A method and apparatus for CBCT sparse angle reconstruction based on a physically-perceived residual-guided diffusion model.

Abstract

This invention discloses a CBCT sparse angle reconstruction method and apparatus based on a physically-perceptive residual-guided diffusion model, belonging to the field of medical image reconstruction technology. Addressing the shortcomings of existing sparse-view CBCT reconstruction methods, such as numerous image artifacts, severe detail loss, and slow reconstruction speed, this invention employs a multi-scale feature extraction network combined with a residual-guided mechanism to optimize the sampling process of the diffusion model. It introduces physically-perceptive constraints based on the BP imaging algorithm, simulating the physical process of real CBCT imaging to enhance the consistency between the reconstructed image and the actual anatomical structure. Furthermore, it improves model inference efficiency based on the DDIM acceleration strategy. This invention significantly reduces artifacts in the reconstructed image while lowering the projection dose, improving image detail resolution and meeting the high-precision, high-efficiency application requirements of real-time intraoperative guidance in clinical practice. It has good clinical translational value and can be widely applied in CBCT imaging-related medical clinical scenarios such as tumor radiotherapy localization, orthopedic surgical guidance, and orthodontic diagnosis.

Description

Technical Field

[0001] This invention relates to the field of medical image reconstruction technology, specifically to a CBCT sparse angle reconstruction method and apparatus based on a physical perception-guided diffusion model. Background Technology

[0002] Cone-beam computed tomography (CBCT) is an imaging technology that combines miniaturization, high scanning efficiency, and adjustable dosage, and has become an indispensable technique in fields such as medical imaging diagnosis, orthodontic treatment, and non-destructive testing of industrial components. Sparse angle reconstruction, as a core approach to optimizing the clinical application value of CBCT, can effectively improve the clinical adaptability of the equipment by reducing the number of projection angles, thereby lowering radiation exposure and shortening scan time. However, the core bottleneck of this technical approach lies in the fact that undersampling of projection data can cause severe image artifacts and loss of detail, directly limiting the application value of the reconstruction results in accurate diagnosis and quantitative analysis.

[0003] For high-quality reconstruction in sparse-angle CBCT, the FDK algorithm is the most widely used among traditional analytical algorithms. Proposed by Feldkamp, ​​Davis, and Kress, this algorithm, with its simple implementation and efficient computational power, has become the industry standard for cone-beam computed tomography (CBCT) reconstruction. When the projection angle is sufficient, the FDK algorithm can generate high-quality 3D images through filtering and weighted backprojection. Subsequently, many scholars have improved the FDK algorithm to address its shortcomings: Katsevich's algorithm is precisely optimized for helical CT scans, improving reconstruction accuracy in this scenario; Li et al.'s C-FDK algorithm innovatively uses curve filtering to optimize the projection resampling process, thereby improving the quality of the reconstructed image. However, these analytical algorithms share a common weakness: they require high integrity of the projection data. In sparse-angle cases, due to the scarcity of effective projection information, the reconstructed image suffers from blurred details and increased artifacts, making it difficult to meet the stringent standards of clinical diagnosis and precision industrial testing.

[0004] In recent years, many scholars have applied diffusion models to denoising or reconstruction of CBCT images. Peng J et al. proposed a conditional denoising diffusion probability model. Utilizing a temporally embedded U-Net architecture, combined with residual and attention modules, they progressively transformed white Gaussian noise samples into a target CT distribution based on CBCT. The reconstructed image showed superior performance in both visual quality and quantitative analysis. Choi Y et al. proposed a novel method for eliminating metal artifacts in CBCT images. They used two diffusion models to model the metal-free CBCT prior and the metal artifact prior, respectively, and enhanced the three-dimensional connectivity of the reconstructed image through dual-domain guidance in the image and projection domains. Xie J et al. proposed a prior frequency-guided diffusion model, using a conditional diffusion model as a regularization term for the reconstruction of CBCT images acquired at finite angles. This condition uses patient-specific prior CT scans as input and extracts high-frequency information from them to guide the denoising process of the diffusion model.

