CN120436785A - A surgical planning method for generating virtual computed tomography images from cone-beam computed tomography images - Google Patents
A surgical planning method for generating virtual computed tomography images from cone-beam computed tomography imagesInfo
- Publication number
- CN120436785A CN120436785A CN202510274601.XA CN202510274601A CN120436785A CN 120436785 A CN120436785 A CN 120436785A CN 202510274601 A CN202510274601 A CN 202510274601A CN 120436785 A CN120436785 A CN 120436785A
- Authority
- CN
- China
- Prior art keywords
- module
- virtual
- images
- cbct
- computed tomography
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Landscapes
- Apparatus For Radiation Diagnosis (AREA)
Abstract
The invention provides a surgical planning method for generating virtual computed tomography images by cone beam computed tomography images, which belongs to the technical field of virtual CT surgical plan planning, wherein the method generates virtual CT by AI algorithm based on CBCT, and then performs surgical plan planning by the virtual CT, the algorithm comprises three different modules, the first module eliminates the influence on Ji Duijie fruits which can not be completely caused, the second module generates weight characteristic graphs of different tissues to distinguish different anatomical structures, the third module completes the generation of the virtual CT, compared with the CBCT, the image quality of the virtual CT is greatly improved, the virtual CT can inhibit artifacts, restore the fine structure of the image, restore the CT value of the CBCT, and in addition, the quality of an operation scheme planned on the virtual CT is equivalent to that planned on the CT.
Description
Technical Field
The invention belongs to the technical field of virtual CT (computed tomography) operation scheme planning, and particularly relates to an operation planning method for generating a virtual computed tomography image by using cone beam computed tomography images.
Background
Cone beam computed tomography (Cone beam computed tomography, CBCT) is a necessary navigation imaging tool for image guided or robot assisted spine surgery, typically with screw trajectories planned directly on or re-registered to the intraoperative CBCT by the physician, but due to the physical imaging characteristics of CBCT, CBCT images have high artifact, low image contrast, poor image quality, low quality CBCT images can increase uncertainty in screw planning, resulting in wrong placement of screws, causing serious consequences, and therefore, acquiring high quality intraoperative CBCT images is a critical process for image guided or robot assisted spine surgery.
In general, a doctor manually performs operation planning on an intraoperative CBCT, low-quality images are unfavorable factors, or performs operation planning on a preoperative CT, then performs registration fusion on the preoperative CT and the intraoperative CBCT, registers the operation planning in the operation, the scheme has a complex process, needs intervention of the doctor, increases the probability of operation failure due to excessive intermediate joints, and causes two non-linear differences between the CT and the CBCT to be not completely aligned, so that the actual operation planning and the planned operation planning are misplaced.
Disclosure of Invention
In order to solve the defects in the prior art, the invention provides a method for generating a high-quality virtual CT image based on a low-quality CBCT image, completing planning of an operation and robot-assisted spinal surgery on the virtual CT image, reducing adverse effects of CBCT image quality problems in the operation on planning of an operation scheme, planning the operation scheme on the CBCT in the operation by generating the virtual CT, and skipping a registration fusion process.
The invention adopts the following technical scheme for solving the problems:
A surgical planning method for generating virtual computed tomography images by cone beam computed tomography images, which is characterized in that virtual CT is generated by AI algorithm based on CBCT, and then surgical plan planning is performed by the virtual CT, wherein the algorithm comprises three different modules, the first module eliminates influence on Ji Duijie fruits, the second module generates weight characteristic diagrams of different tissues to distinguish different anatomical structures, and the third module completes the generation of the virtual CT.
Further, the first module is Deformable image registration (Reg) module, the second module is Unet model (U) module, and the third module is a Diffusion model (DDPM) module.
Further, for the task of generating a virtual CT image by CBCT, a Reg module is introduced before the DDPM framework, so as to ensure that the CBCT and the CT are aligned as much as possible to prevent the generated virtual CT from deforming on the anatomical structure, thereby eliminating the influence on Ji Duijie.
