein the particle swarm takes the average Euclidean distance as the fitness function, the particles in the particle swarm are two-dimensional particles constructed based on the range parameters α and β, and the particle swarm optimization method includes the following steps: Step A1, initialize the parameters of the particle swarm, and the parameters at least include the number of iterations k; Step A2, randomly initialize the positions X H and the changing speeds V H of H particles in the particle swarm; Step A3, calculate the fitness of H particles respectively; Step A4, select the global optimal particle P g and the local optimal particle P N based on the fitness; Step A5, update the positions X g and the changing speeds V N of H particles based on the global optimal particle P H and the local optimal particle P H; Step A6, determine whether the iteration count k has been reached. If the determination is yes, proceed to step A7; if the determination is no, increment the iteration count by 1 and return to step A3. Step A7, output the particle with the minimum fitness, i.e., the optimal solution of the range parameters α and β.

[0019] The image and label asymmetric normalization method applied to the segmentation network provided by the present invention may further have the following technical feature: the parameters further include an inertia coefficient ω, an individual learning factor c1, a population learning factor c2, a local learning rate r1, and a global learning rate r2.

[0020] The image and label asymmetric normalization method applied to the segmentation network provided by the present invention may further have the following technical feature: the velocity of the particle is calculated according to the following formula:

[0021]

[0022] The position of the particle is calculated according to the following formula:

[0023]

[0024] The image and label asymmetric normalization method applied to the segmentation network provided by the present invention may further have the following technical feature: the label image is obtained by manual delineation by a clinician.

[0025] The image and label asymmetric normalization method applied to the segmentation network provided by the present invention may further have the following technical feature: the label image is obtained by an image processing method and verified by a clinician.

[0026] Function and Effect of the Invention

[0027] According to the image and label asymmetric normalization method applied to the segmentation network of the present invention, an asymmetric normalization method is defined, and the original image and the original label image are respectively normalized to different ranges, and the key parameters of the asymmetric normalization method are optimized with the goal of reducing the average Euclidean distance of multiple sample groups. Therefore, after performing asymmetric normalization processing on the original image and the original label image using this asymmetric normalization method, the average distance of the composed sample group is reduced, that is, the similarity between samples is higher. Thus, when training the segmentation network, the convergence speed of the segmentation network can be made faster, and the accuracy of the final segmentation network can be improved. Therefore, the image and label asymmetric normalization method applied to the segmentation network of the present invention improves the efficiency of segmentation network training and provides a reliable and better image preprocessing method for the segmentation of various medical images. Description of the Drawings

[0028] Figure 1It is a flowchart of the image and label asymmetric normalization method applied to the segmentation network in the embodiments of the present invention;

[0029] Figure 2 It is a flowchart of the particle swarm optimization algorithm in the embodiments of the present invention. Specific embodiments

[0030] In order to make the technical means, creative features, achieved purposes and effects of the present invention easy to understand, the image and label asymmetric normalization method applied to the segmentation network of the present invention will be specifically described below in conjunction with embodiments and drawings.

[0031] <Embodiment>

[0032] Figure 1 It is a flowchart of the image and label asymmetric normalization method applied to the segmentation network in the embodiments of the present invention.

[0033] As Figure 1 shown, the image and label asymmetric normalization method applied to the segmentation network includes the following steps:

[0034] Step S1, obtain a plurality of original images and a plurality of original label images corresponding to the plurality of original images.

[0035] In this embodiment, the original image is a medical image. The contour of the target organ in the medical image is outlined, the target organ part is marked as white, and the background part (i.e., the part outside the target organ) is marked as black. The obtained mask image is the corresponding original label image. The original label image is obtained by manual outlining by a clinician or by an image processing method of the prior art. For example, the mask image is automatically obtained through a segmentation network, and in this case, the mask image is further verified by a clinician to ensure the accuracy of the original label image.

[0036] Step S2, use the symmetric normalization method to process the plurality of original images and the plurality of original label images respectively to obtain a plurality of normalized images and a plurality of label images, and form a sample group from the plurality of images and the plurality of labels.

[0037] In this embodiment, the sizes of the original image and the original label image are the same. For the convenience of description, the size of the image is denoted as M×N, where M is the number of pixels in the horizontal direction of the image and N is the number of pixels in the vertical direction.

[0038] The symmetric normalization method is the normalization algorithm in the prior art, which normalizes the gray value of each pixel in the image to the range of [0,1]. The normalized image and label image can be represented by the following formula:

[0039]

[0040] In the formula, A i is the pixel matrix of the normalized image, B i is the pixel matrix of the normalized label image, a is each pixel point in the image, and b is each pixel point in the label image.