[0005] Closely related to this invention is the work of Liu et al. on residual diffusion networks. The core innovation of this method lies in simultaneously introducing residual and noise information during the forward pass of the diffusion model, constructing a directional diffusion path from the target image to the degraded input image. This provides clear guidance for the inverse generation process of image restoration. Furthermore, it combines the DDIM (Denoising Diffusion Implicit Models) strategy to achieve efficient acceleration of diffusion sampling. However, this method is only used for applications such as removing lens raindrops and object shadows, and fails to embed the physical imaging mechanism.

[0006] While traditional pure diffusion models can output high-quality reconstructed images thanks to their powerful generative capabilities, their core flaw lies in confusing the essential task boundaries between "image reconstruction" and "noise removal." These models are essentially end-to-end black-box architectures, relying solely on data-driven methods to learn image feature maps. They completely fail to incorporate prior physical knowledge of CBCT imaging, thus failing to reflect real imaging principles such as X-ray attenuation and projection geometry, and struggling to explain the intrinsic logic of the reconstruction process. Summary of the Invention

[0007] This invention addresses the shortcomings of existing diffusion network-based image reconstruction methods by providing a CBCT sparse angle reconstruction method and apparatus based on a physically-aware residual-guided diffusion model. This invention embeds CBCT imaging physical priors into the diffusion model framework and integrates a physical-aware module based on back-projection (BP) imaging. It not only leverages the advantages of the diffusion model to preserve image details but also relies on physical priors to ensure the anatomical rationality and numerical accuracy of the reconstruction results. Ultimately, it achieves a significant improvement in reconstruction efficiency while maintaining high-quality reconstruction results, fully meeting the cautious requirements of clinical diagnosis for imaging accuracy and interpretability.

[0008] To achieve the above objectives, the present invention adopts the following technical solution:

[0009] On the one hand, this invention provides a CBCT sparse angle reconstruction method based on a physically-aware residual-guided diffusion model, comprising the following steps:

[0010] 1. Obtain sparse view CBCT projection data and preprocess the projection data;

[0011] 2. Construct a residual-guided diffusion model and embed a physical sensing module based on the BP imaging algorithm;

[0012] This step is the core of the invention, aiming to solve the problem of lack of physical consistency in pure data-driven diffusion models by fusing imaging physics priors and diffusion models.

[0013] First, the basic architecture of the residual-guided diffusion model is constructed, which includes a forward diffusion process and a backward diffusion iterative process. In the forward diffusion process, unlike traditional diffusion models that only add Gaussian noise, this invention simultaneously adds Gaussian noise and residual information to real CT images acquired by high-precision CT equipment to simulate the image degradation process under sparse views, providing a clear optimization target for the backward iteration.

[0014] Secondly, a physical perception module based on the BP imaging algorithm is embedded in the model. This module provides physical constraints for the diffusion model by simulating the physical processes of projection geometry and X-ray attenuation in real CBCT imaging. During the back-diffusion iteration process, this module first calculates the difference between the projection of the image generated in the current iteration and the true projection to obtain the projection residual; then, it back-projects the projection residual using the BP imaging algorithm to generate a residual BP image; finally, it adds this residual BP image to the input of the current iteration, guiding the model to accurately estimate the noise and residual of the current iteration, thus achieving refined image reconstruction under the constraints of physical laws.

[0015] 3. The residual-guided diffusion model is trained using coarsely reconstructed CBCT images and real CT images;

[0016] Specifically, the training samples in this step consist of coarsely reconstructed CBCT images and corresponding real CT images. The coarsely reconstructed CBCT images are obtained by performing preliminary reconstruction on preprocessed sparse view projection data using the FDK algorithm or a lightweight NeRF network, providing image priors that meet the basic structural requirements. The real CT images serve as labels, providing accurate supervision information for model training.

[0017] During training, coarsely reconstructed CBCT images are input into a pre-built residual-guided diffusion model. Reconstructed images are generated through alternating forward diffusion and backward iteration. The model parameters are then optimized via backpropagation by calculating the differences between the generated images and the actual CT images. A multi-timestep loss function can be used during training to comprehensively constrain the noise estimation accuracy and residual fitting effect at each diffusion stage, ensuring high-quality reconstruction capabilities after model convergence.