Furthermore, the Unet module is added before the DDPM module, so that different tissue structures can be displayed, the large noise of CT values in the CBCT image is avoided, and soft tissues and partial cancellous bone cannot be clearly distinguished.
Further, some prior information is used in the initialization phase to generate noise to ensure that the generated image has a certain physiological meaning and does not generate erroneous information.
Further, a more sparse variance table is performed at larger variances to reduce the number of iterations DDPM to avoid degradation of the performance of the model trained using the linear program when skipping the multi-step back-diffusion process during the training of the model.
The invention has the beneficial effects that:
1. Compared with CBCT, the image quality and spatial uniformity of sCT (virtual CT) are greatly improved, and meanwhile, the image anatomical structure is kept unchanged, and the image quality of sCT is obviously superior to CBCT.
2. The sCT can not only inhibit the artifact and restore the fine structure of the image, but also restore the CT value of the CBCT.
3. The quality of the surgical plan planned on sCT is comparable to that planned on CT.
Drawings
In order to more clearly illustrate the embodiments of the present invention, the drawings that are required to be used in the description of the embodiments will be briefly described below, and it will be apparent that the drawings in the description below are some embodiments of the present invention and that other drawings may be obtained from these drawings without inventive effort for a person of ordinary skill in the art.
FIG. 1 is a technical roadmap of the invention;
FIG. 2 is a diagram of three modules involved in the present algorithm;
FIG. 3 is a diagram of virtual CT results;
fig. 4 is a diagram of a virtual CT-based plan result.
Detailed Description
The present invention will be described in detail below with reference to the drawings and the specific embodiments, so that those skilled in the art can better understand the technical solutions of the present invention.
Example 1
As shown in fig. 1, the invention provides a surgical planning method for generating virtual computed tomography images by cone beam computed tomography images, which generates virtual CT by AI algorithm based on CBCT, and performs surgical plan planning by virtual CT, wherein the algorithm includes three different modules, the first module eliminates influence on Ji Duijie fruits, the second module generates weight feature maps of different tissues to distinguish different anatomical structures, and the third module completes generation of virtual CT.
As shown in FIG. 2, the first module is Deformable image registration (Reg) module, the second module is Unet model (U) module and the third module is diffration model (DDPM) module, for the task of generating a virtual CT image by CBCT, the Reg module is introduced before DDPM frame, so as to ensure that CBCT and CT are aligned as much as possible to prevent the generated virtual CT from deforming on an anatomical structure so as to eliminate the influence on Ji Duijie fruit, and the Unet module is added before DDPM module so as to display different tissue structures and avoid the existence of larger noise of CT values in the CBCT image and the incapability of clearly distinguishing soft tissues and partial spongy bone.
Some a priori information is used in the initialization phase to generate noise to ensure that the generated image has a certain physiological meaning and does not produce erroneous information.
A more sparse variance table is performed at larger variances to reduce the number of iterations DDPM to avoid degradation of the performance of the model trained using the linear program when skipping the multi-step back-diffusion process during the training of the model.
Example 2
As a preferred embodiment of the invention, the method comprises the following steps:
1. and generating a virtual CT based on the CBCT, and generating the virtual CT by adopting the de-generation probability model as a base frame.
