Improved level set heart MRI image segmentation method based on U-Net network

By introducing an improved horizontal set method in U-Net network, using new area term coefficients, edge indicator functions and constraint terms, the problem of inaccurate and time-consuming segmentation in cardiac MRI image segmentation is solved, achieving more efficient and accurate heart image segmentation effect.

CN119067993APending Publication Date: 2024-12-03QILU UNIVERSITY OF TECHNOLOGY (SHANDONG ACADEMY OF SCIENCES) +2
View PDF 0 Cites 1 Cited by

Patent Information

Application Number
CN202411282421.8
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-09-13
Publication Date
2024-12-03

AI Technical Summary

Technical Problem

The prior art has problems of inaccurate and time-consuming and labor-consuming in cardiac MRI image segmentation, especially in small details processing and initial contour selection, and the cardiac image data is not sufficient for large-scale training of neural networks.

Method used

Adopting an improved level set method based on U-Net network, U-Net preprocessing results are used as the initial contour, and new area term coefficients, edge indicator functions and constraint terms are introduced to adaptively adjust the convergence speed of the evolution curve and fuse the level set and U-Net models to improve segmentation efficiency and accuracy.

Benefits of technology

The segmentation efficiency and accuracy of the model are improved, the problems of oversegment and inadequate details are avoided, the problems of initial contour selection and insufficient data are solved, and the accuracy and calculation efficiency of cardiac image segmentation are significantly improved.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119067993A_ABST
    Figure CN119067993A_ABST
Patent Text Reader

Abstract

The invention discloses an improved level set heart MRI (Magnetic Resonance Imaging) image segmentation method based on a U-Net network, and relates to the field of computer vision, a proposed new edge indicator function can improve the evolution speed at a smooth pixel change position, and when the edge indicator function is close to a target boundary, an evolution curve can accurately stay at the target boundary, so that over-segmentation of a target image is avoided. In addition, a new constraint term is provided, the constraint term carries out linear fusion on a ground true value and a U-Net preprocessing result, the problem that curve convergence is inaccurate due to the fact that the U-Net does not process detail parts of the image in place can be avoided, a U-Net model and an improved level set model are connected, and the accuracy of curve convergence is improved. And the problem of initial contour selection can be solved to a certain extent.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of computer vision, and in particular to an improved level set cardiac MRI image segmentation method based on a U-Net network. Background Art

[0002] In recent years, the technology of image segmentation has become more and more mature, and there are more and more segmentation methods based on neural networks. The image segmentation methods based on neural networks often require a large number of images to train, and the neural network methods are not accurate in segmenting some relatively small details, and output some rough result images. The traditional active contour model is more accurate in segmenting some small details. Cardiovascular disease is still the number one killer of people's health. Segmentation of heart images is a necessary means to solve cardiovascular diseases. However, due to the particularity of heart activity and the differences in imaging equipment, some fuzzy and complex images are difficult to segment, and there is little data on heart images, which is difficult to use for large-scale training of neural networks. The traditional level set method requires manual setting of the initial contour of the segmentation, which is time-consuming and labor-intensive. Summary of the invention

[0003] In order to overcome the shortcomings of the above technologies, the present invention provides an improved level set cardiac MRI image segmentation method based on a U-Net network to improve the model segmentation efficiency and accuracy.

[0004] The technical solution adopted by the present invention to overcome the technical problems is:

[0005] An improved level set cardiac MRI image segmentation method based on U-Net network, comprising:

[0006] a) Select a cardiac MRI image I and a label I1 corresponding to the image;

[0007] b) Inputting the cardiac MRI image I into the U-Net model, and outputting a rough nursing result B1;

[0008] c) obtaining the circumscribed rectangle J of the heart region in the rough care result B1;

[0009] d) The circumscribed rectangle J is set to the cardiac MRI image I according to its coordinates. The area of ​​the circumscribed rectangle J in the cardiac MRI image I is p0. All pixel values ​​in the area p0 of the cardiac MRI image I are assigned to -2. All pixel values ​​outside the area p0 of the cardiac MRI image I are assigned to 2. The pixel values ​​of all pixels in the area p0 are summed to obtain the level set function φ 1 (x);

[0010] e) According to the level set function φ 1 (x) Calculate the initial level set function energy functional E(φ 1(x));

[0011] f) For the level set function φ 1 (x) Iterate k times to get the level set function φ k+1 (x);

[0012] g) If φ k+1 (x)-φ k (x)≤10 -5 , then according to the level set function φ 1 (x) Get the final cardiac MRI region segmentation result. If φ k+1 (x)-φ k (x)>10 -5 Then return to step f).

[0013] Furthermore, in step a), a cardiac MRI image I and a label I1 corresponding to the image I are randomly selected from the ACDC dataset.

[0014] Further, step c) comprises the following steps:

[0015] c-1) Use the np.where command, np.min command, and np.max command in ptcharm to obtain the circumscribed rectangle of the area with a pixel value of 255 in the rough care result B1, which is the circumscribed rectangle J of the heart area; c-2) Sum the pixel values ​​in the area with a pixel value equal to 255 to obtain φ U (x).

[0016] In step e), the formula

[0017] The energy functional of the level set function E(φ 1 (x)), where μ and γ are weight coefficients greater than 0, R p (φ 1 (x)) is the regularization term of the level set function, L gu (φ 1 (x)) is the length term of the level set function, δ(·) is the Dirac delta function, |·| is the absolute value, is the real number space, is the edge indicator function, is the area term coefficient, is the area term of the level set function, H(·) is the Heaviside function, W(φ(x),φ lab (x)) is a constraint item, β is a constant greater than 0,

[0018] Furthermore, through the formula Calculate the edge indicator function Where a and ε are constants greater than 0. The result of filtering the cardiac MRI image I using a two-dimensional Gaussian filter with a standard deviation of 2.0.

[0019] Further, step f) comprises the following steps:

[0020] f-1) through the formula φ 2 (x) = φ 1 (x)+ΔtL(φ 1 (x)) calculates the level set function φ after the first update 2 (x), where Δt = 5, The energy functional E(φ) of the level set function is obtained by the gradient flow equation. 1 (x)) is the result of taking the derivative;

[0021] f-2) The level set function φ 2 (x) Replace the level set function φ in step f-1) 1 (x) and then repeat step f-1) to obtain the second updated level set function φ 3 (x);

[0022] f-3) Repeat step f-2) k-2 times to obtain the level set function φ k+1 (x).

[0023] Furthermore, in step g), if φ k+1 (x)-φ k (x)≤10 -5 , then the level set function φ 1 The pixel values ​​less than 0 in (x) are displayed through the plt.imshow command in the Matplotlib library to obtain the final cardiac MRI region segmentation result.

[0024] The beneficial effects of the present invention are: pre-processing the cardiac image with the U-Net network, taking the circumscribed rectangle of the U-Net pre-processing result as the initial contour of the level set segmentation, and proposing a new area term coefficient, which fuses the level set evolution curve with the U-Net processing result, thereby adaptively adjusting the speed of convergence of the evolution curve, and improving the segmentation efficiency and accuracy of the model. The proposed new edge indicator function can improve the evolution speed at the position where the pixel changes smoothly, and when approaching the target boundary, the evolution curve can be accurately stopped at the target boundary to avoid over-segmentation of the target image. In addition, a new constraint term is proposed, which linearly fuses the ground truth value with the result of the U-Net pre-processing, which can avoid the problem of inaccurate curve convergence caused by the inadequate processing of the image details by U-Net, and connects the U-Net model and the improved level set model, and can also solve the problem of initial contour selection to a certain extent. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] Figure 1 This is a comparison diagram of the effect of the cardiac systolic period experiment of the present invention;

[0026] Figure 2 This is a comparison diagram of the effect of the cardiac diastolic experiment of the present invention;

[0027] Figure 3 It is a comparison chart of the Precision index of the present invention. DETAILED DESCRIPTION

[0028] The following is combined with Figure 1 , Attachment Figure 2 , Attachment Figure 3 The present invention is further described.

[0029] An improved level set cardiac MRI image segmentation method based on U-Net network, comprising:

[0030] a) Select a cardiac MRI image I and a label I1 corresponding to the image.

[0031] b) Input the cardiac MRI image I into the U-Net model and output the rough nursing result B1.

[0032] c) Obtain the circumscribed rectangle J of the heart region in the rough care result B1.

[0033] d) The circumscribed rectangle J is set to the cardiac MRI image I according to its coordinates. The area of ​​the circumscribed rectangle J in the cardiac MRI image I is p0. All pixel values ​​in the area p0 of the cardiac MRI image I are assigned to -2. All pixel values ​​outside the area p0 of the cardiac MRI image I are assigned to 2. The pixel values ​​of all pixels in the area p0 are summed to obtain the level set function φ1 (x).

[0034] e) According to the level set function φ 1 (x) Calculate the initial level set function energy functional E(φ 1 (x)).

[0035] f) For the level set function φ 1 (x) Iterate k times to get the level set function φ k+1 (x).

[0036] g) If φ k+1 (x)-φ k (x)≤10 -5 , then according to the level set function φ 1 (x) Get the final cardiac MRI region segmentation result. If φ k+1 (x)-φ k (x)>10 -5 Then return to step f).

[0037] The image is preprocessed using U-Net to obtain the initial contour of the level set segmentation, which solves the problem of difficulty in selecting the initial contour of the level set, and its preprocessing result is substituted into the edge indicator function for further processing. In one embodiment of the present invention, in step a), a cardiac MRI image I and a label I1 corresponding to the image I are randomly selected from the ACDC data set.

[0038] Attached Figure 1 The first column is the initial image contour, the second column is the DRLSE model segmentation result, the third column is the RSF model segmentation result, the fourth column is the proposed model segmentation result, the fifth column is the ALF model segmentation result, and the last column is the ground truth value. Figure 2 The first column is the initial image contour, the second column is the DRLSE model segmentation result, the third column is the RSF model segmentation result, the fourth column is the proposed model segmentation result, the fifth column is the ALF model segmentation result, and the last column is the ground truth value. Figure 1 and Figure 2 We can clearly see that the model in this paper has better segmentation effect on the left ventricle of the heart than other models, whether in the diastole or systole of the heart. The contour position of the segmented left ventricle is closer to the ground truth, and there is no problem of over-segmentation of the left ventricle.

[0039] Dice indicator comparison table

[0040]

[0041] The dice index shows that the proposed model has a higher accuracy rate in cardiac image segmentation than other models and has a faster computational efficiency. Compared with other models, the model has fewer iterations on the same image and a higher computational efficiency. Figure 3 The box plot also shows intuitively that the model proposed in this paper is more accurate than other models in segmentation.

[0042] In one embodiment of the present invention, step c) comprises the following steps:

[0043] c-1) Use the np.where command, np.min command, and np.max command in ptcharm to obtain the circumscribed rectangle of the area with a pixel value of 255 in the rough care result B1, which is the circumscribed rectangle J of the heart area.

[0044] c-2) Sum the pixel values ​​in the area where the pixel value is equal to 255 to get φ U (x).

[0045] In step e), the formula

[0046] The energy functional of the level set function E(φ 1 (x)), where μ and γ are weight coefficients greater than 0, R p (φ 1 (x)) is the regularization term of the level set function, is the length term of the level set function, δ(·) is the Dirac delta function, |·| is the absolute value, is the real number space, is the edge indicator function, is the area term coefficient, use the area term coefficient This variable replaces the fixed value of the original area term coefficient and can automatically change with the evolution curve, adaptively adjusting the convergence speed of the evolution curve. is the area term of the level set function, H(·) is the Heaviside function, W(φ(x),φ lab (x)) is a constraint item, β is a constant greater than 0, When the first evolution curve approaches the boundary, |φ 1 (x)-φ U (x)|decreases, As it decreases, the convergence speed slows down. When the first evolution curve is far away from the boundary, |φ 1 (x)-φ U(x)|increase, As it increases, the convergence speed is accelerated. Compared with the fixed area term coefficient, it can adaptively change the curve convergence speed according to the distance from the boundary, increase the convergence speed at a distance from the boundary, and reduce the convergence speed near the edge, avoiding problems such as over-segmentation and missed edges caused by being close to the boundary. It improves the overall segmentation efficiency of the model and has strong robustness for images with uneven grayscale. By proposing a new area term coefficient The convergence speed of the evolution curve can be adaptively controlled, the convergence speed is increased in places far from the boundary, and the convergence speed is reduced in places close to the boundary, thereby improving the efficiency of model segmentation. The present invention designs a new edge indication function The evolution speed at the position where the pixel changes slowly can be improved, and when approaching the target boundary, the evolution curve can be accurately stopped at the target boundary to avoid over-segmentation of the target image. lab (x)), the constraint W(φ(x),φ lab (x) Linearly fusing the ground truth value with the result of U-Net preprocessing can avoid the problem of inaccurate curve convergence caused by U-Net's inadequate processing of image details. It also connects the U-Net model and the improved level set model, and can solve the problem of initial contour selection to a certain extent.

[0047] In this embodiment, the formula Calculate the edge indicator function Where a and ε are constants greater than 0. The result of filtering the cardiac MRI image I with a two-dimensional Gaussian filter with a standard deviation of 2.0. The edge indicator function in the traditional distance regularization level concentration evolves slowly in places where the image gradient changes slowly, and when some blurred images are close to the boundary, the edge indicator function cannot control the curve to stay at the boundary well, causing the evolution curve to continue to converge, making the evolution curve invade the target and causing the problem of over-segmentation. The edge indicator function proposed in the present invention Where the first change is gentle Where the gradient changes gently, |φ 1 (x)-φ U The larger the value of (x)| As the value increases, the convergence speed is faster. When it is close to the boundary, |φ 1 (x)-φ U (x)| will decrease sharply, causing Rapidly reduce, control the evolution curve to converge accurately at the boundary, and avoid the evolution curve segmentation invading the boundary.

[0048] In one embodiment of the present invention, step f) comprises the following steps:

[0049] f-1) through the formula φ 2 (x) = φ 1 (x)+ΔtL(φ 1 (x)) calculates the level set function φ after the first update 2 (x), where Δt = 5, The energy functional E(φ) of the level set function is obtained by the gradient flow equation. 1 (x)) is the result of taking the derivative.

[0050] f-2) The level set function φ 2 (x) Replace the level set function φ in step f-1) 1 (x) and then repeat step f-1) to obtain the second updated level set function φ 3 (x).

[0051] f-3) Repeat step f-2) k-2 times to obtain the level set function φ k+1 (x).

[0052] In one embodiment of the present invention, in step g), if φ k+1 (x)-φ k (x)≤10 -5 , then the level set function φ 1 The pixel values ​​less than 0 in (x) are displayed through the plt.imshow command in the Matplotlib library to obtain the final cardiac MRI region segmentation result.

[0053] Finally, it should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art can still modify the technical solutions described in the aforementioned embodiments or replace some of the technical features therein by equivalents. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. An improved level set cardiac MRI image segmentation method based on U-Net network, characterized in that: include: a) Select a cardiac MRI image I and a label I1 corresponding to the image; b) Inputting the cardiac MRI image I into the U-Net model, and outputting a rough nursing result B1; c) obtaining the circumscribed rectangle J of the heart region in the rough care result B1; d) The circumscribed rectangle J is set to the cardiac MRI image I according to its coordinates. The area of ​​the circumscribed rectangle J in the cardiac MRI image I is p0. All pixel values ​​in the area p0 of the cardiac MRI image I are assigned to -2. All pixel values ​​outside the area p0 of the cardiac MRI image I are assigned to 2. The pixel values ​​of all pixels in the area p0 are summed to obtain the level set function φ 1 (x); e) According to the level set function φ 1 (x) Calculate the initial level set function energy functional E(φ 1 (x)); f) For the level set function φ 1 (x) Iterate k times to get the level set function φ k+1 (x); g) If φ k+1 (x)-φ k (x)≤10 -5 , then according to the level set function φ 1 (x) Get the final cardiac MRI region segmentation result. If φ k+1 (x)-φ k (x)>10 -5 Then return to step f).

2. The improved level set cardiac MRI image segmentation method based on U-Net network according to claim 1 is characterized in that: In step a), a cardiac MRI image I and a label I1 corresponding to the image I are randomly selected from the ACDC dataset.

3. The improved level set cardiac MRI image segmentation method based on U-Net network according to claim 1, characterized in that: Step c) comprises the following steps: c-1) Use the np.where command, np.min command, and np.max command in ptcharm to obtain the circumscribed rectangle of the area with a pixel value of 255 in the rough care result B1, which is the circumscribed rectangle J of the heart area; c-2) Sum the pixel values ​​in the area with a pixel value equal to 255 to obtain φ U (x).

4. The improved level set cardiac MRI image segmentation method based on U-Net network according to claim 3 is characterized in that: In step e), the formula The energy functional of the level set function E(φ 1 (x)), where μ and γ are weight coefficients greater than 0, R p (φ 1 (x)) is the regularization term of the level set function, is the length term of the level set function, δ(·) is the Dirac delta function, |·| is the absolute value, is the real number space, is the edge indicator function, is the area term coefficient, is the area term of the level set function, H(·) is the Heaviside function, W(φ(x),φ lab (x)) is a constraint item, β is a constant greater than 0, 5. The improved level set cardiac MRI image segmentation method based on U-Net network according to claim 4 is characterized in that: By formula Calculate the edge indicator function Where a and ε are constants greater than 0. The result of filtering the cardiac MRI image I using a two-dimensional Gaussian filter with a standard deviation of 2.

0.

6. The improved level set cardiac MRI image segmentation method based on U-Net network according to claim 1, characterized in that: Step f) comprises the following steps: f-1) through the formula φ 2 (x) = φ 1 (x)+ΔtL(φ 1 (x)) calculates the level set function φ after the first update 2 (x), where Δt = 5, The energy functional E(φ) of the level set function is obtained by the gradient flow equation. 1 (x)) is the result of taking the derivative; f-2) The level set function φ 2 (x) Replace the level set function φ in step f-1) 1 (x) and then repeat step f-1) to obtain the second updated level set function φ 3 (x); f-3) Repeat step f-2) k-2 times to obtain the level set function φ k+1 (x).

7. The improved level set cardiac MRI image segmentation method based on U-Net network according to claim 1, characterized in that: In step g), if φ k+1 (x)-φ k (x)≤10 -5 , then the level set function φ 1 The pixel values ​​less than 0 in (x) are displayed through the plt.imshow command in the Matplotlib library to obtain the final cardiac MRI region segmentation result.

Citation Information

Cited By

  • Heart nuclear magnetic resonance image segmentation method and system based on supervision signal expansion

    CN120411506A