A radiotherapy patient positioning method and apparatus based on vertebral body recognition

Abstract

The application relates to a radiotherapy patient positioning method and device based on vertebral body recognition, which comprises the following steps: S1, acquiring a planning image and a positioning image of a vertebral body, inputting nn-Unet for feature extraction, and obtaining a planning image vertebral body activation map and a positioning image vertebral body activation map; S2, performing post-processing on each vertebral body activation map to obtain a planning image vertebral body recognition positioning map and a positioning image vertebral body recognition positioning map, wherein the post-processing comprises limiting a cone sequence, fine-tuning a vertebral position and searching for a missing vertebral body; S3, performing global registration on the positioning map through affine transformation and performing local registration by adopting a deformable image registration network based on B-splines; S4, taking similarity measurement and point measurement based on a cone positioning as optimization targets, updating parameters in a registration process through an optimizer, and obtaining a final registration result, and generating a radiotherapy patient positioning adjustment instruction according to the final registration result. Compared with the prior art, the application provides a fast and fully automatic tool to meet the requirements of high speed and high precision of radiotherapy patient positioning, and guarantees the safety of radiotherapy under high flux.

Description

Technical Field

[0001] This invention belongs to the field of medical image processing technology, and in particular relates to a method and device for positioning radiotherapy patients based on spinal body recognition. Background Technology

[0002] Radiation therapy is a crucial method in cancer treatment. By using targeted radiation beams to destroy tumor cells, it can prolong patient survival, improve quality of life, and help some patients achieve partial or complete cancer remission. During radiation therapy, ensuring the radiation beam is correctly aimed at the patient's tumor site and minimizing radiation exposure to surrounding healthy tissues is essential. The spine serves as a reference point, helping doctors determine the correct treatment location; therefore, accurate identification and positioning of the spinal vertebrae is a vital step in achieving this goal.

[0003] Traditional methods involve manual or semi-manual spinal marking and identification, which are prone to subjective judgment errors. Furthermore, due to low image quality, small imaging fields of view, and the complex structure of the spine, as well as the possibility of metal implants, inaccurate vertebral identification and positioning during radiotherapy frequently lead to medical accidents, increasing patient risk. Radiotherapy includes planning images and localization images. Planning images are used to develop radiotherapy plans, determine treatment protocols, and predict treatment outcomes; localization images are used to confirm the patient's position before each treatment session and adjust positioning promptly to ensure accuracy. Image registration between planning and localization images, and adjusting patient positioning based on the registration results, is a crucial means of achieving precise radiotherapy. Typically, compared to planning images, localization images have lower image quality, incomplete image information, and inconsistent intensity between localization and planning images. Traditional image registration techniques are not ideal for registering radiotherapy localization and planning images. Chinese patent application 202311295654.7 discloses a spinal registration method. This method involves 2D-3D registration of preoperative CT images and intraoperative X-ray images, using the registration results and a DRR algorithm to obtain the coordinates of preoperatively selected coarse registration points on the X-ray images. Based on triangulation, the coarse registration points on the X-ray images of the surgical site are reconstructed in 3D and coarsely registered with the preoperatively selected coarse registration points. Then, the preoperatively segmented spinal model is downsampled, and the downsampled points are projected onto the X-ray images. A projection matrix is ​​then used for 3D reconstruction to obtain a spinal point cloud in the array coordinate system of the surgical site. This point cloud is then registered with the preoperatively segmented spinal model using an ICP point cloud registration algorithm to obtain the final spinal registration result. This method establishes the relationship between the array coordinate system of the surgical site and the X-ray image coordinate system through an image calibration module. It is independent of the parameters of the C-arm itself and can be adapted to any type of C-arm for robotic surgical navigation. Furthermore, the registration method provided by this invention is based on a single spine, avoiding the problem of patient posture changes during preoperative CT and intraoperative X-ray imaging. However, due to the sensitivity limitations of the DRR algorithm, this method may produce registration errors. Therefore, it is necessary to design a radiotherapy patient positioning method based on spinal body recognition to improve the registration accuracy of spinal images and ensure the safety of radiotherapy. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings of the prior art by providing a method and device for positioning radiotherapy patients based on spinal body recognition, thereby improving the registration accuracy of spinal images and ensuring the safety of radiotherapy.