The denoising diffusion probability model (Denoising Diffusion Probabilistic Models, DDPM) is a generative model that maps between real images and pure noise through a markov diffusion process of T time steps. In the forward direction, a small amount of gaussian noise is repeatedly added to the real image x 0～q(x0) to obtain samples x T of sufficiently large T from the gaussian distribution of the independent co-distribution, making x T approach to gaussian noise, q (x 1:T|x0) is defined as:
Sample x t is obtained by adding i.i.d. (INDEPENDENT IDENTICALLY Distribution, independent co-Distribution) gaussian noise with variance β t at time step t, and according to a variance table By usingOne significant feature of the forward process is to allow sampling of x t in a closed form at any time step t, obtained by scaling the variance of the previous sample x t-1, using the symbols a t∶＝1-βt andWe obtain:
Is a latent variable model shaped as p θ(x0):＝∫pθ(x0:T)dx1:T, where x 1,....,xT is a latent variable having the same dimensions as data x 0～q(x0). The inverse process p θ(x0:t) is defined as slave Beginning markov chain with learned gaussian transitions:
The diffusion model typically operates each back-diffusion step as a mapping through a neural network that provides estimates of mu θ and/or sigma θ, and then trains by minimizing the lower bound of variation of the log-likelihood:
Wherein the method comprises the steps of Representing the expected value on q, p θ is a network approximation of the reverse transition probability, and the bounds can be further broken down into:
where KL represents the Kullback-Leibler divergence. The best way to predict the cumulative noise e θ added to the current intermediate image x t according to the digitization model as in Ho et al, we therefore obtain the following parameterizations of the predicted average μ θ(xt, t):
thus, the loss function is redefined to be a simpler form as follows:
the working method comprises the following steps:
as shown in FIG. 2, the algorithm comprises three different modules, namely a multi-task model Deformable image registration (Reg) module, a Unet model (U) module and a Diffusion model (DDPM) module for completing different tasks, wherein the Reg module eliminates the influence on Ji Duijie fruits, unet model generates weight characteristic diagrams of different tissues to distinguish different anatomical structures, and DDPM completes the generation of sCT.
For the task of CBCT to generate sCT images, it is necessary to align CBCT with CT as much as possible to prevent deformation of the generated sCT on the anatomy, however, it is difficult to align the images completely for both modalities, so we introduce Reg modules before DDPM framework to eliminate the effect that cannot be done completely on Ji Duijie.
Where λ is a regularization parameter, w Reg is a network parameter of the Reg module, phi is an image offset field output by the registration module, L sim is a normalized cross-correlation similarity measure,Ensure smooth changes in phi.
In order to show different tissue structures, a Unet module is added before a DDPM module, and the module can assist in obtaining the products of weight characteristic graphs x psCT,xpsCT and x CBct of different organs on CBCT as the input of a guide DDPM:
w unet is the network parameter of the Unet module, x psCT＝FUnet(xcbct,wUnet) is the weight feature map output by the Unet module, and L Unet is the MSE loss function. The training objectives of RegUDDPM can thus be defined as:
Finally, the optimization targets are as follows:
CBCT Generation sCT from The process of gradually denoising the sample into an image through a series of diffusion processes is started, obviously, the image generated by initializing different noises is different, and has definite physiological significance for different structures and pixel intensities of a CT image, and random initializing noises can cause the generation of variation of the CT structure to generate error information, so that in order to solve the problem, some prior information is used for generating noise in the initialization stage, so that the generated image has a certain physiological significance, and the error information cannot be generated:
Let x 0＝xcbct be the time, in the reverse process Is replaced byTo better preserve the original anatomical formula (12) on the CBCT image.
In addition, according to the literature, the last noise in the forward noise adding process is too large, and the contribution to the sample quality is small, so that the performance of a model trained by using a linear program is not reduced when a multi-step back diffusion process is skipped, and on the basis, an original linear variance table (linear variance table) is changed, namely, a sparse variance table is performed at a larger beta t, so that the iteration times of DDPM are reduced.
In our study t=100, the variance table strategy is as follows: i.e. 40 values are linearly sampled at (0.0001,0.002), I.e. 60 values are sampled linearly over (0.002,0.02).
The pseudo code of the training process is as follows:
The pseudo code for the generation of the sCT process is as follows:
2. planning an operation plan based on virtual CT
(1) Optimal screw trajectory definition
The pedicle screw trajectory can be defined as a line L passing through the point where it must be located in the pedicle and cannot pass out of the vertebral body, so in this study we define the foramen and the soft tissue around the vertebral body as the avoidance area, the optimal screw trajectory should be as far away from the avoidance area as possible to ensure the safety of the screw, thus, let M be any point in the avoidance area, M 0＝(x0,y0,z0) be the screw insertion point on the surface of the vertebral body,Is the unit vector of the screw, and the optimal screw path can be obtained by the following formula:
(2) Screw trajectory planning based on multi-angle projection
In theory, the optimal three-dimensional screw track is known, multi-angle re-projection is performed on the vertebral body around the pedicle, the projection of the optimal screw insertion track can be obtained on the two-dimensional projection surface of each angle, and conversely, the projection of the optimal screw track on the two-dimensional projection is known, and the three-dimensional optimal screw insertion track can be estimated by reconstructing the projections of the optimal screw tracks of different angles.
