A manifold clustering and weighted sparsity based respiratory motion compensation method

Abstract

The present application provides a kind of respiratory motion compensation method based on manifold clustering and weighted sparse, comprising the following steps: collecting the 4D-CT image of population data set and the 3D-CT image of individual subject expiration-inhalation two phases;Obtain the deformation vector field motion characteristics of population data set and individual subject by image registration;Obtain the sub-population data set similar to individual subject motion characteristics based on manifold clustering;The quantitative relationship between individual and sub-population motion characteristics is established by using weighted sparse method, and individual 4D-CT image is constructed;Obtain the internal deformation vector field and external skin surface motion of individual subject, and construct individualized respiratory motion compensation model to realize real-time prediction of lung image.The present application can reduce the radiation damage of the subject while ensuring accurate prediction of lung movement, and is expected to break through the limitations of existing respiratory motion compensation methods, can real-time image prediction for whole lung, reduce the influence of respiratory motion on lung tissue.

Description

Technical Field

[0001] This invention relates to the fields of medical image processing, computer vision, and machine learning, specifically to a respiratory motion compensation method based on manifold clustering and weighted sparsity. Background Technology

[0002] With the development of technologies such as computers, medical imaging, and high-precision measurement, it is possible to construct multimodal images of the human body. However, due to the influence of respiratory motion, lung tissue undergoes continuous movement along with the entire lung, resulting in significant real-time errors in lung CT images. Therefore, studying the laws governing respiratory motion, especially its impact on lung tissue structure, is of great significance.

[0003] Currently, several methods exist for addressing respiratory motion problems, including breath-holding / breathing gating, motion tracking, and respiratory motion compensation. Breath-holding / breathing gating is the simplest method, but it offers relatively coarse control of the respiratory phase, making it difficult to guarantee accuracy, and some subjects struggle to meet the requirements of breath-holding or breathing gating. Motion tracking requires implanting markers in the lung region, then using imaging equipment (such as X-rays) to monitor respiratory motion by tracking the markers. However, this method is invasive, has a limited monitoring range, and only provides relatively accurate lung motion information near the markers. Furthermore, it carries a significant risk of radiation damage. Due to the limitations of these technologies, researchers have recently shown great interest in how to computationally estimate and correct the impact of respiratory motion on lung deformation and structural movement—that is, respiratory motion compensation models. Respiratory motion compensation models can reduce radiation dose, eliminate the need for breath-holding, and predict lung movement over a wide range, thus becoming a current research hotspot.

[0004] Common respiratory motion compensation methods can be broadly categorized into two types: individualized respiratory motion compensation models and global respiratory motion compensation models. Individualized respiratory motion compensation models first acquire 4D-CT images of individual subjects, using these images to establish a correlation between lung deformation fields and skin surface motion, and then apply this model to predict real-time lung images. A drawback of this method is that the acquisition of 4D-CT images covering the entire respiratory cycle of the subject poses a significant risk of radiation damage. To address this issue, a population-based respiratory motion compensation model has been proposed, establishing a motion compensation relationship between the lung deformation field and skin surface motion of a group of subjects based on statistical models / deep learning. This model extracts common features of lung and skin motion based on the entire population, eliminating the need to acquire individual 4D-CT images when predicting real-time lung images, thus effectively reducing radiation dose. However, due to the variability and randomness of individual lung motion, the individual characteristics that best approximate the actual motion of a specific subject may be overlooked.

[0005] This invention discloses a personalized respiratory motion compensation method based on manifold clustering and weighted sparsity, which effectively combines group motion characteristics with individual phases, achieving a good balance between radiation risk and model accuracy while meeting real-time requirements. Summary of the Invention

[0006] The technical problem solved by this invention is to overcome the shortcomings of existing technologies and provide a personalized respiratory motion compensation method based on manifold clustering and weighted sparsity. This method effectively combines group characteristics and individual phases to perform real-time image prediction of the whole lung, reduce the impact of respiratory motion on lung tissue structure, and provide doctors with a comprehensive, intuitive, and realistic lung scene.

[0007] To achieve the above objectives, the present invention provides the following technical solution:

[0008] A respiratory motion compensation method based on manifold clustering and weighted sparsity is disclosed. The method includes: acquiring 4D-CT images of a population dataset and 3D-CT images of an individual subject during both the expiratory and inspiratory phases; acquiring the deformation vector fields of the population dataset and the individual subject through image registration; obtaining a sub-population dataset with similar motion characteristics to the individual subject based on manifold clustering; establishing a quantitative relationship between the motion characteristics of the individual and the sub-population using a weighted sparsity algorithm to construct the 4D-CT image of the individual subject; acquiring the internal deformation vector field and external skin surface motion of the individual subject to construct an individualized respiratory motion compensation model to achieve real-time prediction of lung images.