[0005] The objective of this invention can be achieved through the following technical solutions:

[0006] A method for positioning radiotherapy patients based on spinal body recognition includes the following steps:

[0007] S1. Obtain the planning image and localization image of the vertebral body of the radiotherapy patient, and input them into nn-Unet for feature extraction to obtain the vertebral body activation map of the planning image and the vertebral body activation map of the localization image.

[0008] S2. Post-process each vertebral activation image to obtain a vertebral identification and positioning image of the planned image and a vertebral identification and positioning image. The post-processing includes restricting the vertebral order, fine-tuning the vertebral position, and finding any missing vertebrae.

[0009] S3. The planned image vertebral body identification and positioning map and the positioning image vertebral body identification and positioning map are globally registered through affine transformation, and local registration is performed using a deformable image registration network based on B-splines.

[0010] S4. Using similarity measurement and cone-based point measurement as optimization objectives, the parameters in the registration process of step S3 are updated through the optimizer to obtain the final registration result. Based on the final registration result, the radiotherapy patient positioning adjustment instructions are generated.

[0011] Furthermore, the specific process of step S2 is as follows:

[0012] S201. Find the centroid of each vertebral cross-section in each vertebral activation image, obtain multiple centroid points, connect all centroid points to obtain the spinal centerline of each vertebral activation image, and extract one-dimensional vertebral activation data.

[0013] S202. Construct an energy function, and perform multiple post-processings on the results obtained in step S201 until the value of the energy function is the lowest, thereby generating a vertebral body center signal.

[0014] S203. Map the vertebral body center signal back to each vertebral body activation map to obtain the planned image vertebral body identification and positioning map and the positioning image vertebral body identification and positioning map.

[0015] Furthermore, the energy function is defined as follows:

[0016]

[0017] Wherein, L(v) l (k) is the energy function, v l For the topmost vertebral label, k i Let i be the position of the i-th vertebra. For vertebral body weight, For an absolute index v = v l +i is the activation value of the vertebral body, R is the regularization term, and N is the number of detected vertebral bodies.

[0018] Furthermore, the process of restricting the cone order involves exhaustively searching for the label v of the apex cone.l Optimize.

[0019] Furthermore, the process of fine-tuning the spinal position involves using a climbing algorithm to adjust the position k of the i-th vertebra. i Optimize.

[0020] Furthermore, the process of finding missing vertebrae involves inserting a new vertebral center between two adjacent vertebrae and searching for the insertion position by minimizing the energy function to find unmarked vertebrae.

[0021] Furthermore, in step S4, mutual information is used as a similarity measure.

[0022] Furthermore, in step S4, point measurement is performed based on the average Euclidean distance.

[0023] Furthermore, in step S4, the optimizer is an iterative stochastic gradient descent optimizer.

[0024] The present invention also provides an electronic device, including a memory, a processor, and a program stored in the memory, wherein the processor executes the program to implement the above-described method.

[0025] Compared with the prior art, the present invention has the following beneficial effects:

[0026] 1. This invention performs post-processing on the activation map obtained after feature extraction from the nn-Unet network, which can more accurately identify and locate the vertebral bodies in the image. Compared with traditional manual or semi-manual methods, this invention greatly reduces subjectivity and human error, and improves the accuracy of vertebral body identification. This invention designs post-processing steps based on the structural features of the vertebral body, including restricting the order of vertebrae, fine-tuning the position of the vertebrae, and finding missing vertebrae. These steps are used to ensure that the identified vertebral labels conform to the order of the real vertebrae, adjust the position of the vertebrae, and identify whether there are any omissions in the identification results of the network. This can further improve the identification results and help overcome the influence of factors such as changes in patient position and the diversity of vertebral morphology on the identification accuracy.

