Weight-bearing cone beam CT enhancement method, device and medium

Abstract

This application discloses a weighted cone-beam computed tomography (CBCT) enhancement method applied to an image processing device, comprising: acquiring a weighted CBCT source image to be processed; performing edge detection on the source image to obtain a contour feature map; performing forward diffusion processing on the source image to obtain a noisy image; concatenating the contour feature map and the noisy image by channel and inputting them into a control encoder to obtain pose features; injecting the pose features into the backbone decoder of a pre-trained diffusion model via a zero convolutional layer through skip connections to obtain fusion features; and performing reverse diffusion sampling based on the fusion features to obtain an enhanced weighted CBCT image.

Description

Technical Field

[0001] This application belongs to the field of medical image processing technology, and in particular relates to a weighted cone-beam CT enhancement method, device and medium. Background Technology

[0002] Weight-bearing CT scans acquire three-dimensional anatomical data of the human body under physiological load, which can accurately reflect the anatomical morphology of the spine, pelvis, lower limbs and other parts under standing pressure. It has important clinical significance for the diagnosis of diseases such as scoliosis, pelvic instability and degenerative joint diseases.

[0003] Currently, cone-beam CT (CBCT) is widely used in weightless imaging due to its low cost and low radiation dose. However, limited by the physical characteristics of flat panel detectors, weightless CBCT images have inherent defects such as severe scattering artifacts and low soft tissue contrast, which affect the accuracy of clinical diagnosis.

[0004] To improve the image quality of weight-bearing CBCT, existing technologies attempt to use deep learning methods to translate low-quality weight-bearing CBCT images into high-quality spiral CT images. However, due to the difficulty in obtaining paired CBCT and spiral CT data of the same patient in the same weight-bearing posture in clinical practice, existing models are usually trained on readily available supine paired data. This leads to a key problem: when the model processes weight-bearing CBCT input, it experiences a "posture illusion" phenomenon, where the model tends to incorrectly reconstruct the anatomical morphology under weight-bearing conditions (such as compressed intervertebral discs and offset pelvic angles) as a supine position, thus losing the true physiological load information that weight-bearing scans aim to capture.

[0005] In summary, existing technologies, when training deep learning models using supine paired data, produce a "posture illusion" phenomenon when processing weight-bearing CBCT images. That is, the model incorrectly restores the anatomical morphology under weight-bearing conditions to the supine state, and cannot effectively maintain the true posture characteristics under weight-bearing conditions. Summary of the Invention

[0006] This application provides a weighted cone-beam computed tomography (CBCT) enhancement method, apparatus, and medium to solve the technical problem that existing technologies produce posture illusions and cannot maintain the true posture characteristics of weighted positions when processing weighted CBCT images after training deep learning models using supine paired data.

[0007] In a first aspect, this application provides a weight-bearing cone-beam computed tomography (CBCT) enhancement method applied to an image processing device, comprising: acquiring a weight-bearing CBCT source image to be processed; performing edge detection on the weight-bearing CBCT source image to obtain a contour feature map; wherein the contour feature map represents the patient's posture information under weight-bearing conditions; performing forward diffusion processing on the weight-bearing CBCT source image to obtain a noisy image; performing channel concatenation of the contour feature map and the noisy image to obtain a concatenation tensor; inputting the concatenation tensor into a control encoder to obtain posture features; injecting the posture features into the backbone decoder of a pre-trained diffusion model via a skip connection through a zero-convolutional layer to obtain fusion features; wherein the pre-trained diffusion model is trained based on a supine paired dataset; and performing reverse diffusion sampling based on the fusion features to obtain an enhanced weight-bearing CT image.

[0008] Furthermore, the pre-trained diffusion model is a denoised diffusion generative adversarial network, including a conditional generator and a discriminator;

[0009] The condition generator models the back transition probability implicitly, and its expression is:

[0010]

[0011] in, For the noisy image at step t, This is the image at step t-1, where t is the diffusion time step. For contour feature map, The image is a negative-weight CBCT source image. For latent variables that follow a standard normal distribution, for The prior distribution, For condition generator, For a clear image of the prediction;

[0012] The condition generator is trained using a variance-preserving marginal distribution that satisfies:

[0013]

[0014] in, For the clear image Noisy image at step t The conditional probability distribution, Indicates a Gaussian distribution. This is the cumulative signal preservation factor. It is the identity matrix. =exp(- ), variance scheduling function >0

[0015] Furthermore, the pre-trained diffusion model also includes a backbone encoder corresponding to the backbone decoder; the step of injecting the pose features into the backbone decoder of the pre-trained diffusion model via skip connections through zero convolutional layers to obtain fused features includes:

[0016] Establish a multi-scale jump connection between the control encoder and the backbone decoder;

[0017] A zero-convolutional layer is set on each of the paths of the skip connections; wherein the weights and biases of the zero-convolutional layer are initialized to zero;

[0018] The fusion features at the i-th resolution scale are obtained using the following formula:

[0019]

[0020] in, To fuse feature maps, This is the locked feature map of the backbone encoder at the i-th scale. To control the encoder's pose feature map at the i-th scale, For zero convolution operators, These are the parameters for a zero convolutional layer.