[0041] Step S3, calculate the average distance of the sample group.

[0042] In this embodiment, the average distance is the average Euclidean distance. Based on the above-mentioned normalized image pixel matrix, for images A i and image B i the pixel-level Euclidean distance matrix is defined as follows:

[0043]

[0044] Based on the above definition, the pixel-level Euclidean distance between images A i and image B i is the square root of the sum of the elements in the pixel-level Euclidean distance matrix D i and is calculated according to the following formula:

[0045]

[0046] The average Euclidean distance of L groups of samples is calculated according to the following formula:

[0047]

[0048] Step S4, define the asymmetric normalization method.

[0049] In this embodiment, the definition of the asymmetric normalization method is as follows:

[0050]

[0051]

[0052] In the formula, a MN is the pixel matrix of the original image, b MN is the pixel matrix of the original label image, γ a , γ b are constants, and α, β are the range parameters.

[0053] To keep the normalized numerical spacing unchanged, usually take γ a = γ b , and taking different values of α, β can normalize the original image and the original label image to different ranges. At this time, the normalization ranges of the original image and the original label image are [α - β, 1 - (α - β)] and [0, 1] respectively.

[0054] Step S5, aiming to reduce the above average distance, use the particle swarm optimization method to optimize the range parameters α and β of the asymmetric normalization method.

[0055] In this embodiment, the particle swarm contains H particles, and each particle is a two-dimensional particle constructed based on the range parameters α and β. The optimization objective of this particle swarm optimization method is to reduce the average Euclidean distance of the above sample group, that is, use the above average Euclidean distance as the fitness function.

[0056] Figure 2 It is the flowchart of the particle swarm optimization method in the embodiment of the present invention.

[0057] As Figure 2 shown, the particle swarm optimization method of this embodiment includes the following steps:

[0058] Step A1, initialize the parameters of the particle swarm. The parameters of the particle swarm include the number of iterations k, the inertia coefficient ω, the individual learning factor c1, the swarm learning factor c2, the local learning rate r1, and the global learning rate r2.

[0059] Step A2, randomly initialize the positions X H and the change speeds V H of the H particles in the particle swarm.

[0060] Step A3, calculate the fitness of the H particles respectively, that is, the above average Euclidean distance.

[0061] Step A4, based on the fitness, select the global optimal particle P g and the local optimal particle P N from the H particles.

[0062] The global optimal particle P g is the particle with the minimum fitness in all historical iteration times, and the local optimal particle P N is the particle with the minimum fitness in the iteration time k.

[0063] Step A5, based on the global optimal particle P g and the local optimal particle P N , update the positions X H and the change speeds V H of the H particles.

[0064] The position X H of the particle is updated according to the following formula:

[0065]

[0066] The change speed V H of the particle is updated according to the following formula:

[0067]

[0068] Step A6: Determine whether the iteration number k is reached. If the determination result is yes, proceed to Step A7; if the determination result is no, increment the iteration number by 1, and then return to Step A3.

[0069] That is, stop the iteration when the iteration number k is reached. When the iteration number k is not reached, return to Step A3 to recalculate the fitness of H particles, find the global optimal particle P g and the local optimal particle P N and update the positions X of H particles H and the change speeds V H .

[0070] Step A7: Output the particle with the minimum fitness, that is, the optimal solution of the range parameters α and β.

[0071] As described above, through Step S5, the key parameters α and β of the asymmetric normalization method are optimized.

[0072] Step S6: Process the original image and the original label image using the asymmetric normalization method.

[0073] That is, perform asymmetric normalization on the original image and the original label image respectively according to the definitions in Step S4:

[0074]

[0075]

[0076] After performing asymmetric normalization, the normalization range of the image is [α - β, 1 - (α - β)], and the normalization range of the label image is [0, 1]. Combine the normalized image and the label image to form a new sample group, and apply the new sample group to the training of the segmentation network.

[0077] Through Step S5, the range parameters α and β of this asymmetric normalization method have been optimized. The optimized range parameters α and β make the average Euclidean distance of the sample group smaller than that before preprocessing, that is, the similarity between samples is improved. Therefore, using the new sample group for the training of the segmentation network can accelerate the convergence speed of the segmentation network and improve the accuracy of the trained segmentation network.

[0078] Functions and effects of the embodiment

[0079] According to the image and label asymmetric normalization method applied to the segmentation network provided in this embodiment, an asymmetric normalization method is defined, and the key parameters of this asymmetric normalization method are optimized with the goal of reducing the average Euclidean distance of multiple sample groups. Therefore, after performing asymmetric normalization processing on the original image and the original label image using this asymmetric normalization method, the average distance of the composed sample group is reduced, that is, the similarity between samples is higher. Thus, when training the segmentation network, the convergence speed of the segmentation network can be made faster, and the final network accuracy can be improved. Therefore, the asymmetric normalization method of this embodiment can improve the efficiency of segmentation network training and provide a reliable and better image preprocessing method for the segmentation of various medical images.

[0080] Specifically, in this embodiment, range parameters α and β are defined in the asymmetric normalization method, so that the original image and the original standard image can be normalized to different ranges based on the range parameters α and β, and with the goal of reducing the average Euclidean distance of the sample group, a particle swarm optimization method is used to optimize the key range parameters α and β. Therefore, through this asymmetric normalization method, the average distance of the sample group can be reduced, that is, the similarity between samples is improved. Thus, when training the segmentation network, the convergence speed of the segmentation network can be made faster, and the accuracy of the final segmentation network can be improved.

[0081] The above embodiments are only used to illustrate the specific implementation manners of the present invention, and the present invention is not limited to the description scope of the above embodiments.

Claims

1. An image and label asymmetric normalization method applied to a segmentation network, characterized in that, Including: Step S1: Obtain a plurality of original images and a plurality of original label images corresponding to the plurality of original images; Step S2: Use the symmetric normalization method to process the plurality of original images and the plurality of original label images respectively, to obtain a plurality of normalized images and a plurality of label images, and form a sample group from the plurality of images and the plurality of label images; Step S3: Calculate the average distance of the sample group; Step S4: Define the asymmetric normalization method; Step S5: With the goal of reducing the average distance, use the particle swarm optimization method to optimize the range parameters of the asymmetric normalization method; Step S6: Use the asymmetric normalization method to process the original images and the original label images, wherein, the asymmetric normalization method is defined as: where a MN is the pixel matrix of the original image, b MN is the pixel matrix of the original label image, γ a , γ b are constants, and α and β are the range parameters.

2. The image and label asymmetric normalization method applied to a segmentation network according to claim 1, characterized in that: Among them, The symmetric normalization method is a normalization algorithm, The sizes of the image and the label image are both M×N, The normalized image and the label image are represented by the following formula: a MN ∈ [0, 1], b MN ∈ [0, 1] Where, A i is the pixel matrix of the said image, B i is the pixel matrix of the said label image, a is the pixel of the said image, and b is the pixel of the said label image.

3. The image and label asymmetric normalization method applied to a segmentation network according to claim 2, characterized in that: Among them, The average distance is the average Euclidean distance, and is calculated according to the following formula: Where D i is the pixel-level Euclidean distance matrix, d is an element in the pixel-level Euclidean distance matrix, and DS ij is the pixel-level Euclidean distance between the said image and the said label image, and L is the number of samples.

4. The method for asymmetric normalization of images and labels applied to a segmentation network according to claim 3, Characterized in that: wherein, the particle swarm takes the average Euclidean distance as the fitness function, The particle swarm has H particles, The particle is a two-dimensional particle constructed based on the range parameters α and β, The particle swarm optimization method includes the following steps: Step A1: Initialize the parameters of the particle swarm, and the parameters at least include the number of iterations k; Step A2, randomly initialize the positions X of H particles in the particle swarm H and the velocity of change V H ; Step A3: Calculate the fitness of each of the H particles respectively; Step A4, select the global optimal particle P based on the fitness g and the local optimal particle P N ; Step A5, based on the global optimal particle P g and the local optimal particle P N , update the positions X H and the change speeds V H ; Step A6: Determine whether the number of iterations k has been reached. When the determination is yes, enter step B7. When the determination is no, return to step B3; Step A7: Output the particle with the minimum fitness, that is, the optimal solution of the range parameters α and β.

5. The image and label asymmetric normalization method applied to a segmentation network according to claim 4, characterized in that: Among them, The parameters further include the inertia coefficient ω, the individual learning factor c1, the group learning factor c2, the local learning rate r1, and the global learning rate r2.

6. The image and label asymmetric normalization method applied to a segmentation network according to claim 5, characterized in that: Among them, The velocity of the particle is calculated according to the following formula: The position of the particle is calculated according to the following formula:

7. The image and label asymmetric normalization method applied to a segmentation network according to claim 1, characterized in that: Among them, The label image is obtained by manual delineation by a clinician.

8. The image and label asymmetric normalization method applied to a segmentation network according to claim 1, characterized in that: Among them, The label image is obtained by an image processing method and verified by a clinician.

Images