CBCT image scattering artifact correction method and system

By adding attention module CBAM and MIND structural consistency loss functions to the CycleGAN network model, the AttMind-CycleGAN network model solves the problem of anatomical details fidelity in CBCT image scattering artifact removal, achieving high-quality artifact correction effect .

CN119067893APending Publication Date: 2024-12-03JIANGXI ZHONGKE JIUFENG ZHIHUI MEDICAL TECH CO LTD

Patent Information

Application Number
CN202411079460.8
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-08-07
Publication Date
2024-12-03

AI Technical Summary

Technical Problem

The existing CBCT image scattering artifact removal methods have shortcomings in calibration accuracy and anatomical detail fidelity, which is difficult to meet the needs of medical image processing.

Method used

Using the AttMind-CycleGAN network model, the structure between the CBCT image and the generated image is constrained by adding the attention module CBAM to the generator and the structural consistency loss function LMIND of the modal-independent domain descriptor MIND in the loss function. Consistency, artifact correction is performed in combination with the structural consistency of PCT images and generated images.

Benefits of technology

Effectively remove scattering artifacts from CBCT images while preserving the anatomical details of the image, improving image quality and accuracy of CT values.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119067893A_ABST
    Figure CN119067893A_ABST
Patent Text Reader

Abstract

The invention relates to a CBCT image scattering artifact correction method and system, and the method comprises the steps: obtaining the radiotherapy CBCT image data of a pelvic cavity patient and the corresponding PCT image data, carrying out the preprocessing, and building a CBCT image scattering artifact correction data set; inputting the CBCT image scattering artifact correction data set into an AttMind-CycleGAN network model, and carrying out scattering artifact correction by constraining the structural consistency between the CBCT image and the generated # imgabs0 # image and the structural consistency between the PCT image and the generated # imgabs1 # image; wherein an attention module (CBAM) is added in a generator of the AttMind-CycleGAN network model, and a structural consistency loss function (LMIND) of a mode-independent neighborhood descriptor (MIND) is added in a loss function of the generator. Compared with the prior art, images with higher fidelity can be generated, and the correction performance is better.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of computer-aided medicine, and in particular to a CBCT image scattering artifact correction method and system. Background Art

[0002] In recent years, cone beam computed tomography has been increasingly used, especially in adaptive radiotherapy. The real-time cone beam CT (CBCT) images obtained during radiotherapy contain a large number of scattering artifacts due to the large cone angle and low dose of the airborne CBCT scanning system, which hinders the development of precise radiotherapy and adaptive radiotherapy (ART). Compared with PCT images, CBCT images have more noise and imaging artifacts, making it difficult to accurately guide subsequent fractionated radiotherapy plans. Therefore, it is very important to remove scattering artifacts from CBCT images and improve image quality and CT value accuracy.

[0003] At present, many excellent traditional methods have been developed for the removal of scattering artifacts in CBCT images: convolution method, Monte Carlo method, scatter correction plate method, frequency modulation method, etc. However, various methods also have certain limitations in correction accuracy, model complexity, imaging dose, time consumption, etc. With the continuous improvement of computer computing power and the development of artificial intelligence in the field of medical image processing, some studies have applied deep learning methods to image artifact removal, which can obtain good correction results in a very short time. It is a very promising method for removing scattering artifacts. Among them, the unsupervised deep learning CycleGAN network model with the advantage of free matching in image data sets is widely used. However, this model has defects in the fidelity of image anatomical structure detail information, which is not allowed for medical image processing. Summary of the invention

[0004] The purpose of the present invention is to overcome the defects of the above-mentioned prior art and to provide a CBCT image scattering artifact correction method and system to achieve correction of CBCT image scattering artifacts.