[0021] Furthermore, during the training process of the pre-trained diffusion model, the loss function includes adversarial loss and cycle consistency loss:

[0022]

[0023] in, For the total loss, To combat the losses, For cycle consistency loss, These are the weighting coefficients.

[0024] Furthermore, the step of performing edge detection on the negative-weight CBCT source image to obtain a contour feature map includes:

[0025] The negative CBCT source image is processed using the Canny edge detection operator. Based on preset high and low thresholds, the edges are connected using the hysteresis thresholding method to obtain a binarized contour feature map.

[0026] Furthermore, the inverse diffusion sampling based on the fused features includes performing mask diffusion sampling:

[0027] Adaptive threshold segmentation is performed on the negative-position CBCT source image to obtain an anatomical mask; wherein, in the anatomical mask, the pixel value of the tissue region is 1, and the pixel value of the background region is 0;

[0028] In each iteration of the backdiffusion sampling, the predicted image is projected according to the following formula:

[0029]

[0030] in, The projected image. This is the anatomical mask. For the image predicted by the model, For background pixels, For Hadamard products.

[0031] Furthermore, the reverse diffusion sampling based on the fusion features includes performing alternating fusion sampling:

[0032] Parallel evolution of structure-preserving trajectories and bone-enhancing trajectories; wherein the structure-preserving trajectory is used to preserve anatomical geometric boundaries, and the bone-enhancing trajectory is used to enhance bone texture details;

[0033] When executing the structure-preserving trajectory, the contour feature map is used as a conditional input;

[0034] When executing the bone enhancement trajectory, set the conditional input to empty;

[0035] Threshold segmentation is performed on the output image of the bone enhancement trajectory to obtain the bone mask;

[0036] The outputs of the structure-preserving trajectory and the bone enhancement trajectory are semantically fused according to the following formula to obtain the final enhanced CT image:

[0037]

[0038] in, To ultimately enhance the image, For the skeleton mask, The output image of the bone enhancement trajectory. The output image that preserves the trajectory of the structure. For Hadamard products.

[0039] Secondly, this application provides a weighted cone-beam CT enhancement device, comprising:

[0040] The image acquisition module is used to acquire the negative-position CBCT source image to be processed;

[0041] The edge detection module is used to perform edge detection on the CBCT source images under load to obtain contour feature maps; the contour feature maps represent the patient's posture information under load.

[0042] The forward diffusion module is used to perform forward diffusion processing on negative CBCT source images to obtain noisy images;

[0043] The channel stitching module is used to stitch the contour feature map with the noisy image through channels to obtain the stitching tensor.

[0044] The attitude feature extraction module is used to input the spliced ​​tensor into the control encoder to obtain attitude features;

[0045] The feature fusion module is used to inject pose features into the backbone decoder of the pre-trained diffusion model via skip connections through zero convolutional layers to obtain fused features; wherein, the pre-trained diffusion model is trained based on the supine position paired dataset.

[0046] The image generation module is used to perform reverse diffusion sampling based on fusion features to obtain enhanced negative-weight CT images.

[0047] Thirdly, this application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the above-described method.

[0048] This application obtains contour feature maps representing the posture information of the weighted state by performing edge detection on the weighted CBCT source image, and then injects the posture features extracted based on the contour feature maps into the backbone decoder of the pre-trained diffusion model via skip connections through a zero convolutional layer. Since the pre-trained diffusion model is trained on a supine paired dataset, it already has the ability to generate high-quality spiral CT textures. By injecting weighted posture constraints during the decoding stage, the diffusion model is constrained within the posture boundary defined by the weighted source image while generating high-definition textures, thus effectively solving the posture illusion problem. Attached Figure Description

[0049] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below.

[0050] Figure 1 This is a schematic diagram of the structure of the weighted cone-beam CT enhancement system provided in the embodiments of this application;

[0051] Figure 2 A schematic flowchart illustrating the weighted cone-beam CT enhancement method provided in this application embodiment;

[0052] Figure 3 This is a comparison chart of HU value intensity distribution curves in the embodiments of this application;

[0053] Figure 4 This is a comparison diagram of ablation experiments in the embodiments of this application;

[0054] Figure 5 A performance comparison table of different algorithms on the pelvic and spine datasets;

[0055] Figure 6 Comparison table of Dice coefficients for segmentation of major anatomical sites;

[0056] Figure 7 This is a table comparing the computational complexity of different methods. Detailed Implementation

[0057] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application.