[0018] 4. During the inference process, inputting a coarsely reconstructed CBCT image will yield a finely reconstructed image;

[0019] Specifically, in the model inference stage, the sparse view CBCT projection data collected clinically is first processed through the preprocessing procedure in step 1, and then a coarsely reconstructed CBCT image is generated using a coarse reconstruction method; subsequently, the coarsely reconstructed CBCT image is input into the trained residual-guided diffusion model.

[0020] During the inverse iteration process, the model calculates the backpropagation (BP) image of the projection residual in real time through the physical sensing module and applies guidance. By adjusting the iteration step size of the diffusion process, the number of iterations is reduced, improving inference efficiency. Ultimately, the model outputs finely reconstructed CBCT images with high detail and high physical consistency, meeting the application needs of clinical diagnosis or industrial testing.

[0021] Furthermore, the preprocessing includes dark field correction, geometric correction, and noise removal on the original sparse angle CBCT projection data.

[0022] Furthermore, the lightweight NeRF network is DIFNet.

[0023] Furthermore, the forward diffusion process simultaneously adds Gaussian noise and residual information to the real CT image, as detailed below:

[0024]

[0025] in, The degraded image obtained during the forward diffusion process at step t, For coarse reconstruction of CBCT images, The difference between the real image and the original image. It is standard Gaussian noise. They are the first The cumulative retention factor of the residual and the cumulative retention factor of the noise step.

[0026] Furthermore, the aforementioned From single-step coefficients It is obtained by recursion, which satisfies , .

[0027] Furthermore, the reverse diffusion iteration process of the residual-guided diffusion model can be represented as follows:

[0028]

[0029] in, , To generate process control coefficients, when At that time, the reverse process is random; when hour, If so, then the reverse process is deterministic.

[0030] Furthermore, in the reverse iterative diffusion process of the residual-guided diffusion model, each iteration involves a deep learning network estimating and removing noise and residual components from the current iteration step to generate an intermediate image for the next iteration step. The deep learning network is either a UNet or a Vision Transformer network structure. Further, the reverse iterative process of the residual-guided diffusion model can employ a DDIM acceleration strategy to improve inference efficiency. This DDIM acceleration strategy adjusts the iteration step size of the diffusion process, skipping some intermediate diffusion time steps and directly estimating the image state at key time steps, significantly reducing the number of iterations required for model inference. This strategy improves model inference speed by 10-100 times without sacrificing the quality of the reconstructed image.

[0031] On the other hand, the present invention provides a CBCT sparse angle reconstruction device based on a physically-aware residual guided diffusion model, comprising one or more processors and a memory, wherein the memory is coupled to the processor; wherein the memory is used to store program data, and the processor is used to execute the program data to implement the above-mentioned CBCT sparse angle reconstruction method based on a physically-aware residual guided diffusion model.

[0032] On the other hand, the present invention provides a computer-readable storage medium having a program stored thereon, which, when executed by a processor, is used to implement the above-described CBCT sparse angle reconstruction method based on a physically-aware residual guided diffusion model.

[0033] Compared with the prior art, the present invention has the following beneficial effects:

[0034] (1) This invention is the first to deeply embed the physical perception module based on the BP imaging algorithm into the residual guided diffusion model architecture, realizing the deep integration of imaging physical priors and diffusion model generation capabilities. It not only ensures the anatomical consistency and numerical accuracy of the reconstruction results by leveraging physical priors, but also makes up for the lack of information in sparse view projection data by utilizing the powerful detail completion capability of the diffusion model, effectively solving the problems of black-box modeling, poor physical consistency, and severe artifacts in sparse scenes of traditional reconstruction algorithms.

[0035] (2) By synchronously introducing residual information in the forward and backward processes of the diffusion model and training the model with a multi-time-step loss function, precise guidance of the diffusion iteration process is achieved. Compared with the traditional diffusion model, the convergence speed and reconstruction accuracy of the model are greatly improved. At the same time, the DDIM acceleration strategy is adopted to further improve the inference efficiency, which can meet the needs of real-time guidance during surgery.