[0005] The purpose of the present invention can be achieved by the following technical solutions:

[0006] A CBCT image scattering artifact correction method comprises the following steps:

[0007] Obtain pelvic patients' radiotherapy CBCT image data and corresponding PCT image data, perform preprocessing, and establish a CBCT image scattering artifact correction data set;

[0008] The CBCT image scattering artifact correction dataset is input into the AttMind-CycleGAN network model. Structural consistency between images and PCT images and generation Scattering artifact correction was performed based on structural consistency between images;

[0009] The attention module CBAM is added to the generator of the AttMind-CycleGAN network model, and the structural consistency loss function L of the modality-independent neighborhood descriptor MIND is added to the loss function of the generator. MIND .

[0010] Furthermore, the preprocessing includes: removing non-anatomical structures in the image and registering the CBCT image with the PCT image.

[0011] Furthermore, the attention module CBAM includes a channel attention module and a spatial attention module.

[0012] Furthermore, the expression of the modality-independent neighborhood descriptor MIND is as follows:

[0013]

[0014] Where I is a 2D grayscale image, is the pixel position of the MIND currently being calculated, is an element in the search space, R is the set of all pixels in the search space, n is a normalization constant, is the similarity measure between blocks, and is the descriptor pixel and search space pixels The distance between and The squared intensity error of the two blocks centered is expressed as follows:

[0015]

[0016] So Similarity measure of N-neighborhood search space centered on The average value is expressed as follows:

[0017]

[0018] In the formula, is an element in the search space, and N is the set of all pixels in the search space.

[0019] Furthermore, the structural consistency loss function LMIND The expression of is as follows:

[0020]

[0021] Where |R| is the similarity measure between the image patch in the search space and the image patch of interest in MIND The number of CBCT is the number of pixels in the CBCT image, n PCT is the number of pixels in the PCT image, is the pixel position of the MIND currently being calculated, is an element in the search space.

[0022] Furthermore, the AttMind-CycleGAN network model includes two generators F (PCTtoCBCT) and G (CBCTtoPCT) And two discriminators D PCT and D CBCT ;

[0023] Among them, the generator F (PCTtoCBCT) Used to generate Image domain data, the generator G (CBCTtoPCT) Used to generate Image domain data; the discriminator D CBCT For distinguishing real CBCT image domain data and the The authenticity of the image domain data, the discriminator D PCT For distinguishing true PCT image domain data and the The authenticity of image domain data.

[0024] Furthermore, the loss function of the AttMind-CycleGAN network model includes: the generator loss function L G and the discriminator loss function L D .

[0025] Furthermore, the generator loss function L G Also includes: cycle consistency loss function L cyc , domain identity recognition loss function L id , loss function L GAN ;

[0026] Among them, the cycle consistency loss function L cyc Used to constrain the real CBCT domain image data and the generated Image domain data and real PCT domain image data and generation Image domain data, its expression is as follows:

[0027] Lcyc (G,F)=E x~pdata(CBCT) [||F(G(CBCT))-CBCT||1]

[0028] +E y~pdata(PCT) [||G(F(PCT))-PCT||1]

[0029] In the formula, E is the mathematical expectation, F(G(CBCT)) is Image domain data, CBCT is the real CBCT domain image data, G(F(PCT)) is Image domain data, PCT is the real PCT domain image data;

[0030] The domain identity loss function L id It is used to constrain the real CBCT domain image data through the generator G (CBCTtoPCT) After generation, it is still The domain image data and the real PCT domain image data are passed through the generator F (PCTtoCBCT) After generation, it is still Domain image data, its expression is as follows:

[0031] L id (G,F)=E x~pdata(CBCT) [||G(CBCT)-CBCT||1]

[0032] +E y~pdata(PCT) [||F(PCT)-PCT||1]

[0033] In the formula, E is the mathematical expectation, G(CBCT) is domain image data, CBCT is the real CBCT domain image data, and F(PCT) is Domain image data, PCT is the real PCT domain image data;

[0034] The loss function L GAN For making the generator G (CBCTtoPCT) Generated Domain image data can deceive the discriminator D PcT And through the generator F (PCTtoCBcT) Generated Domain image data can deceive the discriminator D cBCT , which is expressed as follows:

[0035] L GAN (G, F, CBCT, PCT) = E x~pdata(CBCT) [(D PCT (G(CBCT))-1) 2 ]

[0036] +Ey~pdata(PCT) [(D CBCT (F(PCT))-1) 2 ]

[0037] In the formula, E is the mathematical expectation, D PCT (G(CBCT)) is The image passes through the discriminator D PCT Domain image data, D CBCT (F(PCT)) is The image passes through the discriminator D CBCT domain image data.