[0058] For ease of understanding, the following explanations are provided for the terminology related to the diffusion model in this application:

[0059] Forward diffusion refers to the process of starting with a clear image and gradually adding Gaussian noise until the image becomes pure noise. This process transforms the complex image data distribution into a simple standard normal distribution.

[0060] Reverse diffusion refers to the reverse process of forward diffusion. Starting from pure noise, it uses neural networks to predict and remove noise at each step, gradually recovering a clear image.

[0061] The score function refers to the gradient of the logarithmic probability density of the data distribution, used in diffusion models to guide the direction of backsampling. Neural networks achieve denoising and reconstruction by learning the score function.

[0062] Sampler: refers to an algorithm that performs reverse diffusion iterative updates, updating the current noisy image to a less noisy image based on the score estimate from the output of the neural network.

[0063] like Figure 1As shown, this application provides a weight-bearing cone-beam computed tomography (CBCT) enhancement system, which includes a medical image acquisition device 101 and an image processing device 102. The medical image acquisition device 101 is a weight-bearing CBCT scanning device used to acquire CBCT image data of a patient in a standing, weight-bearing state. The image processing device 102 is a server or workstation with deep learning computing capabilities, connected to the medical image acquisition device 101 via a network. The medical image acquisition device 101 sends the acquired weight-bearing CBCT source images to the image processing device 102. The image processing device 102 executes the image enhancement method provided in this application to process the weight-bearing CBCT source images, obtaining enhanced high-quality weight-bearing CT images, and returns the enhanced images to the medical image acquisition device 101 or sends them to a PACS system for clinicians to review and diagnose.

[0064] The PPDM (Posture Preserving Diffusion Model) provided in this application consists of three stages: pre-training stage, fine-tuning stage, and sampling stage.

[0065] Pre-training phase: A denoising diffusion generative adversarial network (DDGAN) is trained using a supine paired dataset (paired CBCT and spiral CT images). DDGAN employs an adversarial learning mechanism, learning the mapping relationship between the CBCT and spiral CT image domains through game-theoretic optimization between the generator and discriminator. During training, adversarial loss and cycle consistency loss are used for optimization, enabling the model to recover high-resolution spiral CT textures from low-quality CBCT images.

[0066] Fine-tuning Phase: A Controlled Noise Conditional Score Network (CNCSN) is introduced based on the pre-trained model. A FreeD parameter isolation strategy is employed, freezing all parameters of the discriminator, backbone encoder, and backbone decoder in the pre-trained model, training only the control encoder and zero-convolutional layer parameters in the CNCSN. During fine-tuning, the contour feature map of the weighted CBCT image is used as conditional input, and the pose constraint signal is injected into the diffusion module generator through the CNCSN, enabling the model to learn the ability to maintain the weighted pose. This parameter isolation strategy ensures that the model does not experience "catastrophic forgetting" when learning new pose constraints, thus maintaining high-fidelity contrast in the generated images.

[0067] Sampling Phase: For the input negative-weight CBCT source image, a contour feature map is first extracted as an attitude condition, and then dual-trajectory parallel sampling is performed. The structure-preserving trajectory is conditioned on the contour feature map and estimated using the CNCSN score estimation function. Guided generation, preserving anatomical geometric boundaries; the skeletonized trajectory sets conditions to empty, estimating the function through unconditional score. Guided generation enhances bone texture details. Finally, the extracted bone mask is used... Semantic-level fusion is performed on the outputs of the two trajectories to obtain enhanced negative-weight CT images.

[0068] This application achieves zero-shot translation through cross-pose domain transfer: the paired image dataset used for training in the pre-training stage consists of CBCT images and spiral CT images in the supine position; the fine-tuning and sampling stages process unpaired CBCT images in the weighted position. Traditional models, after training on supine CT, forcibly restore the skeleton to the supine position when encountering weighted CBCT input (producing pose illusion). This application uses CNCSN to lock the contour boundary of the weighted position, forcing the diffusion model to generate high-resolution textures only within the boundary without changing the boundary shape, thereby achieving pose preservation from texture knowledge in the supine position domain to the weighted position domain.

[0069] The CNCSN module adopts a dual-path parallel architecture, including a trunk path and a control path.