[0027] 2. This invention performs global registration of the planned image vertebral body recognition and positioning image and the positioning image vertebral body recognition and positioning image through affine transformation, so that the planned image and the positioning image are initially aligned and maintain spatial consistency; a deformable image registration network based on B-splines is used for local registration, and the similarity measurement incorporates the information of vertebral body recognition; by combining the two, the accuracy and precision of registration can be improved.

[0028] 3. Since the intensity and pixel count of the localization image and the planning image are different, this invention uses nnU-Net for feature extraction. nnU-Net can automatically adjust all hyperparameters according to the attributes of the given dataset. The whole process does not require manual intervention, reducing the dependence on parameter tuning. This invention also uses mutual information as a similarity measure, which is applicable to situations with different pixel intensities, improving the practicality of this invention and giving it broad application prospects. Attached Figure Description

[0029] Figure 1 This is a flowchart of the method of the present invention;

[0030] Figure 2 The activation map of the spinal centerline obtained by nn-Unet;

[0031] Figure 3 A diagram for vertebral body identification and localization.

[0032] (3a) is the vertebral body identification and positioning map of the planned image, and (3b) is the vertebral body identification and positioning map of the positioning image;

[0033] Figure 4 This is a schematic diagram of the final registration result. Detailed Implementation

[0034] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. These embodiments are based on the technical solution of the present invention and provide detailed implementation methods and specific operating procedures. However, the scope of protection of the present invention is not limited to the following embodiments.

[0035] Example:

[0036] This embodiment provides a method for positioning radiotherapy patients based on spinal body recognition, such as... Figure 1 As shown, it includes the following steps:

[0037] S1. Obtain the planning image and localization image of the vertebral body of the radiotherapy patient, and input them into nn-Unet for feature extraction to obtain the vertebral body activation map of the planning image and the vertebral body activation map of the localization image.

[0038] nnU-Net is a medical image segmentation framework based on deep learning and neural network technology that adapts to any new dataset. This framework automatically adjusts all hyperparameters according to the attributes of a given dataset, without manual intervention. The input planning and localization images are used to initially identify the location and structure of the patient's vertebrae. The vertebrae are automatically detected in the images, and the probability of each voxel corresponding to each vertebral center is obtained, i.e., a vertebral activation map. This method can identify 7 cervical vertebrae (C1-C7), 12 thoracic vertebrae (T1-T12), 5 lumbar vertebrae (L1-L5), and 2 sacral vertebrae (S1, S2). Therefore, the activation map consists of 26 3D images.

[0039] S2. Post-process each vertebral activation image to obtain the planned image vertebral identification and positioning image vertebral identification and positioning image. The post-processing includes restricting the vertebral order, fine-tuning the vertebral position, and finding any missing vertebrae.

[0040] After obtaining the 3D vertebral activation map, the centerline of the spine is extracted and merged along the centerline to generate a 1D vertebral activation signal. The specific process is as follows: 26 activation maps of a single vertebra are merged into one activation map. This represents the probability of vertebral body centers without distinguishing labels. Due to similar image patterns, individual activation maps often incorrectly activate in the wrong vertebrae, but these activations usually only appear near the vertebral body center, so merged activation maps can typically encompass all vertebral centers. This is then achieved by calculating... For each 2D cross-sectional data point, the centroid is obtained, resulting in a series of centroid points. These centroid points are then connected to form a series of centroid points. Figure 2 The central axis of the spine is shown. (Activation diagram) 1D data along the spinal centerline represents the probability that the center of the vertebral body is located at a given position on the spinal centerline.

[0041] For 1D vertebral activation signals Let N represent the number of vertebrae detected, and v l Represented as the lowest label (the highest vertebral label), since the detected vertebrae must be anatomically continuous, their indices can be represented as [v l ,v l +N-1]. The detected vertebral body position is represented as k = {k i}, i = [0, N-1], where i is relative to label v l The index of k. Therefore, k i This indicates that it has an absolute index (v l The position of the vertebral body is given by (+i). The energy function is defined as follows:

[0042]

[0043] Wherein, L(v) l (k) is the energy function. For an absolute index v = v l +i is the activation value of the vertebra, and R is a regularization term used to limit the distance between adjacent vertebrae and encourage the distance between adjacent vertebrae to remain similar, expressed as: R(a,b)=exp(max(a / b,b / a)). The weights for the vertebral bodies are set to 2. Since the vertebral bodies located at both ends of the spine (C1, C2, S1, S2) have a more prominent appearance, the two vertebral bodies at both ends of the spine (C1: C1, C2: C2, S1: first sacrum, S2: last sacrum) are considered anchor vertebrae throughout the experiment, and their weights are set to 2. The weights for the other vertebrae are set to 1. It should be noted that since N can be represented by k, N can be removed from the parameters for symbol simplification.

[0044] Parameter (v) l ,k) search in the following space: v l ∈[1,26],k i ∈[0,L]. Let v l Initialize to 1, and initialize the vertebral body center k to the 1D vertebral body activation signal. Find the coordinates of the local maxima, and ensure that k i+1 >k i After initialization, the parameters are updated iteratively using three post-processing methods: restricting the order of the cones, fine-tuning the position of the spines, and finding any missing spines.

[0045] The process of restricting the order of the cones is to exhaustively search for the label v of the apex cone. l Optimize:

[0046]

[0047] The process of fine-tuning the spine position involves adjusting the vertebral body center to minimize the activation probability and the total energy of distance regularization. This is achieved by using a climbing algorithm to adjust the index k of the i-th vertebra. i Optimize:

[0048]

[0049] The process of finding a missing vertebra involves inserting a new vertebral center between two adjacent vertebral bodies k(u,u+1):

[0050]

[0051] The insertion position u is searched by minimizing the energy function to find unlabeled missing vertebrae.

[0052]

[0053] These three operations are applied iteratively until the energy function begins to increase (i.e., convergence). During this process, the parameter (v) associated with the lowest energy... l k) is considered as the optimized output. After obtaining the vertebral body centers located in the 1D signal, their coordinates are mapped back to the 3D image to generate the final 3D localization result, as shown below. Figure 3The vertebral body identification and localization diagram shown in (3a) and as follows Figure 3 The positioning image shown in (3b) is a vertebral body identification and positioning diagram.

[0054] S3. The planned image vertebral body identification and positioning map and the positioning image vertebral body identification and positioning map are globally registered through affine transformation to perform initial alignment of the target image and the source image; then, a deformable image registration (DIR) network based on B-splines is used for local registration to improve the initial alignment result.

[0055] S4. Using similarity measurement and cone-based point measurement as optimization objectives, the parameters in the registration process of step S3 are updated by iterative stochastic gradient descent optimizer to obtain the final registration result. Based on the final registration result, the radiotherapy patient positioning adjustment instructions are generated.

[0056] Image registration aims to find a transformation T that transforms the source image I... M Spatially aligned to the target image I F For intensity-based image registration, optimal alignment is typically determined by optimizing a similarity metric between image intensities. In this invention, considering that planning and localization images may be obtained from different devices, mutual information (MI) is used as the similarity metric because it is applicable to cases where corresponding anatomical structures have different pixel intensities. Furthermore, considering the similarity of adjacent vertebral structures, spinal registration misalignment may occur during actual radiotherapy. Therefore, a point metric is introduced based on the spinal center point positions of the two obtained vertebral identification and localization images. The point metric is used to minimize the distance between two point sets with known correspondences, defined as the average Euclidean distance:

[0057]

[0058] Where P is point x i Quantity, and These are the corresponding points from the target image and the source image point sets, respectively. The point set is the center point of the corresponding cone detected in the two images.

[0059] The final registration result is as follows Figure 4 As shown.

[0060] If the above methods are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0061] The above description of the embodiments is provided to enable those skilled in the art to understand and use the invention. It will be apparent to those skilled in the art that various modifications can be made to these embodiments, and the general principles described herein can be applied to other embodiments without inventive effort. Therefore, the present invention is not limited to the above embodiments, and any improvements and modifications made by those skilled in the art based on the disclosure of the present invention without departing from the scope of the invention should be within the protection scope of the present invention.

Claims

1. A method for positioning radiotherapy patients based on spinal body recognition, characterized in that, Includes the following steps: S1. Obtain planning and localization images of the vertebral body of the radiotherapy patient, input them into nn-Unet for feature extraction, and obtain vertebral activation maps of the planning and localization images. The planning and localization images are used to initially identify the position and structure of the patient's vertebral body, automatically detect the vertebral body in the images, and obtain the probability of each voxel corresponding to each vertebral center, i.e., the vertebral activation map. S2. Post-process each vertebral activation image to obtain a vertebral identification and positioning image of the planned image and a vertebral identification and positioning image. The post-processing includes restricting the vertebral order, fine-tuning the vertebral position, and finding any missing vertebrae. S3. The planned image vertebral body identification and positioning map and the positioning image vertebral body identification and positioning map are globally registered through affine transformation, and local registration is performed using a deformable image registration network based on B-splines. S4. Using mutual information as a similarity measure, and the distance between the spine center point positions of the planned image vertebral body identification and positioning map and the positioning image vertebral body identification and positioning map as a point measure, the similarity measure and point measure are used as optimization targets. The parameters in the registration process of step S3 are updated through the optimizer to obtain the final registration result. The radiotherapy patient positioning adjustment instruction is generated based on the final registration result.

2. The method for positioning radiotherapy patients based on spinal body recognition according to claim 1, characterized in that, The specific process of step S2 is as follows: S201. Find the centroid of each vertebral cross-section in each vertebral activation image, obtain multiple centroid points, connect all centroid points to obtain the spinal centerline of each vertebral activation image, and extract one-dimensional vertebral activation data. S202. Construct an energy function, and perform multiple post-processings on the results obtained in step S201 until the value of the energy function is the lowest, thereby generating a vertebral body center signal. S203. Map the vertebral body center signal back to each vertebral body activation map to obtain the planned image vertebral body identification and positioning map and the positioning image vertebral body identification and positioning map.

3. The method for positioning radiotherapy patients based on spinal body recognition according to claim 2, characterized in that, The energy function is defined as follows: in, Let be the energy function. The label for the topmost vertebral body. For the first The location of each vertebra. For vertebral body weight, For those with absolute indexes The activation value of the vertebral body, For regularization terms, The number of vertebrae detected.

4. The method for positioning radiotherapy patients based on spinal body recognition according to claim 2, characterized in that, The process of restricting the order of the cones involves exhaustively searching for the label of the apex cone. Optimize.

5. A method for positioning radiotherapy patients based on spinal body recognition according to claim 2, characterized in that, The process of fine-tuning the spine position involves using a climbing algorithm to adjust the first... The position of each vertebra Optimize.

6. A method for positioning radiotherapy patients based on spinal body recognition according to claim 2, characterized in that, The process of finding missing vertebrae involves inserting a new vertebral center between two adjacent vertebrae and searching for the insertion position by minimizing the energy function to find unmarked vertebrae.

7. A method for positioning radiotherapy patients based on spinal body recognition according to claim 1, characterized in that, In step S4, point measurement is performed based on the average Euclidean distance.

8. A method for positioning radiotherapy patients based on spinal body recognition according to claim 1, characterized in that, In step S4, the optimizer is an iterative stochastic gradient descent optimizer.

9. An electronic device comprising a memory, a processor, and a program stored in the memory, characterized in that, When the processor executes the program, it implements the method as described in any one of claims 1-8.

Images