[0038] Furthermore, the generator loss function L G The expression of is as follows:

[0039] L G (G, F, CBCT, PCT) = λ cyc *L cyc +λ id *L id +L GAN +λ MIND *L MIND

[0040] In the formula, λ cyc , id and λ MIND is the weight parameter;

[0041] The discriminator loss function L D The expression of is as follows:

[0042] L D (D CBCT ,D PCT ,CBCT,PCT)=

[0043] E x~pdata(CBCT) [(D(CBCT)-1) 2 ]+E x~pdata(CBCT) [(D(G(CBCT)) 2 ]

[0044] +E y~pdata(PCT) [(D(PCT)-1) 2 ]+E y~pdata(PCT) [(D(F(PCT)) 2 ]

[0045] In the formula, E is the mathematical expectation, D is the discriminator, and G(CBCT) is Domain image data, F(PCT) is Domain image data.

[0046] According to another aspect of the present invention, a CBCT image scattering artifact correction system is provided, comprising:

[0047] A data processing module, which obtains and preprocesses the CBCT image data of pelvic patients undergoing radiotherapy and the corresponding PCT image data, and establishes a CBCT image scattering artifact correction data set;

[0048] Image correction module, using AttMind-CycleGAN network model, constrains CBCT images and generates Structural consistency between images and PCT images and generation Structural consistency between images, scatter artifact correction;

[0049] The attention module CBAM is added to the generator of the AttMind-CycleGAN network model, and the structural consistency loss function L of the modality-independent neighborhood descriptor MIND is added to the loss function of the generator. MIND .

[0050] Compared with the prior art, the present invention has the following beneficial effects:

[0051] 1. The present invention adds the MIND structure consistency loss function L MIND To directly constrain CBCT images and generate Structural consistency of images and PCT images and generation The structural consistency between images preserves the structural details of CBCT images when correcting artifacts, allowing the trained network model to generate higher-fidelity images.

[0052] 2. The present invention adds an attention module CBAM to the generator of the CycleGAN network model, so that the model focuses more on the anatomical structure information of the image and enhances the model's ability to learn anatomical features. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] Figure 1 It is a schematic diagram of the process of the present invention;

[0054] Figure 2 Schematic diagram of the structure of the AttMind-CycleGAN network model of the present invention, wherein (2a) is a schematic diagram of the structure of the generator of the AttMind-CycleGAN network model, (2b) is a schematic diagram of the structure of the discriminator of the AttMind-CycleGAN network model, and (2c) is a schematic diagram of the structure of the attention module in the generator of the AttMind-CycleGAN network model;

[0055] Figure 3 The network model generated in the embodiment A cross-section of an image;

[0056] Figure 4 Generated for the network model in the embodiment Horizontal, coronal, and sagittal planes of the image;

[0057] Figure 5 The network model generated in the embodiment Residual map of image and labeled PCT image. DETAILED DESCRIPTION

[0058] The present invention is described in detail below in conjunction with the accompanying drawings and specific embodiments. This embodiment is implemented based on the technical solution of the present invention, and provides a detailed implementation method and specific operation process, but the protection scope of the present invention is not limited to the following embodiments.

[0059] Abbreviations involved:

[0060] Cone Beam CT: Cone Beam CT, CBCT

[0061] Plan CT: Plan CT, PCT

[0062] Modality Independent Neighborhood Descriptor: MIND

[0063] Convolutional Block Attention Module, CBAM

[0064] Example 1

[0065] This embodiment provides a CBCT image scattering artifact correction method, such as Figure 1 As shown, the following steps are included:

[0066] S1, obtaining CBCT image data and corresponding PCT image data of pelvic patients undergoing radiotherapy, and performing preprocessing to establish a CBCT image scattering artifact correction data set;

[0067] In this embodiment, a total of 58 patients who underwent pelvic radiotherapy, and each patient included CBCT images and PCT images of the same field of view, and finally 52 cases that met the requirements of the study were screened out.