[0070] The backbone path includes a backbone encoder and a backbone decoder. The backbone encoder receives noisy images. The embedded information at time step t is processed through a 3×3 convolutional layer, followed by multiple encoding levels (Encoder Level 0 to Encoder Level N) and a middle layer. Each level contains a residual attention block (Res+Attn Block). The backbone decoder starts from the middle layer and proceeds through multiple decoding levels (Decoder Level N to Decoder Level 0), with skip connections between each decoding level and its corresponding encoding level. The parameters of the backbone path are locked during the fine-tuning phase to preserve the high-resolution texture generation capability obtained during pre-training. Because the backbone encoder is locked, its number of input channels remains consistent with that during pre-training, processing only noisy image signals and ensuring the stability of the underlying feature extraction logic.

[0071] The control path consists of a control encoder and zero-convolutional layers. The control encoder has the exact same structure as the backbone encoder, serving as a trainable copy of it. The control encoder receives contour feature maps obtained by processing the weighted CBCT source image y. With noisy images The concatenated tensor is used as input, along with the time step. Embedded information. Contour feature map. First, the system processes the data through 3×3 convolutional layers and zero convolutional layers to extract initial spatial constraint features. Then, it sequentially passes through multiple encoding levels (Encoder Level 0 to Encoder Level N) and intermediate layers, identical to the backbone encoder structure. The parameters of the control path are trainable during the fine-tuning phase and are specifically designed to be learned from the input. Extract constraint features related to the current weight position and learn how to transform them into signals generated by guided diffusion.

[0072] Feature injection method: The pose features output by the control encoder at each resolution level are processed by a zero-convolutional layer and then injected element-wise into the skip connections of the corresponding resolution level of the backbone decoder. Specifically, the output of the intermediate layer is injected into the beginning of the backbone decoder (i.e., the input of Decoder Level N) after passing through a zero-convolutional layer; the output of the control encoder at Encoder Level i is injected into the skip connections of the corresponding level Decoder Level i of the backbone decoder after passing through a zero-convolutional layer, where i is the level index decreasing from N to 0. Through the above feature injection mechanism corresponding to the resolution scale, it is ensured that the pose constraint signals extracted by the control encoder under different receptive fields can be accurately semantically fused with the feature flow of the corresponding scale in the backbone decoder. The zero-convolutional layer uses 1×1 convolution, and its weights and biases are initialized to zero to ensure that the output of the control path is zero in the early stage of fine-tuning, so as not to interfere with the prior knowledge of high-frequency texture and material composition distribution already learned by the backbone path. With training iterations, the weights of the zero-convolutional layer are gradually updated, so that the pose constraint signals can be smoothly injected into the decoding stage, forcibly locking the negative position pose geometry features.

[0073] Data flow process: During inference, the negative CBCT source image is used to obtain a contour feature map through Canny edge detection. , with initial noise Input CNCSN together. Noisy image. Extracting multi-scale features from the backbone encoder; contour feature map The attitude features are extracted from the control encoder. The attitude features of each level of the control encoder are injected into the backbone decoder after passing through a zero convolutional layer. They are fused with the features of the backbone encoder, and finally the backbone decoder outputs the conditional score estimation result, which guides the diffusion sampling to evolve in the direction that conforms to the attitude of the negative position.

[0074] like Figure 2 As shown in the figure, this application provides a weighted cone-beam computed tomography (CBCT) enhancement method, applied to an image processing device 102. The method includes the following steps:

[0075] Step S201: Obtain the negative-position CBCT source image to be processed.

[0076] In this step, the image processing device 102 receives the weight-bearing CBCT source image sent by the medical image acquisition device 101. This weight-bearing CBCT source image is a cone-beam CT image acquired when the patient is in a natural standing position (weight-bearing state), containing the true anatomical morphology information of the human body under physiological load. Because weight-bearing CBCT uses a large-size flat panel detector, and is limited by physical imaging characteristics, this source image usually has inherent defects such as severe scattering artifacts, low soft tissue contrast, and unstable HU value.

[0077] Preferably, after acquiring the weight-bearing CBCT source image to be processed, the method further includes preprocessing the image. Specifically, the weight-bearing CBCT source image undergoes CT value truncation, limiting the pixel values ​​to a preset display window range. For example, the display window can be set to [-1000, 1000] HU values, which can be adjusted according to actual clinical needs. Pixel values ​​exceeding this range are truncated to reduce interference from metal implants or air noise. Then, the truncated image is normalized by using linear mapping to scale the truncated HU values ​​to the [0, 1] interval, and then linearly mapping them to the [−1, 1] interval to match the input requirements of the deep neural network. In addition, bilinear interpolation algorithm and zero-padding technology can be used to uniformly adjust the image to a preset resolution, for example, 256×256 pixels.

[0078] Step S202: Perform edge detection on the negative CBCT source image to obtain a contour feature map.