Therefore, the invention adopts a method of calculating the optimal screw track on the projection surface and then reconstructing the optimal screw track into a three-dimensional screw track to plan the optimal placement track of the screw.
The simulation projection (DIGITALLY RECONSTRUCTED RADIOGRAPH, DRR) can simulate the process that X rays pass through a three-dimensional object and accumulate X-Ray attenuation values falling on a detector plane, the method for generating the simulation projection is mainly divided into two types, namely a Voxel driving (Voxel-driven) projection algorithm, and the other type is a Ray driving (Ray-driven) projection algorithm, compared with the Voxel driving projection algorithm, the result obtained by the Ray driving projection algorithm is more accurate, so that the simulation projection is generated by adopting the parallel light Ray driving projection algorithm in the research, the Siddon algorithm is the most classical algorithm in the Ray driving projection method, the projection P θ under the angle theta is shown as the following formula, ρ θ (i, j, k) is the CT value of a Voxel (i, j, k) where light rays intersect with the projected object, l θ (i, j, k) is the length of the light rays passing through the Voxel, and theta represents the projection angle.
Pθ＝∑i∑j∑klθ(i,j,k)ρθ(i,j,k) (14)
Assuming that P θ is the projection of the CT image at angle θ, the optimal screw placement trajectory in P θ can be determined from ω and b, where ω and b are parameters of the screw trajectory, assuming that M ' = (S| S＝xcosθ+ysinθ, z) is the boundary of the projection of M on P θ, then the two boundaries of M ' can be marked with either-1 (avoidance area) or +1 (safety area), which can be expressed as T= { M ', y|Y= +1or-1}, and then the optimal screw implantation trajectory in P θ can be expressed as follows:
the constraint optimization problem described above can be converted to unconstrained optimization by a lagrangian multiplier:
where α is the Lagrangian multiplier and α is ≡0, it can be solved by Support Vector Machine (SVM) method.
Assuming that P' θ is the θ projection of the optimal screw implantation trajectory, where the point on the optimal screw implantation trajectory is marked as 1 and the other points are marked as 0, the optimal screw implantation trajectory can be obtained by (equation 14) and (equation 15), so that a three-dimensional stereo image, that is, a three-dimensional screw trajectory of the screw, can be reconstructed layer by layer using a filtered back projection method in the reconstruction process, and the filtered back projection algorithm is shown in the following equation.
In the above formula, i is the ith layer of the three-dimensional image, P' iθ(s) represents the projection of a screw track of the ith layer under the angle theta, omega is a slope filter, f i (x, y) is the three-dimensional image containing the screw track, in fact, the final result is a binary three-dimensional image only containing the screw track, the image value at the screw track is 1, the other image value is 0, the filtering process can be omitted during reconstruction, and the final three-dimensional screw optimal track can be obtained by simplifying the following modes:
(3) Screw trajectory planning simplification
In fact, the end result is a binary three-dimensional image containing only the screw trajectory, where the image value is 1 and the others are 0, so in practical deployment we can estimate the screw trajectory by only two projections, since the screw trajectory is simple enough as a line segment, in this work we project with the horizontal directionAnd sagittal plane direction projectionTo estimate the optimal screw implantation trajectory, assuming M 'axial and M' sagittal are each M inAndBoundary points on the projection can then be usedThe sum of the optimal screw implantation trajectory parameters { omega axial,baxial }, aboveIn the model, the optimal screw implantation track parameter { omega sagittal,bsagittal } is used for determining the three-dimensional screw track
More succinctly, formula (19) can be equivalently represented as
[y,z]·ωsagittal T+bsagittal=0
[x,z]·ωaxial T+baxial=0 (20)
Where x, y and z represent coordinates on the CT image.