[0068] The CBCT images and PCT images are preprocessed, including the following steps:

[0069] To reduce the impact of the scanning bed in the image on network training, the scanning bed part in the CBCT and PCT images was removed;

[0070] Image registration of CBCT and PCT is performed to facilitate comparison of the effects after correction of the test set;

[0071] The intensity normalization operation is used to scale the pixel values ​​of the image to [-1, 1] to facilitate network processing and calculation.

[0072] Based on the image data of these 52 patients, a CBCT image scattering artifact correction dataset was established, which was divided into a training set of 43 patients (2895 images) and a test set of 9 patients (636 images). The characteristic of this dataset is that it cannot be completely matched, mainly because the CBCT images and PCT image data obtained by the patient at different times and on different devices during radiotherapy cannot be completely matched even after registration.

[0073] S2. Input the CBCT image scattering artifact correction dataset into the AttMind-CycleGAN network model. By constraining the CBCT image and generating Structural consistency between images and PCT images and generation Structural consistency between images for scatter artifact correction of CBCT images;

[0074] The AttMind-CycleGAN network model is constructed as follows:

[0075] Firstly, the unsupervised learning CycleGAN network model is adopted, and the model is trained using unmatched CBCT and PCT images. The generator of the CycleGAN network model is designed, and an attention module CBAM is added to the generator. The schematic diagram of the attention module structure is shown in Figure (2c), which includes a channel attention module and a spatial attention module. The purpose is to weight the shallow features of the generator and fuse them into the deep features of the network to emphasize the important features with anatomical information in the image. The loss function of the CycleGAN network model is designed, and the modality-independent neighborhood descriptor MIND structural consistency loss function L is added to the generator loss function of the CycleGAN network model. MIND , and the improved network model is named AttMind-CycleGAN network model.

[0076] We designed a loss function L based on the CBAM attention module and the MIND structure consistency loss function. MIND The improved CycleGAN network model is named AttMind-CycleGAN. The AttMind-CycleGAN network model includes two generators F (PCTtoCBCT) and G (CBCTtoPCT) And two discriminators D PCT and D CBCT .

[0077] Among them, the generator F (PCTtoCBCT) For generation from PCT artifact images Artifact-free image domain data, generator G (CBCTtoPCT) For generation from CBCT artifact images The artifact-free image domain data, the schematic diagram of the generator structure is shown in Figure (2a), including three downsampling and three upsampling structures. CBCT For distinguishing real CBCT image domain data and the The authenticity of image domain data, discriminator D PCT For distinguishing true PCT image domain data and the The authenticity of image domain data, the schematic diagram of the discriminator structure is shown in Figure (2b).

[0078] In the generator loss function L of the CycleGAN network model G Add MIND structural consistency loss function L MIND , used to constrain the real CBCT image and generate Images and real PCT images and generated The structural consistency between images can further preserve the anatomical structure of the artifact-corrected image while removing scattering artifacts. The loss function of the network model includes: a generator loss function L G and a discriminator loss function L D .

[0079] Among them, MIND is an image registration algorithm that aims to extract the structural information of the local area around each pixel or voxel. The expression of MIND is as follows:

[0080]

[0081] Where I is a 2D grayscale image, The pixel position of the MIND currently being calculated (such as the pixel in the 3rd column and the 6th row: ), is an element in the search space (such as ), R is the set of all pixels in the search space (such as the 4-neighborhood search space in a 2D image: {(1,0)(0,1)(-1,0)(0,-1)}), n is a normalization constant, is the similarity measure between blocks and is the descriptor pixel and search space pixels The distance between and The squared intensity error of the two blocks centered is expressed as follows:

[0082]

[0083] So Similarity measure of N-neighborhood search space centered on The average value is expressed as follows:

[0084]

[0085] In the formula, is an element in the search space, and N is the set of all pixels in the search space.

[0086] Generator loss function L G Including the generator loss function L G Includes: Cycle consistency loss function L cyc , domain identity recognition loss function L id , loss function L GAN And the structural consistency loss function L MIND ;

[0087] Cycle consistency loss function L cyc , used to constrain the real CBCT domain image data and the generated Image domain data as well as real PCT domain image data and generated Image domain data, the function expression is as follows:

[0088] L cyc (G,F)=E x~pdata(CBCT) [||F(G(CBCT))-CBCT||1]

[0089] +E y~pdata(PCT) [||G(F(PCT))-PCT||1]

[0090] In the formula, E is the mathematical expectation, F(G(CBCT)) is Image domain data, CBCT is the real CBCT domain image data, G(F(PCT)) is Image domain data, PCT is the real PCT domain image data;

[0091] Domain identity loss function L id , which is used to constrain the real CBCT domain image data to pass through the generator F (PCTtoCBCT) The generated The domain image data and the real PCT domain image data are passed through the generator G (CBCTtoPCT) The generated Domain image data, the function expression is as follows:

[0092] L id(G,F)=E x~pdata(CBCT) [||G(CBCT)-CBCT||1]

[0093] +E y~pdata(PCT) [||F(PCT)-PCT||1]

[0094] In the formula, E is the mathematical expectation, G(CBCT) is domain image data, CBCT is the real CBCT domain image data, and F(PCT) is Domain image data, PCT is the real PCT domain image data;

[0095] Loss function L GAN , we hope that the generator G (CBCTtoPCT) Generated Domain image data can deceive the discriminator D PCT , generator F (PCTtoCBCT) Generated Domain image data can deceive the discriminator D CBCT , the function expression is as follows:

[0096] L GAN (G, F, CBCT, PCT) = E x~pdata(CBCT) [(D PCT (G(CBCT))-1) 2 ]

[0097] +E y~pdata(PCT) [(D CBCT (F(PCT))-1) 2 ]

[0098] In the formula, E is the mathematical expectation, D PCT (G(CBCT)) is The image passes through the discriminator D PCT Domain image data, D CBCT (F(PCT)) is The image passes through the discriminator D CBCT Domain image data;

[0099] Structural consistency loss function L MIND The expression of is as follows:

[0100]

[0101] Where |R| is the similarity measure between the image patch in the search space and the image patch of interest in MIND The number of CBCT is the number of pixels in the CBCT image, n PCT is the number of pixels in the PCT image.

[0102] The generator loss function L G The expression is as follows:

[0103] L G (G, F, CBCT, PCT) = λ cyc *L cyc +λ id *L id +L GAN +λ MIND *L MIND

[0104] In the formula, λ cyc , id and λ MIND is the weight parameter.

[0105] Setting Lambda cyc =10,λ id =5,λ MIND Appropriate weight parameter values ​​need to be tested in experiments.

[0106] The discriminator loss function L D The expression is as follows:

[0107] L D (D CBCT ,D PCT ,CBCT,PCT)=

[0108] E x~pdata(CBCT) [(D(CBCT)-1) 2 ]+E x~pdata(CBCT) [(D(G(CBCT)) 2 ]

[0109] +E y~pdata(PCT) [(D(PCT)-1) 2 ]+E y~pdata(PCT) [(D(F(PCT)) 2 ]

[0110] In the formula, E is the mathematical expectation, D is the discriminator, and G(CBCT) is Domain image data, F(PCT) is Domain image data.

[0111] The AttMind-CycleGAN network model can remove scattering artifacts from CBCT images without the need for fully matched image data, while retaining the patient's anatomical details in the image and avoiding secondary artifacts.

[0112] When setting the hyperparameters in the network model training, the Adam optimizer algorithm was used, with parameters β1 = 0.5, β2 = 0.999; the network initialization weights were all set to random normal initialization, with a mean of 0 and a variance of 0.02; Batchsize was set to 1, Epoch was set to 100, and the initial learning rate was set to 0.0002, which remained unchanged in the first 50 Epochs and decreased linearly to 0 in the next 50 Epochs. The weight parameter λ of the MIND structural consistency loss function of the CycleGAN network model MIND The weights are set to 0, 0.5, 1, 2, 5, and 10 for testing to find the most appropriate weight parameters, while the image block Patch block inside MIND is set to 7*7 and the neighborhood search space is set to 9*9 for calculation and processing.

[0113] In the model evaluation stage, this embodiment uses 2895 patient pelvic image data to train the network model, and tests the network model on a test set of 636 pairs of CBCT images and PCT images. First, the original CycleGAN network model and the AttMind-CycleGAN network model are trained, and different weight parameters are set for the MIND loss function in the AttMind-CycleGAN model: MIND =0, 0.5, 1, 2, 5, 10, and then compare the effects of each model on the test set. The experimental results are shown in Table 1.

[0114] Table 1 Different weight parameters λ of MIND loss function MIND Comparison of the effects

[0115]

[0116] As can be seen from Table 1, the weight ratio λ of the added MIND structural consistency loss function in the generator loss function MIND Discussion. MIND =5, the generated The image quality is good. Compared with the uncorrected CBCT images, the three image quality evaluation indicators of PSNR, SSIM, and MSE are improved by 30.28%, 6.47%, and 76.87% respectively. Compared with the correction results of the CycleGAN network model, the three evaluation indicators of PSNR, SSIM, and MSE are improved by 2.24%, 1.31%, and 14.19% respectively. MIND =10, the generated Image Quality and Lambda MIND =5 Compared with the image quality, the differences in PSNR, SSIM, and MSE are 0.04%, 0.10%, and 0.90% respectively. However, the larger the weight of the MIND loss function, the more stripe artifact structures will be retained while retaining the anatomical structure. Therefore, λ is selected MIND The experiment was conducted with a weight parameter of 5.

[0117] Based on the experiments in Table 1, different models are compared to correct CBCT image artifacts. The quality evaluation of the images is shown in Table 2.

[0118] Table 2 Generated based on different correction methods Image and label PCT image quality assessment ratio

[0119]

[0120] As shown in Table 2, compared with the correction results of the CycleGAN network model, the correction results of the AttMind-CycleGAN model are improved by 2.24%, 1.31%, and 14.19% in the three evaluation indicators of PSNR, SSIM, and MSE, respectively. However, the corrected images obtained by the U-Net network model perform best in terms of the three evaluation indicators of PSNR, SSIM, and MSE.

[0121] Based on the experiments in Table 2, different models are compared to correct CBCT image artifacts. The structural consistency of the image, the experimental results are shown in Table 3.

[0122] Table 3 Generated based on different correction methods Structural consistency assessment of images and original CBCT images

[0123]

[0124] As shown in Table 3, the AttMind-CycleGAN network model performs well in both the structural consistency index SSIM and the modality-independent neighborhood descriptor MIND. Compared with the CycleGAN network model, the correction results are improved by 2.43% and 32.03%, respectively, and compared with the U-Net network model, the correction results are improved by 2.57% and 49.82%, respectively, indicating that this method can better retain the anatomical details of the image while removing the scattering artifacts of the CBCT image. However, even if the U-Net network model obtains good generated image quality, the structural consistency of the generated image is poor due to the incompletely matched data set used to train the network model.

[0125] Generated for different network models The image is visualized, and the profile diagram, three-section diagram and residual diagram of the image generated by different correction methods are shown as follows Figure 3 , Figure 4 , Figure 5 As shown in Figure 2, the intuitive comparison of the visualization results shows that the AttMind-CycleGAN network model performs well in both artifact removal and anatomical structure preservation.

[0126] In summary, this embodiment uses the CycleGAN network model with the CBAM attention module and the MIND structural consistency loss function to correct the scattering artifacts of radiotherapy CBCT images. This model can remove the scattering artifacts of CBCT images without completely matching experimental data PCT and CBCT images. Compared with the original CycleGAN network model, this method can better preserve the anatomical structure while improving the quality of CBCT images.

[0127] Example 2

[0128] This embodiment provides a CBCT image scattering artifact correction system, including:

[0129] A data processing module, which obtains and preprocesses the CBCT image data of pelvic patients undergoing radiotherapy and the corresponding PCT image data, and establishes a CBCT image scattering artifact correction data set;

[0130] Image correction module, using AttMind-CycleGAN network model, constrains CBCT images and generates Structural consistency between images and PCT images and generation Structural consistency between images, scatter artifact correction;

[0131] The attention module CBAM is added to the generator of the AttMind-CycleGAN network model, and the MIND structure consistency loss function L is added to the loss function of the generator. MIND .

[0132] The rest is the same as in Example 1.

[0133] The preferred specific embodiments of the present invention are described in detail above. It should be understood that a person skilled in the art can make many modifications and changes based on the concept of the present invention without creative work. Therefore, any technical solution that can be obtained by a person skilled in the art through logical analysis, reasoning or limited experiments based on the concept of the present invention on the basis of the prior art should be within the scope of protection determined by the claims.

Claims

1. A CBCT image scattering artifact correction method, characterized in that: The following steps are involved: Obtain pelvic patients' radiotherapy CBCT image data and corresponding PCT image data, perform preprocessing, and establish a CBCT image scattering artifact correction data set; The CBCT image scattering artifact correction dataset is input into the AttMind-CycleGAN network model. Structural consistency between images and PCT images and generation Scattering artifact correction was performed based on structural consistency between images; The attention module CBAM is added to the generator of the AttMind-CycleGAN network model, and the structural consistency loss function L of the modality-independent neighborhood descriptor MIND is added to the loss function of the generator. MIND .

2. The CBCT image scattering artifact correction method according to claim 1, characterized in that: The preprocessing includes: removing non-anatomical structures in the image and registering the CBCT image with the PCT image.

3. The CBCT image scattering artifact correction method according to claim 1, characterized in that: The attention module CBAM includes a channel attention module and a spatial attention module.

4. The CBCT image scattering artifact correction method according to claim 1, characterized in that: The expression of the modality-independent neighborhood descriptor MIND is as follows: Where I is a 2D grayscale image, is the pixel position of the MIND currently being calculated, is an element in the search space, R is the set of all pixels in the search space, n is a normalization constant, is the similarity measure between blocks, and is the descriptor pixel and search space pixels The distance between and The squared intensity error of the two blocks centered is expressed as follows: So Similarity measure of N-neighborhood search space centered on The average value is expressed as follows: In the formula, is an element in the search space, and N is the set of all pixels in the search space.

5. The CBCT image scattering artifact correction method according to claim 4, characterized in that: The structural consistency loss function L MIND The expression of is as follows: Where |R| is the similarity measure between the image patch in the search space and the image patch of interest in MIND The number of CBCT is the number of pixels in the CBCT image, n PCT is the number of pixels in the PCT image, is the pixel position of the MIND currently being calculated, is an element in the search space.

6. The CBCT image scattering artifact correction method according to claim 1, characterized in that: The AttMind-CycleGAN network model includes two generators F (PCTtoCBCT) and G (CBCTtoPCT) And two discriminators D PCT and D CBCT ; Among them, the generator F (PCTtoCBCT) Used to generate Image domain data, the generator G (CBCTtoPCT) Used to generate Image domain data; the discriminator D CBCT For distinguishing real CBCT image domain data and the The authenticity of the image domain data, the discriminator D PCT For distinguishing true PCT image domain data and the Authenticity of image domain data.

7. The CBCT image scattering artifact correction method according to claim 6, characterized in that: The loss function of the AttMind-CycleGAN network model includes: the generator loss function L G And the discriminator loss function L D .

8. The CBCT image scattering artifact correction method according to claim 7, characterized in that: The generator loss function L G Also includes: cycle consistency loss function L cyc , domain identity loss function L id , loss function L GAN ; Among them, the cycle consistency loss function L cyc Used to constrain the real CBCT domain image data and the generated Image domain data and real PCT domain image data and generation Image domain data, its expression is as follows: L cyc (G,F)=E x~pdata(CBCT) [||F(G(CBCT))-CBCT||1] +E y~pdata(PCT) [||G(F(PCT))-PCT||1] In the formula, E is the mathematical expectation, F(G(CBCT)) is Image domain data, CBCT is the real CBCT domain image data, G(F(PCT)) is Image domain data, PCT is the real PCT domain image data; The domain identity loss function L id It is used to constrain the real CBCT domain image data through the generator G (CBCTtopCT) After generation, it is still The domain image data and the real PCT domain image data are passed through the generator F (PCTtoCBCT) After generation, it is still Domain image data, its expression is as follows: L id (G,F)=E x~pdata(CBCT) [||G(CBCT)-CBCT||1] +E y~pdata(PCT) [||F(PCT)-PCT||1] In the formula, E is the mathematical expectation, G(CBCT) is domain image data, CBCT is the real CBCT domain image data, and F(PCT) is Domain image data, PCT is the real PCT domain image data; The loss function L GAN For making the generator G (CBCTtoPCT) Generated Domain image data can deceive the discriminator D PCT And through the generator F (PCTtoCBCT) Generated Domain image data can deceive the discriminator D CBCT , which is expressed as follows: L GAN (G,F,CBCT,PCT)=E x~pdata(CBCT) [(D PCT (G(CBCT))-1) 2 ] +E y~pdata(PCT) [(D CBCT (F(PCT))-1) 2 ] In the formula, E is the mathematical expectation, D PCT (G(CBCT)) is The image passes through the discriminator D PCT Domain image data, D CBCT (F(PCT)) is The image passes through the discriminator D CBCT domain image data.

9. The CBCT image scattering artifact correction method according to claim 7, characterized in that: The generator loss function L G The expression of is as follows: L G (G,F,CBCT,PCT)=λ cyc *L cyc +λ id *L id +L GAN +λ MIND *L MIND In the formula, λ cyc , id and λ MIND is the weight parameter; The discriminator loss function L D The expression of is as follows: L D (IS CBCT ,IS PCT ,CBCT,PCT)= E x~pdata(CBCT) [(D(CBCT)-1) 2 ]+E x~pdata(CBCT) [(D(G(CBCT)) 2 ] +E y~pdata(PCT) [(D(PCT)-1) 2 ]+E y~pdata(PCT) [(D(F(PCT)) 2 ] In the formula, E is the mathematical expectation, D is the discriminator, and G(CBCT) is Domain image data, F(PCT) is Domain image data.

10. A CBCT image scattering artifact correction system, characterized in that: include: A data processing module, which obtains and preprocesses the CBCT image data of pelvic patients undergoing radiotherapy and the corresponding PCT image data, and establishes a CBCT image scattering artifact correction data set; Image correction module, using AttMind-CycleGAN network model, constrains CBCT images and generates Structural consistency between images and PCT images and generation Structural consistency between images, scatter artifact correction; The attention module CBAM is added to the generator of the AttMind-CycleGAN network model, and the structural consistency loss function L of the modality-independent neighborhood descriptor MIND is added to the loss function of the generator. MIND .

Citation Information

Patent Citations

  • Method for correcting scattering artifact in CBCT based on planned CT

    CN113177991A

  • Cone beam CT image scattering artifact correction processing method and device

    CN116543070A

  • Arc evaporator having filtering net for macroparticle

    KR102602127B1

Cited By

  • CBCT scattering correction method based on rapid Monte Carlo simulation

    CN119679445A

  • Intelligent sketching method for cervical cancer radiotherapy target area and endangered organs

    CN121331396A