[0079] In this step, edge detection is performed on the weight-bearing CBCT source images to extract contour feature maps representing the patient's posture information under weight-bearing conditions. These contour feature maps contain precise geometric posture information of the patient under weight-bearing conditions, such as spinal curvature under compression, pelvic angle deviation, and anatomical features like joint space under stress. Explicitly extracting this posture boundary information provides guiding signals for subsequent posture constraint injection, which is a crucial step in resolving the posture illusion problem.

[0080] Preferably, the edge detection is implemented as follows: the Canny edge detection operator is used to process the negative-position CBCT source image, dual threshold parameters (low threshold and high threshold) are set, edges are connected by hysteresis thresholding, and binarized contour feature maps are extracted based on the preset high and low thresholds. For example, the low threshold can be set to 50 and the high threshold can be set to 150. This parameter can be adjusted according to the image contrast, and the ratio of the high threshold to the low threshold is generally maintained between 2:1 and 3:1.

[0081] Step S203: Perform forward diffusion processing on the negative-weight CBCT source image to obtain a noisy image.

[0082] In this step, a forward diffusion process is performed on the negative-weight CBCT source image, gradually adding Gaussian noise to the image. Forward diffusion is one of the core operations of the diffusion model, its purpose being to gradually transform the complex data distribution of the original image into a simple standard normal distribution, laying the foundation for subsequent inverse reconstruction. Furthermore, this process can be modeled based on the marginal distribution of stochastic differential equations (SDEs), defined as:

[0083]

[0084] in, The image is the original negative-weight CBCT source image. For the noisy image at step t, The predefined variance scheduling parameters control the noise intensity added at each step. It is the identity matrix. As the number of steps t increases, the original information in the image is gradually overwhelmed by noise.

[0085] In this embodiment, for example, the forward diffusion process can be set to a total number of steps T=1000, and the variance scheduling range can be set to... =0.1 to =20. The variance scheduling parameter mentioned here. This refers to the noise intensity rate over a continuous time frame. (By...) By setting the above range and combining it with the aforementioned continuous marginal distribution formula, the diffusion trajectory can more efficiently cover the data manifold from low-contrast CBCT to high-resolution CT while ensuring numerical stability.

[0086] This represents the cumulative preservation factor from the initial state to step t, which is related to the variance scheduling parameter. The relationship is =exp(- This formula, through an exponential decay mechanism, ensures that even... The range of values ​​is relatively large (e.g.) When =20), the coefficient It remains within the real number interval of (0, 1), thus ensuring the self-sufficiency of the digital logic.

[0087] Step S204: Perform channel concatenation between the contour feature map and the noisy image to obtain the concatenation tensor.

[0088] In this step, the contour feature map (1 channel) obtained in step S202 and the noisy image (1 channel) obtained in step S203 are concatenated along the channel dimension to form a 2×H×W concatenation tensor, which serves as the input to control the encoder. Here, H and W represent the height and width of the image, respectively; in this embodiment, H=W=256.

[0089] Step S205: Input the splicing tensor into the control encoder to obtain attitude features.

[0090] In this step, the spliced ​​tensor is input to the control encoder for feature extraction to obtain a multi-scale attitude feature map containing attitude constraint information.

[0091] Preferably, the control encoder is a structural copy of the backbone encoder of the pre-trained diffusion model. Specifically, this application constructs a Controlled Noise Conditional Score Network (CNCSN) using a dual-path parallel architecture. The backbone encoder inherits from the pre-trained diffusion model encoder, and its weight parameters are completely locked during the fine-tuning phase to preserve prior knowledge of the high-frequency texture and material composition distribution of the helical CT images. The control encoder is a complete topological copy of the backbone encoder, and its parameters are trainable to extract pose constraint signals. This design allows the model to learn new pose constraints without "catastrophic forgetting," thereby maintaining high-fidelity contrast in the generated images.

[0092] Step S206: Inject pose features into the backbone decoder of the pre-trained diffusion model via skip connections through zero convolutional layers to obtain fused features.

[0093] In this step, zero-convolutional layers are used to inject the pose features extracted by the control encoder into the backbone decoder of the pre-trained diffusion model via skip connections, thereby achieving the fusion of pose constraint signals and high-quality texture priors. The pre-trained diffusion model is trained on a supine paired dataset and has the ability to recover high-resolution spiral CT textures from low-contrast, noisy CBCT images. Through the injection mechanism of zero-convolutional layers, the pose geometry features of the negative weight position can be forcibly locked without destroying these learned high-quality texture priors. It is worth noting that in this embodiment, the pose features are injected into the backbone decoder rather than the backbone encoder. By performing feature fusion during the decoding stage, the destruction of the underlying texture features of the source image is effectively avoided, thereby maximizing the preservation of the high-definition contrast characteristics of spiral CT while maintaining the pose of the negative weight position.