[0036] (3) The coarse reconstructed images of this invention can be flexibly generated using various methods such as FDK algorithm and lightweight CNN, adapting to the sparse view CBCT reconstruction needs of different hardware platforms and application scenarios. The entire method takes into account reconstruction efficiency, quality and clinical interpretability, and can be widely applied to multiple medical clinical scenarios such as tumor radiotherapy localization, orthopedic surgery guidance, and orthodontic diagnosis, with strong robustness and versatility. Attached Figure Description

[0037] Figure 1 This is a schematic diagram of the CBCT sparse angle reconstruction method based on the physical perception residual guided diffusion model described in this invention.

[0038] Figure 2 This is an example diagram illustrating the forward diffusion and backward diffusion iteration processes in a residual-guided diffusion model. Figure 2 (a) in the diagram is an example of the forward diffusion process. Figure 2 (b) in the diagram is an example of the reverse diffusion iteration process.

[0039] Figure 3 For initial reconstructed image and residual guidance terms Example image.

[0040] Figure 4 This is a schematic diagram illustrating the effect of the CBCT sparse angle reconstruction method described in this invention. Detailed Implementation

[0041] The technical content of the present invention will now be described in detail with reference to the accompanying drawings and specific embodiments.

[0042] The flowchart of the CBCT sparse angle reconstruction method based on the physical perception residual guided diffusion model described in this invention is as follows: Figure 1 As shown, the core technology is based on the physical prior of the BP imaging algorithm, constructing a residual-guided diffusion model with an embedded physical perception module, realizing accurate residual correction and physical consistency constraints, and completing high-quality real-time sparse view CBCT reconstruction.

[0043] Step 1: Data Preparation and Preprocessing

[0044] During the acquisition process, sparse angle CBCT projection data is easily affected by factors such as X-ray scattering, detector electronic noise, and mechanical jitter of the scanning system. Directly using it for reconstruction will cause problems such as image artifacts and blurred details.

[0045] First, raw sparse angular CBCT projection data is obtained from the dataset. The data format is DICOM, and the resolution is 1024×1024 pixels. Preprocessing operations such as dark field correction, geometric correction, and noise suppression are performed on this data, including the following steps:

[0046] (1.1) Even without X-ray incident (i.e., zero exposure), the detector pixels will still output non-zero signals. This is mainly due to slight differences in the detector manufacturing process (leading to inconsistent pixel responsivity) and fluctuations in the bias voltage of the readout circuit. Dark-field correction can be used to correct this inconsistency, specifically using the following physical model:

[0047] (1)

[0048] in, For the original projection data, As a dark field reference image, For flat field reference image, This is the system gain factor.

[0049] (1.2) Geometric Calibration aims to correct the geometric distortion of the ray path in CBCT imaging systems caused by factors such as mechanical installation errors, rotation axis offset, and detector tilt. Typically, a phantom-based calibration method is used. This involves acquiring projection data from a known geometric structure (such as a steel ball array or specific marker points) and inversely solving for the actual geometric parameters (such as source-detector distance, source-axis distance, detector pixel offset, and rotation center drift). Subsequently, the distorted projection data is resampled into an ideal geometric space to ensure that the projection data at each viewpoint conforms to an accurate central ray geometric model.

[0050] (1.3) Perform noise removal operation to effectively suppress noise interference while preserving key structural information such as edges and textures in the projection data to the maximum extent. Then perform normalization processing to map the gray values ​​of the projection data to the [0,1] interval, unify the data scale, eliminate the data distribution differences caused by different acquisition batches and different scanning parameters, and provide high-quality input data for subsequent model training and image reconstruction.

[0051] Step 2: Initial Image Reconstruction

[0052] The classic FDK algorithm or the lightweight NeRF network (DIFNet) is used to reconstruct images from the preprocessed sparse angular CBCT projection data to obtain coarsely reconstructed CBCT images.

[0053] (a) Specifically, the classic FDK algorithm is adopted. First, based on the scanning geometry parameters of the CBCT equipment, key parameters such as the relative position of the X-ray source and detector, and the projection angle interval are determined, providing accurate geometric basis for filtering and backprojection operations. Then, weighted filtering is performed on the projection data of each angle. Finally, based on the geometric principle of cone-beam imaging, the filtered projection data is backprojected along the ray direction with weights, mapping the information of the projection domain to the three-dimensional voxel space. The gray value of each voxel is obtained through integration calculation, and then stacked to generate the initial three-dimensional reconstructed image (coarse reconstructed CBCT image). The classic FDK calculation formula is as follows:

[0054] (2)

[0055] In the formula, voxel points in 3D space after reconstruction grayscale value, For projection angle parameters, This indicates the projection data at voxel points after weighted filtering. The value at the projection position, Indicates after filtering, This represents the redundant weighting function in the back projection process.

[0056] (b) A lightweight NeRF network (DIFNet) is adopted. DIFNet models the three-dimensional imaging space based on implicit neural representations. It can fit the three-dimensional voxel gray intensity with a small amount of sparse projection data, which is suitable for sparse angle CBCT scanning scenarios. Its reconstruction process is mainly divided into three core steps: 1) Projection feature extraction: A lightweight 2D convolutional encoder is used to encode the features of the preprocessed multi-view sparse projection image, quickly extract effective features such as image edges and tissue structures, remove noise and sampling artifact interference, and output multi-scale two-dimensional projection features; 2) Cross-view feature fusion: Combine the geometric parameters of CBCT cone-beam scanning to establish the mapping relationship between three-dimensional voxel coordinates and two-dimensional projection features. The features of each view are queried through geometric back projection, and then the effective information of multiple views is aggregated by the cross-view fusion module to make up for the data loss of sparse views and suppress reconstruction stripe artifacts and structural defects; 3) Intensity field reconstruction inference: The fused voxel features are input into a lightweight MLP network to learn the nonlinear mapping relationship between three-dimensional coordinates and CBCT gray intensity, and construct a continuous and smooth three-dimensional depth gray field.

[0057] Step 3: Construct a residual-guided diffusion model embedded with a physical sensing module

[0058] This step is the core of this invention, aiming to solve the problem of lack of physical consistency in purely data-driven diffusion models by fusing imaging physics priors and diffusion models. It constructs the basic architecture of a residual-guided diffusion model, which includes a forward diffusion process and a backward diffusion iterative process. A specific schematic diagram is shown below. Figure 2 .

[0059] Noise and residual information are added simultaneously during the forward pass of the residual network to simulate the image degradation process under sparse views, providing a clear optimization objective for the inverse iteration. The process is as follows:

[0060] (3)

[0061] in, The degraded image obtained during the forward diffusion process at step t, These are the original images, acquired using high-precision CT equipment. This is the residual image (the difference between the real image and the original image). It is standard Gaussian noise. They are the first The cumulative residual retention factor and cumulative noise retention factor of each step are derived from the single-step coefficients. It is obtained by recursion, which satisfies , .

[0062] Therefore, during the forward diffusion process, from... arrive The transition probability can be defined as:

[0063] (4)

[0064] The reverse process can be represented as follows:

[0065] (5)

[0066] in, , To generate process control coefficients, when At times, the reverse process is random, resulting in diverse outcomes; when hour, If so, then the reverse process is deterministic.

[0067] Therefore, in the reverse diffusion iteration process, the transition probability can be defined as:

[0068] (6)

[0069] In the residual-guided diffusion model with an embedded physical sensing module designed in this invention, each iteration of the reverse process predicts residual information. and noise information , can be represented as:

[0070] (7)

[0071] (8)

[0072] in, For predictive networks, network architectures such as Unet or Vision Transformer can be selected. The initial T residual guide term and the time point t residual guide term, respectively, can be expressed as:

[0073] (9)

[0074] (10)

[0075] in, For the back projection operator, For projection operators, This is the initial reconstruction result. Figure 3 for A schematic diagram (5 projection angles).

[0076] The final loss function can be expressed as:

[0077] (11)

[0078] Where m is the number of time points spanned. These are the weighting coefficients. The residual prediction loss and the noise prediction loss are respectively expressed as follows:

[0079] (12)

[0080] (13)

[0081] Step 4: Collect data to train the residual-guided diffusion model.

[0082] Specifically, the training samples in this step consist of coarsely reconstructed CBCT images and corresponding real CT images, which are acquired using high-precision CT equipment. The coarsely reconstructed CBCT images are obtained by initially reconstructing preprocessed sparse angular CBCT projection data using the FDK algorithm or a lightweight NeRF network (see step 2 for details), providing image priors that meet the basic structural requirements. The real CT images serve as labels, reconstructed through full-angle scanning, providing accurate supervision information for model training.

[0083] During training, multi-timestep forward diffusion is performed on the input coarsely reconstructed CBCT images: 1000 diffusion timesteps are predefined, each corresponding to a different noise intensity parameter; Gaussian noise and residual information are added to the images sequentially according to the timesteps, simulating the degradation process from a coarsely reconstructed image to complete noise, thus constructing image sequences with different degrees of degradation. Subsequently, using the multi-timestep noisy images as input, inverse denoising training is carried out through a residual-guided mechanism: the residual-guided diffusion model network receives the noisy image of the current timestep and the corresponding timestep embedding information as input, predicts the residuals and residual information that should be removed in the current step, thereby generating intermediate reconstruction results close to real CT; during training, each training batch randomly selects multiple timesteps from the 1000 timesteps for parallel training, enabling the network to learn the denoising capabilities under different noise intensities and improve the model's generalization ability.

[0084] Step 5: Image reconstruction.

[0085] A coarsely reconstructed CBCT image is generated using a coarse reconstruction method (see steps 1 and 2 for details). This coarsely reconstructed CBCT image is then input into a trained residual-guided diffusion model. During the inverse iteration process, the model calculates the backpropagation (BP) image of the projection residual in real time using a physical sensing module and applies guidance. By adjusting the iteration step size of the diffusion process, the number of iterations is reduced, improving inference efficiency. Ultimately, the model outputs a finely reconstructed CBCT image with high detail and high physical consistency, meeting the application requirements of clinical diagnosis or industrial testing.

[0086] In practical deployments, the DDIM method can be used for acceleration. The DDIM acceleration strategy adjusts the iteration step size of the diffusion process, skipping some intermediate diffusion time steps and directly estimating the image state at key time steps, significantly reducing the number of iterations for model inference. This strategy improves model inference speed by 10-100 times without sacrificing the quality of the reconstructed image.

[0087] Accordingly, the present invention also provides a CBCT sparse angle reconstruction device based on a physically-aware residual guided diffusion model, comprising one or more processors and a memory, wherein the memory is coupled to the processor; wherein the memory is used to store program data, and the processor is used to execute the program data to implement the above-described CBCT sparse angle reconstruction method based on a physically-aware residual guided diffusion model.

[0088] Accordingly, the present invention also provides a computer-readable storage medium having a program stored thereon, which, when executed by a processor, is used to implement the above-described CBCT sparse angle reconstruction method based on a physically-aware residual guided diffusion model.

[0089] Example 1

[0090] This embodiment uses CT data for simulation and compares the reconstruction results of the DIFNet-reconstructed image with those of the proposed method. Figure 4 shows three sets of key comparison results: the left side is the real high-resolution reference image (ground truth), the middle side is the initial coarse reconstructed image obtained by the DIFNet reconstruction algorithm, and the right side is the fine reconstruction result of the proposed physical perception residual guided diffusion model. The peak signal-to-noise ratio (PSNR) of the corresponding image is labeled for quantitative evaluation of reconstruction quality. The PSNR of the reconstructed image obtained by the DIFNet method is 31.7125 dB. After optimization by the proposed physical perception residual guided diffusion model, the PSNR is improved to 36.6208 dB. The spatial resolution of the image is significantly improved, noise and artifacts are significantly suppressed, and the grayscale distribution is highly consistent with the real image, verifying the advantages of this method in detail restoration and physical consistency.

[0091] The key parameter configurations for the physically-aware residual-guided diffusion model used in this embodiment are as follows:

[0092] Diffusion process settings: The original diffusion time step is set to 1000 steps, and full-time-step training is used during the training phase; the DDIM acceleration strategy is used during the inference phase to compress the effective iteration steps to 50 steps; Backbone network: Adopts U-Net architecture, containing 4 downsampling / upsampling modules, each layer is configured with residual convolutional units; the basic number of channels is 64, the number of channels doubles with increasing depth, and the maximum number of channels is 512; the convolutional kernel size is 3×3, and the activation function is GELU.

[0093] Training hyperparameters: Initial learning rate is 1×10 -4 The training uses cosine annealing with learning rate decay; the batch size is 8; the number of training epochs is 100; and the weight decay coefficient is 1×10⁻⁶. -5 Use the AdamW optimizer.

[0094] Obviously, the above embodiments are merely illustrative examples for clear explanation and are not intended to limit the implementation. Those skilled in the art will recognize that other variations or modifications can be made based on the above description. It is neither necessary nor possible to exhaustively list all possible implementations here. However, obvious variations or modifications derived therefrom are still within the scope of protection of this invention.

Claims

1. A CBCT sparse angle reconstruction method based on a physically-perceived residual-guided diffusion model, characterized in that, Includes the following steps: (1) Obtain the original sparse angle CBCT projection data and preprocess the projection data; (2) The preprocessed sparse angle CBCT projection data is reconstructed using a reconstruction algorithm to obtain a coarsely reconstructed CBCT image; (3) Construct a residual-guided diffusion model that embeds a physical sensing module based on the BP imaging algorithm; (4) The residual-guided diffusion model is trained using coarsely reconstructed CBCT images and corresponding real CT images acquired by high-precision CT equipment as training sample pairs; (5) During the inference process, a coarse reconstructed CBCT image is generated and input into the trained residual guided diffusion model, and a fine reconstructed CBCT image is output. The residual-guided diffusion model includes a forward diffusion process and a backward diffusion iteration process. In the forward diffusion process, Gaussian noise and residual information are added to the real CT image simultaneously. In the backward diffusion iteration process, the difference between the projection of the image generated in the current iteration and the real projection is calculated using the physical perception module to obtain the projection residual. The projection residual is then back-projected using the BP algorithm to generate a residual BP image, which is then added to the input of the current iteration to guide the model to accurately estimate the noise and residual of the current iteration. Finally, the Gaussian noise and residual information of the current iteration are estimated.

2. The CBCT sparse angle reconstruction method according to claim 1, characterized in that, In step (1), the preprocessing includes dark field correction, geometric correction and noise removal of the original sparse angle CBCT projection data.

3. The CBCT sparse angle reconstruction method according to claim 1, characterized in that, The reconstruction algorithm is either the classic FDK algorithm or a lightweight NeRF network, wherein the lightweight NeRF network is DIFNet.

4. The CBCT sparse angle reconstruction method according to claim 1, characterized in that, The forward diffusion process simultaneously adds Gaussian noise and residual information to the real CT image, and the specific process is as follows: ； in, The degraded image obtained during the forward diffusion process at step t, For coarse reconstruction of CBCT images, The difference between the real image and the original image. It is standard Gaussian noise. They are the first The cumulative retention factor of the residuals and the cumulative retention factor of the noise of each step.

5. The CBCT sparse angle reconstruction method according to claim 4, characterized in that, The From single-step coefficients It is obtained by recursion, which satisfies , .

6. The CBCT sparse angle reconstruction method according to claim 1, characterized in that, In the reverse iterative diffusion process of the residual-guided diffusion model, each iteration is performed by a deep learning network to estimate and remove the noise and residual components of the current iteration step, generating an intermediate image for the next iteration step. The deep learning network is a Unet or VisionTransformer network structure.

7. The CBCT sparse angle reconstruction method according to claim 1, characterized in that, The reverse diffusion iteration process of the residual-guided diffusion model can be accelerated by the DDIM acceleration strategy to improve inference efficiency. By adjusting the iteration step size of the diffusion process, the number of iterations of model inference can be reduced.

8. A CBCT sparse angle reconstruction device based on a physically-aware residual-guided diffusion model, comprising one or more processors and a memory, characterized in that, The memory is coupled to the processor; wherein the memory is used to store program data, and the processor is used to execute the program data to implement the CBCT sparse angle reconstruction method based on the physical perception residual guided diffusion model as described in any one of claims 1-7.

9. A computer-readable storage medium, characterized in that, It stores a program that, when executed by a processor, is used to implement the CBCT sparse angle reconstruction method based on the physical perception residual guided diffusion model as described in any one of claims 1-7.

Images