[0009] Furthermore, the method specifically includes the following steps:

[0010] Step 1: Obtain 4D-CT images P of the population dataset containing N subjects. i ,i=1...N, where each 4D-CT image contains 3D-CT images of M respiratory phases, as well as 3D-CT images of two phases, Ts: expiratory phase (EE) and inspiratory phase (EI), for the individual subject;

[0011] Step 2: Register the 4D-CT images obtained in Step 1 to a standard template space to obtain a population dataset. Internal deformation vector field corresponding to the exhalation-inhalation phase of an individual subject The standard template space is an expiratory phase (EE) image for each subject;

[0012] Step 3: Perform manifold clustering on the deformable vector field dataset obtained in Step 2 to obtain a subgroup dataset S with similar respiratory movement characteristics to the individual subjects. k ∈(P i (i = 1...N);

[0013] Step 4: A weighted sparsity method is used to study the correlation between the motion characteristics of the deformable vector field of subgroups and individual subjects. The motion characteristics of subgroups with similar respiratory motion features obtained in Step 3 are fused into those of individual subjects, thereby constructing 4D-CT sequence images I1, I2, ..., I of the respiratory phase of the individual subject. M ;

[0014] Step 5: Process the 4D-CT sequence images of the individual subject obtained in Step 4 to extract the internal deformation vector field of the individual subject throughout the entire respiratory cycle. and skin surface movement Establish a quantitative relationship between the two and construct an individualized respiratory motion compensation model to predict lung images in real time.

[0015] Furthermore, in step 3, manifold clustering is performed on the deformable vector field dataset, specifically including the following steps:

[0016] (3a) Using the deformed vector field corresponding to the expiratory-inspiratory phase of an individual subject As a respiratory motion feature, motion features corresponding to the phase are also selected from the group deformation vector field dataset.

[0017] (3b) Through sparse optimization, the deformable vector field is transformed from high-dimensional data to low-dimensional data to obtain sparse vector c. n The optimal solution is obtained; based on this, the weighted vector ω is calculated from the sparse vector. n This represents the difference in motion characteristics between individual subjects and the group dataset, further generating a similarity matrix W = [W0, W1, ..., W...]. n ];

[0018] (3c) Perform k-means-based spectral clustering on the similarity matrix to obtain a subgroup dataset S with similar respiratory movement characteristics to the individual subjects. k ∈(P i (i = 1...k).

[0019] Furthermore, step 4 specifically includes the following steps:

[0020] (4a) The deformable vector field of individual subjects Describing the deformable vector field corresponding to the subgroup dataset With weighted sparsity coefficient ω * The combination;

[0021] (4b) Using the obtained optimal weighted sparsity coefficient ω * Describes the intermediate deformation vector field corresponding to other respiratory phases of an individual subject. The other respiratory phases are those excluding the inspiratory and expiratory phases;

[0022] (4c) Calculate the 4D-CT sequence images I1, I2, ..., I of the individual subject based on the expiratory phase and M-1 intermediate deformation vector fields. M .

[0023] Furthermore, in step 5, the individual subject T is extracted. S Lung deformation vector field throughout the respiratory cycle and skin surface movement Establishing a quantitative relationship between the two and constructing an individualized respiratory movement compensation model includes the following steps:

[0024] (5a) Extracting the lung deformation vector field throughout the respiratory cycle from 4D-CT sequence images of individual subjects and skin surface movement

[0025] (5b) Deep learning was used to study the correlation between skin surface motion and deformation vector field of individual subjects, establish the quantitative relationship between the two, and construct an individualized respiratory motion compensation model.

[0026] The advantages of this invention compared to the prior art are:

[0027] (1) This invention can effectively combine group motion characteristics with individual phases to study the intrinsic relationship between group motion characteristics and individuals, and can track the movement of lung organs and tissue structures in real time, while improving the accuracy and robustness of prediction.

[0028] (2) The respiratory motion compensation method in this invention reconstructs individual 4D-CT images by using two phase images of exhalation and inhalation, which helps to reduce radiation damage to the subject. Attached Figure Description

[0029] Figure 1 This is an overall flowchart of the present invention;

[0030] Figure 2 for Figure 1 The flowchart for step 4 of the weighted sparsity method. Detailed Implementation

[0031] To make the objectives, technical solutions, and effects of this invention clearer and more explicit, the invention is further described in detail below. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. The invention will be described in detail below with reference to specific embodiments and the accompanying drawings.

[0032] like Figure 1As shown, a preferred embodiment of the present invention provides a personalized respiratory motion compensation method based on manifold clustering and weighted sparsity, comprising the following steps:

[0033] Step 1: Acquire 4D-CT images of the population dataset and individual subjects. Preferably, the specific process is as follows:

[0034] 4D-CT images and corresponding expiratory-inspiratory phase images were collected from 40 groups of subjects. Each group's data covered 3D-CT images of M respiratory phases throughout the respiratory cycle (M=12, where EE represents the expiratory phase and EI represents the inspiratory phase). The data were then divided into population datasets (30 groups, P...). i (i = 1...30) and individual subject test sets (10 groups, T j (j = 1...10). The test set uses images from only the exhalation and inhalation phases to build the model, while the remaining 10 intermediate phases are used to validate the model's accuracy.

[0035] Step 2: Preprocess the 4D-CT data from Step 1. Register the 4D-CT images from the population dataset and the individual subject test set to a standard template space (expiratory phase image for each subject) to obtain the lung deformation vector field datasets for the population dataset and the individual subject test set:

[0036] Step 3: Using the deformable vector field dataset obtained in step S2, obtain a subgroup dataset similar to the motion characteristics of individual subjects based on manifold clustering, including the following steps:

[0037] (3a) The model is constructed using two-phase images of exhalation and inhalation from individual subjects in the test set. Define n∈{1,...,30} as the selected population dataset P. i i = 1...30, n = 0 is an individual subject T selected from the individual subject test set. s The expiratory phase image was selected as the reference template. Respiratory features were extracted. Simultaneously, features corresponding to the same phase are selected from the group deformation vector field dataset.

[0038] (3b) By using sparse optimization, the deformable vector field is transformed from high-dimensional data to low-dimensional data, thus transforming it into a sparse optimization problem:

[0039]

[0040] Among them, c n Let Mn and Qn be sparse vectors, and let Mn and Qn be diagonal matrices, defined as follows:

[0041]

[0042] The optimal solution for the obtained sparse vector is calculated. Based on this, from the sparse vector c... n Calculate the weighted vector ω n , representing the difference in characteristics from individual subjects to the group. After calculating the weight vectors of all subjects, a similarity matrix W = [W0, W1, ..., W...] is generated. n ].

[0043] (3c) Perform k-means-based spectral clustering on the similarity matrix to obtain a subgroup dataset S with similar respiratory characteristics to the individual subjects. k ∈(P i (i = 1...30).

[0044] Step 4: Establish the quantitative relationship between individual and subgroup motion characteristics using the weighted sparsity method, construct individual 4D-CT images, and refer to... Figure 2 As shown, it includes the following steps:

[0045] (4a) Obtain individual subject T from step 3 s The deformable vector field of the two phases of exhalation and inhalation is: and the deformable vector field of the corresponding phase in the subgroup The deformed vector field of an individual subject can be described as the deformed vector field corresponding to a subgroup of subjects and the weighted sparsity coefficients ω. * Combinations:

[0046]

[0047]

[0048] Where Sim is the correlation coefficient similarity measure, I EE and I EI It is an image of the exhalation and inhalation phases of an individual subject, where x represents the internal coordinates of the image.

[0049] (4b) By changing the sparse weight coefficient ω * =[ω1,ω2,...,ω k The optimal solution is sought using gradient descent. Each coefficient represents the respiratory motion characteristics from expiration to inspiration for different subjects in a subgroup. A deformation matrix is ​​created using the obtained sparse system to describe the deformation vector field corresponding to the middle M-2 respiratory phases of an individual subject:

[0050]

[0051] in, and These are the deformation vector fields from the expiratory phase to other phases for individual subjects and subgroup subjects, respectively. Based on the expiratory phase and M-2 intermediate deformation fields, the 4D-CT images I1, I2, ..., I of the individual subject are calculated. M The 4D-CT images of the other nine individual subjects in the test set were generated in the same manner.

[0052] Step 5: Construct an individualized respiratory motion compensation model to achieve real-time prediction of lung images, including the following steps:

[0053] (5a) Step S4 acquires the 4D-CT images I1, I2, ..., I of the individual subject. M Extract the lung deformation vector field of an individual subject throughout the entire respiratory cycle. and skin surface movement

[0054] (5b) Preferably, a conditional generative adversarial network is used to establish a respiratory motion compensation model, comprising two parts: a generative model and a discriminative model. The generative model G learns the mapping from the individual subject's skin surface motion ν and random noise vector z to the deformation vector field μ: G:{ν,z}→μ. The discriminative model D classifies between real and synthetic images. The objective function of the model is:

[0055]

[0056]

[0057] in, The expected value is denoted as G. By obtaining the optimal generative model G, a quantitative relationship between the internal deformation vector field and the skin surface motion is established, and an individualized respiratory motion compensation model is constructed.

[0058] (5c) Validate the model on the test set, using the skin surface motion ν corresponding to the test set. P Calculate the internal deformation field μ P Lung images can be predicted in real time using internal deformation fields.

[0059] Following the steps of this preferred embodiment, a subgroup with similar respiratory movements to individual subjects is selected from the population dataset. Global statistical features and individual phase differences are effectively combined to reconstruct individual 4D-CT images. Based on this, a quantitative relationship between the individual lung deformation field and skin surface movement is constructed, resulting in a new respiratory movement compensation model. In this preferred embodiment, the average lung respiratory movement prediction error for 10 test set subjects across 10 respiratory phases is shown in Table 1. The average error for the subjects ranged from 0.16 mm to 0.87 mm, with an average value of 0.43 ± 0.19 mm. The results indicate that this invention can effectively predict lung movement.

[0060] Table 1

[0061] Test set subjects Average value (mm) 1 0.49 2 0.16 3 0.87 4 0.34 5 0.40 6 0.29 7 0.41 8 0.63 9 0.25 10 0.47

[0062] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for respiratory motion compensation based on manifold clustering and weighted sparsity, characterized in that, The method comprises: Step 1: Obtain 4D-CT images P of a population dataset of N subjects i , i = 1...N, where each 4D-CT image contains M 3D-CT images of respiratory phases, and individual subject T s 3D-CT images of two phases, expiration phase (EE) and inspiration phase (EI); Step 2: Register the 4D-CT image obtained in Step 1 to a standard template space to obtain a population deformation vector field dataset and internal deformation vector fields corresponding to the expiratory- inspiratory phase of individual subjects , the standard template space being an expiratory phase (EE) image for each subject; Step 3: manifold clustering is performed on the population deformation vector field dataset obtained in Step 2 to obtain a sub-population deformation vector field dataset with similar breathing motion characteristics as the individual subject ; comprising: (3a) using the T s deformation vector field corresponding to the expiratory-inspiratory phase as a breathing motion feature, simultaneously in the population deformation vector field dataset selecting the motion feature corresponding phase ; (3b) converting the deformation vector field from high-dimensional data to low-dimensional data by sparse optimization to obtain sparse vectors of the optimal solution; on this basis, the weighted vectors representing the difference value of the individual subject to the motion characteristics of the population data set, further generating a similarity matrix ; (3c) performing k-means based spectral clustering on the similarity matrix to obtain a sub-population dataset with similar respiratory motion characteristics for individual subjects ; Step 4: Study the correlation between the motion characteristics of the sub-population and the deformation vector field of the individual subject using the weighted sparse method, and fuse the motion characteristics of the sub-population with similar respiratory motion characteristics obtained in step 3 into the individual subject, so as to construct the 4D-CT sequence images I1, I2,..., I M ; comprising: (4a) a deformation vector field for the individual subject described as a deformation vector field corresponding to the sub-population dataset in combination with weighted sparse coefficients in combination with weighted sparse coefficients (4b) using the obtained optimal weighted sparse coefficients , describing intermediate deformation vector fields corresponding to other breathing phases of the individual subject , the other breathing phases being phases other than the expiration and inspiration phases. (4c) computing a 4D-CT sequence of images I1, I2,..., I of the individual subject from the individual subject's expiration phase and M-1 intermediate deformation vector fields M ; Step 5: processing the 4D-CT sequence images of the individual subject obtained in the step 4, extracting the internal deformation vector field of the individual subject throughout the respiratory cycle and skin surface motion , establishing a quantitative relationship between the two, constructing an individualized respiratory motion compensation model, thereby predicting the lung image in real time, comprising the steps of: (5a) extracting a lung deformation vector field throughout a respiratory cycle from 4D-CT sequence images of an individual subject and skin surface motion ; (5b) using deep learning to study the correlation between the skin surface motion and the deformation vector field of the individual subject, establishing a quantitative relationship between the two, and constructing an individualized respiratory motion compensation model.

Images