As shown in FIG. 3, the virtual CT result graph has greatly improved image quality, and the virtual CT can not only inhibit artifacts and restore the fine structure of the image, but also restore the CT value of the CBCT.
The plan result based on the virtual CT is shown in fig. 4, and the screw obtained by planning is in the clinical safety range and can be directly applied to clinical practice.
The present invention has been described in detail by way of examples, but the description is only of the preferred embodiments of the present invention and should not be construed as limiting the scope of the invention. All equivalent changes and modifications within the scope of the present invention are intended to be covered by the present invention.
Claims (6)
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202510274601.XA CN120436785A (en) | 2025-03-10 | 2025-03-10 | A surgical planning method for generating virtual computed tomography images from cone-beam computed tomography images |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202510274601.XA CN120436785A (en) | 2025-03-10 | 2025-03-10 | A surgical planning method for generating virtual computed tomography images from cone-beam computed tomography images |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN120436785A true CN120436785A (en) | 2025-08-08 |
Family
ID=96622864
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202510274601.XA Pending CN120436785A (en) | 2025-03-10 | 2025-03-10 | A surgical planning method for generating virtual computed tomography images from cone-beam computed tomography images |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN120436785A (en) |
-
2025
- 2025-03-10 CN CN202510274601.XA patent/CN120436785A/en active Pending
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| US12387395B2 (en) | System and method for artifact reduction in an image | |
| EP3525171B1 (en) | Method and system for 3d reconstruction of x-ray ct volume and segmentation mask from a few x-ray radiographs | |
| US8897514B2 (en) | Imaging method for motion analysis | |
| Abdolali et al. | Automatic segmentation of mandibular canal in cone beam CT images using conditional statistical shape model and fast marching | |
| Tomazevic et al. | 3-D/2-D registration by integrating 2-D information in 3-D | |
| CN106920246B (en) | Uncertainty map for segmentation in the presence of metal artifacts | |
| US20070167784A1 (en) | Real-time Elastic Registration to Determine Temporal Evolution of Internal Tissues for Image-Guided Interventions | |
| WO2024165508A1 (en) | A computer-implemented method, device, system and computer program product for processing anatomic imaging data | |
| US11989830B2 (en) | Imaging reconstruction system and method | |
| WO2012017375A2 (en) | In-plane and interactive surface mesh adaptation | |
| CA2831079A1 (en) | Model-based processing of image data | |
| Baka et al. | Statistical shape model-based femur kinematics from biplane fluoroscopy | |
| Rossi et al. | Image‐based shading correction for narrow‐FOV truncated pelvic CBCT with deep convolutional neural networks and transfer learning | |
| CN114558251B (en) | Automatic positioning method, device and radiotherapy equipment based on deep learning | |
| CN117611695A (en) | Dental cone beam CT metal artifact correction method and device based on diffusion model | |
| Birklein et al. | Motion correction for separate mandibular and cranial movements in cone beam CT reconstructions | |
| CN117522747A (en) | Metal artifact correction method and system for CT image | |
| Zollei | 2D-3D rigid-body registration of X-ray Fluoroscopy and CT images | |
| Thies et al. | A gradient-based approach to fast and accurate head motion compensation in cone-beam CT | |
| Sadowsky et al. | Hybrid cone-beam tomographic reconstruction: Incorporation of prior anatomical models to compensate for missing data | |
| CN120436785A (en) | A surgical planning method for generating virtual computed tomography images from cone-beam computed tomography images | |
| CN120500705A (en) | CBCT simulation for training AI-based CT-to-CBCT registration and CBCT segmentation | |
| KR102248191B1 (en) | Method and apparatus for correcting cone-beam artifacts in a cone-beam computed tomography image, cone-beam computed tomography apparatus including the same | |
| Balogh et al. | Comparison of iterative reconstruction implementations for multislice helical CT | |
| Bora et al. | Digitally reconstructed radiograph generation for enabling AI/ML in medical imaging |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination |