High resolution image segmentation method based on region probability label driven active contour

Abstract

The application provides a high-resolution image segmentation method based on region probability label driven active contour, and solves the problem that an active contour model cannot process high-resolution images.The steps are as follows: firstly, a high-resolution image is down-sampled to obtain images from low to high resolution, and the following processing is performed on the images from low to high resolution: an initial level set function and statistical parameters are obtained, and a region probability label and geometric prior are calculated by using the level set function; secondly, a segmentation contour is updated under the driving of the region probability label, and the corresponding level set function and statistical parameters are updated based on the region probability label update driving; and finally, the level set function and statistical parameters of the current resolution image are transmitted to the next high-resolution image and contour segmentation is performed on the next high-resolution image.The application can greatly reduce the calculation cost and enhance the robustness of the algorithm, so that the high-resolution image can be processed.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of medical image analysis and remote sensing image processing, and particularly relates to a high-resolution image segmentation method based on region probability label driven active contour. BACKGROUND

[0002] The rapid development of various professional imaging equipment enables professional technicians to obtain larger resolution images, so as to obtain more abundant and comprehensive information. Traditional image processing methods regard the image as a multi-dimensional matrix for processing, and the typical application object is an image with a resolution lower than 500*500. However, the existing high-resolution image reaches 10000*10000 or even higher. The most direct problem faced by these methods is that many models cannot be used due to the multiplied calculation amount. The common solution is to divide the large-resolution image into blocks, but the block processing destroys the global information of the image.

[0003] At present, the super-high-resolution images mainly include medical images and remote sensing images, such as whole-slide pathological images. The whole-slide digital image refers to scanning a complete microscope slide and creating a single high-resolution digital file. The emergence of this technology enables efficient data storage in clinical applications, enables cross-experimental repeatable data analysis, and enables simple and instant data sharing, but also faces some problems mentioned above. With more and more application of artificial intelligence related technologies in medical images, the underlying image processing becomes more and more important. For example, when training and learning pathological images, the pathological images need to be preprocessed to remove the background other than the target, which requires a segmentation algorithm for high-resolution images to obtain an accurate edge.

[0004] The active contour method is difficult to process high-resolution images, and the essence is the difficulty in balancing precision and speed. The large calculation amount and low execution efficiency have always plagued the development of the level set algorithm. At present, some scholars have also proposed some fast level set algorithms. Some models propose a continuous parameter model on the span space of B-spline, and perform filtering through convolution to evolve the level set function, thereby reducing the calculation amount by avoiding solving partial differential equations, but it is still not fast enough to obtain accurate results in complex natural images. There are also local appearance models using additional level sets, so as to divide the hierarchy of object shape into multiple connected regions with different appearance models, and use the form of pixel probability map obtained from the object detection process to represent the shape. However, the images processed by these methods are basically below 500*500.

[0005] Through the above analysis, there is currently a lack of an image segmentation algorithm for high-resolution images, which can obtain a smooth and continuous segmentation contour. SUMMARY

[0006] To address the shortcomings of the aforementioned background technology, this invention proposes a high-resolution image segmentation method based on region probability label-driven active contours, which solves the problem that active contour models cannot handle high-resolution images.

[0007] The technical solution of this invention is implemented as follows:

[0008] A high-resolution image segmentation method based on region probability label-driven active contouring, comprising the following steps:

[0009] Step 1: Process the original high-resolution image I K Downsampling is performed to obtain a series of images I k k∈{1,2,3,…K}, from I1 to I K Increasing resolution;

[0010] Step 2: Load Image I k Preprocessing is performed: when k = 1, step three is executed; when k > 1, step four is executed; when k > K, the operation ends.

[0011] Step 3: Determine the initial contour interactively, based on the initial contour and image I. k The size of the level set function Φ is initialized, and the statistical parameters are initialized according to the level set function Φ.

[0012] Step 4: Based on image I k The size of the pair is determined by I k-1 The obtained level set function Upsampling is performed to obtain And will be by I k-1 Statistical parameters as image I k Statistical parameters;

[0013] Step 5: Calculate the region probability labels based on the level set function and statistical parameters, and update the segmentation contours driven by the region probability labels;

[0014] Step 6: Update the level set function and statistical parameters based on the region probability labels;

[0015] Step 7: Determine if the stopping condition is met. If not, return to Step 5. If it is met, let k = k + 1. When k > K, the operation ends and the result is output. Otherwise, return to Step 2.

[0016] The initialization methods for the level set function and statistical parameters are as follows:

[0017] Let the observed values ​​of the image be d-dimensional variables x j It follows a multivariate Student's-t distribution, j={1,2,...,N}, Ω iis the 2D image space, i ∈ {1, 2}, and each pixel x j ∈Ω i , for each pixel x

[0018]

[0019] where S(x j |θ i ) represents the distribution function, θ i = {μ i ,∑ i ,ν i} is the statistical parameter, μ i is the mean,∑ i is the variance matrix, and ν i

[0020] The pixel x j corresponding to the level set function is φ j ∈ R, and when k = 1, it is initialized by an interactive method.

[0021] Let H(·) represent the Heaviside step function:

[0022]

[0023] The mean μ i is initialized as:

[0024]

[0025] The variance matrix∑ i is initialized as:

[0026]

[0027] The degree of freedom ν i is initialized as: ν1 = ν2 = 1.

[0028] The implementation method of step four is:

[0029] When k > 1, the level set and statistical parameters of the image I k are determined by the segmentation result of the image I k-1 , and the specific method is:

[0030] First, the level set function is obtained by I k-1 According to the size of I k , the level set function is up-sampled to obtain

[0031]

[0032] The statistical parameters of the image I k-1 are taken as the initial values of the statistical parameters of the segmentation I k :

[0033]

[0034]

[0035]

[0036] After the above initialization, a new round of segmentation is performed on the image I k .

[0037] The calculation method of the region probability label is:

[0038] The geometric prior w i is calculated by the current level set function:

[0039]

[0040] w2=1-w1

[0041] Wherein, ε is an approximation coefficient;

[0042] The region Student's-t distribution is proposed:

[0043]

[0044] Wherein, p LSt (x j |θ i ) represents the region probability distribution, represents the Student's t distribution function, and α j is the region centered on x j ; the region Student's-t distribution is brought into the Bayes formula, and the region probability label p ji of x j belonging to Ω i is obtained:

[0045]

[0046] The updating method of the level set function is:

[0047]

[0048] Wherein n represents the nth iteration, Δt represents the time step, and β and γ are used to adjust the weights of the contour length and area in the energy function, respectively. This represents the gradient.

[0049] The method for updating the statistical parameters is as follows:

[0050] mean μ i The update formula is:

[0051]

[0052] in, These are the weighting coefficients for the statistical distribution;

[0053] variance matrix ∑ i The update formula is:

[0054]

[0055] Degrees of freedom ν i (k) The update formula is:

[0056]

[0057] in, It is the derivative of the logarithm of the gamma function.

[0058] Compared with existing technologies, the beneficial effects of this invention are as follows: This invention obtains a low-resolution image of the original image through downsampling, and then transfers the information obtained from the low-resolution image to the segmentation process of the high-resolution image. This "relay" approach using images of different resolutions drives the evolution of the level set at a lower cost. By processing the downsampled low-resolution image, an initial contour that closely approximates the target is obtained at a lower cost, thus reducing the running time in the original image space. This invention can significantly reduce computational costs and enhance the robustness of the algorithm, thereby processing ultra-high resolution images. Attached Figure Description

[0059] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0060] Figure 1 This is a flowchart of a method according to an embodiment of the present invention.

[0061] Figure 2 It is the initial contour image.

[0062] Figure 3 is the initial level set image.

[0063] Figure 4 is the final result of the method of the present application.

[0064] Figure 5 is the final level set of the method of the present application.

[0065] Figure 6 is the local map of the method of the present application.

[0066] Figure 7 is the label comparison map of the method of the present application and other algorithms; wherein (a) the method of the present application, (b) the EM-SMM algorithm, (c) the K-means algorithm. DETAILED DESCRIPTION

[0067] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only some of the embodiments of the present application, but not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative effort fall within the scope of the present application.

[0068] As shown in Figure 1 , the embodiment of the present application provides a high-resolution image segmentation method based on region probability label driven active contour. The present application obtains a low-resolution image of the original image through downsampling, transmits the information obtained in the low-resolution image to the segmentation process of the high-resolution image, and uses the "relay" mode of different resolution images to drive the level set evolution at a lower cost. The core idea is to obtain an initial contour as close to the target as possible at a lower cost through processing of the low-resolution image obtained by downsampling, so as to reduce the running time in the original image space. The algorithm aims to minimize the active contour energy function, and the algorithm can be roughly divided into three steps: 1) interactive initialization; 2) evolution of the level set function driven by the region probability label; and 3) the level set obtained by the low-resolution image is upsampled and used as the initial value of the high-resolution image segmentation together with the statistical parameters obtained. The second step and the third step are alternately performed until the stop condition is met. The specific steps are as follows:

[0069] Step one: downsample the original high-resolution image I K to obtain a series of images I k with decreasing resolution, until the resolution of I1 is below 1000x1000, wherein k∈{1,2,3,…K}, and the pixel of the original high-resolution image I K is about 10000x10000.

[0070] Step two: for the image I kRead, load and pre-process the image according to the resolution from low to high, k e {1,2,3}; when k = 1, execute step three, when k > 1, execute step four, when k > K, end the operation.

[0071] Step three: determine the initial contour through interaction, initialize the level set function Φ according to the initial contour and the image I k , and initialize the statistical parameters according to the level set function Φ;

[0072] Let the observation value d of the image x j be subject to a multivariate Student's-t distribution, j = {1,2,...,N}, Ω i be a 2D image space, i e {1,2}, and each pixel x j e Ω i , for each pixel, there is:

[0073]

[0074] where S(x j | θ i ) represents a distribution function, θ i = {μ i ,∑ i ,ν i} is a statistical parameter, μ i is a mean value,∑ i is a variance matrix, and v i is a degree of freedom; Γ(·) is a Gamma function, Δ ji 2 is a Mahalanobis distance: where x j is a pixel observation value, and the image uses three channels d = 3 in the present example.

[0075] Φ = (φ1,φ2,...,φ N ) is a level set function, φ j e R is a signed distance function, and represents the corresponding level set function of x j , the level set function Φ can be automatically initialized by K-means, and is initialized by interaction in the present example, for example, the initial value is a circle with a diameter equal to the short side of the image. Generally, a unit matrix with the same size as the image is set, and the values inside the initial curve are set to zero.

[0076]

[0077] Let H(·) represent the Heaviside step function:

[0078]

[0079] mean μ i i∈{1,2} is initialized using the following formula:

[0080]

[0081] Square difference matrix ∑ i Perform initialization:

[0082]

[0083] For degrees of freedom ν i Initialize the values: ν1 = ν2 = 1.

[0084] Step 4: Based on image I k The size of the pair is determined by I k-1 The obtained level set function Upsampling is performed to obtain And will be by I k-1 Statistical parameters as image I k Statistical parameters;

[0085] When k>1, image I k The level set and statistical parameters are derived from image I k-1 The segmentation result is used to determine the specific method as follows:

[0086] First, through I k-1 The obtained level set function According to I k Size of the level set function Upsampling is performed, which involves linear interpolation to obtain the result.

[0087]

[0088] Active contouring methods evolve from an initial contour to a target contour, thus the initial contour largely determines the final result. In this invention, initializing and segmenting in a high-resolution image would generate a huge computational load. Therefore, the core idea of ​​the algorithm is to obtain a relatively accurate contour in a low-resolution image and upsample it to serve as the initial contour for the high-resolution image, thereby reducing the computational load.

[0089] In image I k-1 After segmentation, both the level set function and statistical parameters are obtained. The statistical parameters are not directly related to the resolution; their size is related to the dimension of the data. Therefore, this invention uses image I... k-1 The resulting final statistics are used as the starting values ​​for segmentation of the next image at a higher resolution:

[0090]

[0091] After the above initialization, the image I k is segmented again.

[0092] Step five: calculate the region probability label based on the level set function and the statistical parameter, and update the segmentation contour under the driving of the region probability label;

[0093] The present application proposes a region student t distribution:

[0094]

[0095] Wherein, p LSt (x j |θ i ) represents the region probability distribution, represents the student t distribution function, and alpha j is the region centered on x j ; the above formula is brought into the Bayes inference to obtain the region probability label p ji of x i belonging to Omega j :

[0096]

[0097] Wherein w is a geometric prior, which is calculated by the current level set function:

[0098]

[0099] In the above formula, H (phi) is defined by formula (3), and it is known that H (·) is hard classification, which is 1 or 0, and can be understood as the proportion of the pixel points in the segmentation curve to all the pixel points. This kind of prior is relatively rough and cannot represent more abundant geometric information such as the transition of the edge in the level set, and H (·) is discrete and non-differentiable. Therefore, an approximation commonly used in H (·) is introduced in the present example:

[0100]

[0101] Wherein, epsilon is an approximation coefficient, when epsilon tends to zero, the equation is established, which is also a commonly used way to approximate formula (3) in machine learning.

[0102] Therefore, the prior containing more level set geometric information can be obtained:

[0103]

[0104] Step six: update the level set function and the statistical parameter based on the region probability label;

[0105] Level set updating equation driven by region probability label. The cost function of the probabilistic Chan-Vese model is used in this embodiment:

[0106]

[0107] The first two terms are data terms, and the third and fourth terms are used to restrict the length and area of the segmentation contour, respectively. The parameters β and γ are used to adjust the relationship between the terms.

[0108] In addition, the present application uses the logarithm of the odds ratio to give another definition of the level set function φ:

[0109]

[0110] The evolution equation of the n+1 iteration of the level set can be obtained by the gradient descent method as:

[0111]

[0112] In the above formula, the can be brought into the formula (11) by the formula (14) represents the gradient, and Δt represents the time step which controls the The contribution of the narrow band of Φ to each iteration.

[0113] In order to reduce the singular value in the level set function, the formula (17) is rewritten as:

[0114]

[0115] In the above formula, It can be seen from the above formula that the updating of the level set function is driven by the probability label and the statistical parameters {p ji ,θ i}.

[0116] Statistical parameter updating based on region probability label. The parameter updating is realized by minimizing the energy function. First, the formula (14) is substituted into the formula (11) to obtain Substitute the approximate expression into the energy function (13) and discretize to obtain:

[0117]

[0118] The goal of this part is to obtain the statistical parameters θ i of the t-distribution, so as to bring the evolved level set function into the formula (16). Since the student t-distribution S(x|θ i ) can be written as a weighted sum of Gaussian distributions, the weight is given by the Gamma function. Therefore, in order to facilitate operation, first introduce a hidden variable weight coefficient ui and let the random variable u i ~ Gam(ν i / 2,ν i / 2), where the Gamma function is defined by where Γ(a) is defined previously.

[0119] Under the condition of the newly added weight u i , (17) can be rewritten as

[0120]

[0121] Since this part is to obtain the updating formula of the statistical parameters, it is assumed that p ji is fixed and known, so the above formula is removed from the irrelevant quantity, and if the above formula is minimized, it is to solve

[0122]

[0123] The above formula is the same as the objective function of the expectation maximum value, so the hidden variable u ij is replaced by its conditional expectation:

[0124]

[0125] Then the updating of the parameters can be solved by taking the first-order derivative of the parameters with respect to u ij and setting it equal to 0, and the solution is:

[0126]

[0127] The updating formula of the variance matrix ∑ i is:

[0128]

[0129] The updating formula of the degree of freedom is:

[0130]

[0131] where ψ(y) is the derivative of the logarithm of the gamma function. is a nonlinear equation, and a bisection method is used to obtain an approximate solution.

[0132] Step seven: judge whether the stopping condition is met, if not, return to step five; if yes, let k = k + 1, when k > K, the operation is completed, and the result is output, otherwise, return to step two.

[0133] Example analysis

[0134] To verify the performance of the method, for a pair of SVS format whole slide histopathology images from a public dataset, Fig. I k , k∈{1, 2, 3}, are different scale images with the image, and the resolutions are 768x975, 2171x2757, 8687x11031 respectively. According to the size of the image, α j In this example, it varies from 5x5, 7x7, 9x9, ε=0.001, β=1, γ=-0.25. The initial contour and the level set function are shown in Fig. 2 Figure 2 、 3 The segmentation results are shown in Fig. 3 Figure 4-5 , Figure 6 and Fig. 4 is a local image of the segmentation result.

[0135] Table 1 Comparison of segmentation time

[0136] The method of the invention CV model EM-SMM K-means Time (sec) 86.36 - 319.40 73.39

[0137] Table 1 shows the time consumption of four algorithms for segmenting the image, among which the typical active contour method Chan-Vese model cannot be processed due to the large image, and the comparison algorithms also include expectation maximization based on student's t mixture model (EM-SMM) and K-means. Figure 7 The experimental comparison of the three images shows that the segmentation performance of the method is better.

[0138] Based on the same inventive concept, the application further discloses a high-resolution image segmentation device based on a region probability label driven active contour, which comprises a memory, a processor and a computer program stored in the memory and capable of running on the processor, and the computer program realizes the high-resolution image segmentation method based on the region probability label driven active contour when loaded into the processor.

[0139] The above only describes the preferred embodiments of the application and is not used to limit the application, and any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the application shall be included in the protection scope of the application.

Claims

1. A method for high resolution image segmentation based on region probability label driven active contour, characterized in that, The steps are as follows: Step one: downsample the original high resolution image I K to a series of images I k , k e {1,2,3,…K}, with increasing resolution from I K to I Step two: load image I k and pre-process, when k = 1, perform step three, when k > 1, perform step four, When k>K, the operation is ended; Step three: Determine the initial contour by interactive means, initialize the level set function Φ according to the size of the initial contour and the image I k , And initialize the statistical parameters according to the level set function Φ; Step four: upsample the level set function obtained from I k-1 according to the size of image I k and take the statistical parameters of I k-1 as the statistical parameters of image I k ;​​ Step five: calculate the region probability label based on the level set function and the statistical parameters, and update the segmentation contour under the driving of the region probability label; Step six: update the level set function and the statistical parameters based on the region probability label; Step seven: judge whether the stop condition is met, if not, return to step five; if yes, let k=k+1, when k>K, The operation is ended, and the output result is output, otherwise, return to step two; The calculation method of the region probability label is: Calculating the geometric prior w by the current level set function i : w2=1-w1 where ε is an approximation coefficient; φ j is the pixel x j the corresponding level set function; Proposed region Student's-t distribution: where p LSt (x j |θ i ) represents the region probability distribution, θ i ={μ i ,Σ i ,ν i} is the statistical parameter, μ i is the mean, Σ i is the variance matrix, and ν i is the degree of freedom; is the Student’s t distribution function, and α j is the region centered at pixel x j . The region Student’s-t distribution is brought into the Bayes formula to obtain the region probability label of pixel x j belonging to Ω i : The updating method of the level set function is: wherein n denotes the nth iteration, At denotes the time step, and β and γ are used to adjust the weight of the contour length and area in the energy function, respectively, denotes the gradient; The updating method of the statistical parameters is: The mean μ i The update formula is: wherein, is a weight coefficient for the statistical distribution; d is the dimension of the pixel x j is the Mahalanobis distance; Δ ji 2 is the Mahalanobis distance; Δ covariance matrix i The update formula for is Degrees of freedom v i The update formula is: wherein is the derivative of the log of the gamma function Γ(·).

2. The region probability label driven active contour based high resolution image segmentation method according to claim 1, characterized in that, The initialization method of the level set function and the statistical parameters is: Let d be the observed value of an image of dimension x j obeys a multivariate Student's-t distribution, j = {1, 2,..., N}, Ω i is the 2D image space, i ∈ {1, 2}, and each pixel x j ∈ Ω i For each pixel has: where S(x j |θ i ) denotes the distribution function, θ i = {μ i ,Σ i ,ν i} is the statistical parameter, μ i is the mean, Σ i is the variance matrix, and ν i is the degree of freedom; Γ(·) is the Gamma function, Δ ji 2 is the Mahalanobis distance: Pixel x j The corresponding level set function is φ j ∈R, when k = 1, initialized by the method of interaction; Let H(·) represent the Heaviside step function: Initialize the mean μ i Initialize: covariance matrix i Initialization is performed: for the degrees of freedom v i Initial values are assigned: v1 = v2 = 1.

3. The region probability label driven active contour based high resolution image segmentation method according to claim 2, characterized in that, The implementation method of step four is: When k > 1, the level set and statistical parameters of the image I k are determined by the segmentation result of the image I k-1 , and the specific method is as follows: First, by I k-1 the level set function According to I k the size of the level set function up-sampling to obtain The statistical parameters of the image I k-1 are determined as I k The starting values of the statistical parameters for the segmentation are: After the above initialization, the image I k is subjected to a new round of segmentation.

Images