[0094] Preferably, the specific implementation of injecting pose features into the pre-trained diffusion model's backbone decoder via skip connections through zero-convolutional layers to obtain fused features can further include: step S2061, establishing multi-scale skip connections between the control encoder and the backbone decoder; step S2062, setting zero-convolutional layers on the paths of each skip connection, the zero-convolutional layers using 1×1 convolution operators, with their weights and biases initialized to zero; step S2063, obtaining the fused features at the i-th resolution scale according to the following formula:

[0095]

[0096] in, To fuse feature maps, the pre-trained diffusion model also includes a backbone encoder corresponding to the backbone decoder. This is the locked feature map of the backbone encoder at the i-th scale. To control the encoder's pose feature map at the i-th scale, For zero convolution operators, The parameters are set to zero for the convolutional layers. This design ensures that, in the early stages of fine-tuning, any random initialization noise controlling the encoder output will not contaminate the feature flow of the backbone network. As iterations proceed, the model can smoothly learn how to progressively inject the pose constraint signal into the backbone decoder through non-zero gradient updates.

[0097] Furthermore, the pre-trained diffusion model is a Denoising Diffusion Generative Adversarial Network (DDGAN), comprising a conditional generator and a discriminator. Compared to the traditional Denoising Diffusion Probabilistic Model (DDPM), DDGAN utilizes adversarial learning mechanisms to compress the number of sampling steps in the inference phase from thousands to less than 8 steps while maintaining high fidelity in the generated images. This solves the technical problem of traditional diffusion models having excessively long inference iterations, making it difficult to meet the needs of real-time clinical diagnosis. The conditional generator models the back transition probability through implicit parameterization, and its expression is:

[0098]

[0099] in, For the noisy image at step t, This is the image at step t-1, where t is the diffusion time step. Let c be a latent variable that follows a standard normal distribution, and let c be a profile feature map. The source image is a negative-weight CBCT image. for The prior distribution, For condition generator, For the posterior distribution, For a clear image of the prediction.

[0100] The condition generator is trained using a marginal distribution that preserves variance. Variance preservation refers to the design of a noise scheduling strategy during the forward diffusion noise generation process, ensuring that the noise-generated image at each time step is optimized. The total variance remains constant (normalized to 1), thus ensuring energy conservation between the signal and noise. The marginal distribution of its forward noise addition process satisfies:

[0101]

[0102] in, Given a clear image Image with added noise at step t The conditional probability distribution; This represents a multidimensional Gaussian (normal) distribution; The mean term of this Gaussian distribution represents the original sharp image. The signal components after cumulative attenuation; Let be the covariance matrix of the Gaussian distribution. Let the identity matrix have the same dimensions as the image, representing noise that is independent and homoscedastic in each dimension, with a variance of size . .

[0103] The cumulative signal retention factor represents the proportion of the original signal retained when the signal evolves from the initial state to step t through forward diffusion. Its value ranges from (0, 1) and decreases monotonically with the increase of time step t. That is, as the diffusion process progresses, the original signal gradually decays and the noise gradually increases.

[0104] By using the variance scheduling function Earn points by playing: =exp(- );

[0105] in, The variance scheduling function defines the instantaneous noise injection rate at each time step s. >0. The above exponential integral form ensures that It is always a positive number, thus making the coefficient of the mean term... It is always a real number. In traditional DDPM, the variance scheduling parameter... Typically, values ​​are limited to the range (0,1) to ensure numerical stability; however, in the DDGAN framework used in this application embodiment, due to... Defined using the exponential form of continuous integrals, even if the variance scheduling parameter When >1, It is still a positive number. It remains a real number, so there will be no issue of numerical invalidation. Meanwhile, the coefficient of the mean term... With variance term (1− The sum of ) satisfies This means that the sum of signal energy and noise energy is always 1, which reflects the variance preservation characteristic.

[0106] By introducing latent variables This application bypasses the strong assumption in traditional diffusion models that "the denoising distribution must conform to a unimodal Gaussian distribution," thereby allowing for larger sampling step sizes and achieving highly efficient real-time reconstruction. Although the pre-trained model sets the total number of steps T=1000, after optimization through adversarial training, for example, the actual number of sampling steps in the inference sampling stage can be compressed to a step interval of k=250, meaning that only 4-8 discrete diffusion steps are needed to reconstruct a clear image. The specific number of steps can be adjusted according to the image quality requirements.

[0107] Furthermore, during the training process of the pre-trained diffusion model, the loss function includes adversarial loss and cycle consistency loss:

[0108]

[0109] in, For the total loss, To combat the losses, For cycle consistency loss, These are the weight coefficients. The training objective is to minimize the total loss. Combating losses The specific form is as follows:

[0110]

[0111] in For example, a discriminator. The output can be activated by a sigmoid function, with values ​​ranging from (0,1). Adversarial loss is used to train the discriminator to distinguish between genuine denoised sample pairs and pseudo-denoised sample pairs constructed by the conditional generator. Cycle consistency loss. The L1 norm is used to measure the reconstruction error of the source image after bidirectional mapping. This ensures that the image returns to its original domain after cyclic mapping, constrains the consistency of the bidirectional mapping between the source and target image domains, and prevents the loss of anatomical structures during unpaired training. For example, the weight coefficients... It can be set to 20, and this parameter can be adjusted according to the experimental results.

[0112] For example, the dataset structure and usage in this embodiment are as follows:

[0113] Pelvic Dataset: A publicly available pelvic CT dataset was used, containing a total of 11,660 images. Simulated CBCT images were generated by applying physical simulation processing such as scattering noise and contrast degradation to the spiral CT images, forming the paired training data required for the pre-training phase. Approximately 8,660 paired image pairs were selected for the pre-training phase, enabling DDGAN to learn the mapping ability from low-quality CBCT to high-resolution spiral CT; approximately 3,000 unpaired simulated weight-bearing CBCT images were selected for the fine-tuning phase, enabling CNCSN to learn its ability to maintain posture constraints.

[0114] Abdominal dataset: A publicly available abdominal CT dataset containing a total of 12,211 images was used. Paired data were generated using the same method as the pelvic dataset to validate the model’s generalization performance across different anatomical sites.

[0115] Validation set for weight-bearing positions: Three real CBCT scans of patients in weight-bearing positions were used as test examples. Each set of data included the three-dimensional projection reconstruction of the human body under natural standing physiological load, which was used to evaluate the enhancement effect of the model on real clinical weight-bearing images.

[0116] Cross-domain generalization set: using slices of anatomical sites such as the lungs that were not involved in the training, to validate the model's zero-shot generalization ability.

[0117] For validation cases containing corresponding spiral CT scans, affine transformation and nonlinear registration were performed using advanced normalization tools (ANTs) to map the spiral CT images to the physical coordinate system of the negative-position CBCT, thereby constructing a reference benchmark for quantitative performance evaluation.

[0118] Furthermore, the engineering configuration for the pre-training phase is as follows: The network backbone adopts an improved U-Net architecture (NCSN++ structure), with multiple (e.g., 6) resolution scale encoder and decoder blocks, each scale containing residual attention blocks (each scale can contain 2-3 residual blocks and attention modules). The Adam optimizer is used; for example, the initial learning rate can be set to 10. -4 The number of training iterations can be set according to the convergence results. Model training can be implemented using deep learning frameworks such as PyTorch, and the training hardware can be a GPU device with sufficient video memory.

[0119] Furthermore, the fine-tuning phase employs a FreeD parameter isolation strategy: freezing the discriminator. All parameters, freeze trunk encoder E locked The original parameters of the backbone decoder D are only for the control encoder E. ctrlThe parameters of each residual block along the path and the parameters of the zero-convolutional layers at each scale are backpropagated and updated. The fine-tuning stage uses approximately 3,000 unpaired negative-position CBCT images for training, without the corresponding helical CT images. For example, the fine-tuning learning rate can be set to 10. -5 The learning rate is approximately one-tenth of that in the pre-training phase, and cosine annealing can be used to control the decay of the learning rate. The number of training epochs can be fine-tuned based on the convergence results. This parameter isolation strategy ensures that the model does not suffer from "catastrophic forgetting" when learning new pose constraints, thereby maintaining high-fidelity contrast in the generated images.

[0120] Step S207: Perform reverse diffusion sampling based on fusion features to obtain enhanced negative weight CT images.

[0121] In this step, a reverse diffusion sampling process is performed to gradually restore the noisy image to a clear image. Reverse diffusion is the inverse process of forward diffusion: forward diffusion gradually adds noise to the clear image until it becomes pure noise, while reverse diffusion starts from the noise and uses a neural network to predict and remove noise at each step to gradually recover the clear image.

[0122] Based on the fusion features obtained in step S206, the backbone decoder outputs conditional score estimation results. Its expression is:

[0123]

[0124] in, As a backbone decoder, The locked feature map is the output of the backbone encoder. To control the attitude feature map output by the encoder, (·) is the zero convolution operator. This is a contour feature map.

[0125] This result characterizes the current noisy image. The gradient direction evolves towards a clearer image. The sampler performs iterative updates based on the score estimation results, moving the image from its current noisy state at each step. Update to a less noisy state This process is repeated iteratively until the initial noise is reduced. Gradually reconstruct the enhanced heavy-duty CT images .

[0126] Because this application employs the DDGAN architecture, after optimization through adversarial training, the inverse sampling process can use a larger step interval, requiring only a small number of discrete diffusion steps to complete reconstruction. Compared to the thousands of sampling steps required by traditional diffusion models, this significantly improves inference efficiency. The enhanced image maintains the patient's true anatomical posture under load, while possessing the texture quality and contrast of high-resolution spiral CT.

[0127] Preferably, inverse diffusion sampling is performed based on fusion features, including mask diffusion sampling. Since negative-weight CBCT images typically exhibit severe scattering artifacts outside anatomical boundaries, this application utilizes anatomical topological priors to spatially constrain the diffusion trajectory, thereby suppressing scattering artifacts outside the anatomical contour. Specifically, the implementation includes: step S2071, performing adaptive thresholding segmentation on the negative-weight CBCT source image to obtain an anatomical mask, wherein the pixel value of the tissue region in the anatomical mask is 1, and the pixel value of the background region is 0. In this embodiment, the Otsu adaptive thresholding segmentation algorithm is used to automatically determine the optimal segmentation threshold based on the image histogram, thereby generating an anatomical mask M∈{0,1} for human tissue. H×W M i,j =1 indicates that the coordinate point belongs to the tissue-supported region, M i,j =0 represents the background air region. In step S2072, at each iteration t∈[T,0] of the diffusion model inverse denoising, the predicted image is projected according to the following formula:

[0128]

[0129] in, The projected image. For dissection masking, For the image predicted by the model, The background pixel constant (for example, first normalized to [0,1], then linearly mapped to the [−1,1] interval to match the input requirements of the diffusion model). This is the Hadamard product (element-wise multiplication). This strategy forces the model to "remove" artifact signals outside the contour during each generation of denoising, ensuring that the generated CT image is strictly consistent with the weighted CBCT source image in terms of external pose contour, preventing background noise from being incorrectly translated into false tissue texture.

[0130] Preferably, inverse diffusion sampling is performed based on fused features, including alternating fusion sampling. Considering the competition between strong structural constraints and high-frequency detail enhancement, this application designs a dual-trajectory alternating sampling logic to reconstruct sharp skeletal details while suppressing scattering artifacts. Specifically, this includes parallel evolution of structure-preserving trajectories and bone-enhancing trajectories, where the structure-preserving trajectory preserves anatomical geometric boundaries, and the bone-enhancing trajectory enhances skeletal texture details. The sampling process starts from a shared initial noise distribution x T Starting from ~N(0,I), two complementary reconstruction trajectories are advanced in parallel within the same iterative chain. The inference algorithm performs conditional score estimation at each iterative sampling step using the following formula:

[0131]

[0132] in, The conditional input for the structure-preserving trajectory is a contour feature map. The conditional input for the bone enhancement trajectory is empty. When executing the structure-preserving trajectory, the contour feature map is used as the conditional input, and the score estimation is strongly constrained by CNCSN to lock the anatomical geometric boundaries of the generated image and ensure that the soft tissue is not deformed. When executing the bone enhancement trajectory, the conditional input is left empty, disabling CNCSN, thereby utilizing the original generative capability of the adversarial diffusion pre-trained model to enhance the high-frequency texture details of the trabeculae and cortex in the reconstructed image, resulting in a sharper bone texture.

[0133] Furthermore, the method also includes semantic-level fusion of the outputs of the structure-preserving trajectory and the skeleton-enhanced trajectory. At the sampling termination time (t=0), the system obtains the output images of the structure-preserving trajectory, respectively. Output image of bone enhancement trajectory To combine the advantages of both, the following operation is performed: Step S2073, threshold segmentation is performed on the output image of the bone enhancement trajectory to obtain the bone mask. In this embodiment, adaptive threshold segmentation (the threshold is denoted as ) is applied to the output image of the bone enhancement trajectory. Extract the bone mask representing high-density bone regions. Threshold The threshold is automatically determined using the Otsu adaptive threshold algorithm, but can also be manually set according to the specific anatomical location. For example, the typical value range is 0.6-0.8 after normalization. This threshold range corresponds to the typical HU value range of the bone cortex (approximately 200-400 HU), and can be adjusted according to the bone mineral density characteristics of different anatomical locations. Step S2074: Obtain the final enhanced CT image according to the following formula:

[0134]

[0135] in, To ultimately enhance the image, For skeletal mask, The output image for the skeleton augmentation trajectory. The output image preserves the trajectory of the structure. Through this semantic-level Hadamard fusion, the final output image uses enhanced sharp details in the skeletal regions and structurally stable, fidelity-preserving features in the non-skeletal regions, thus solving the two major challenges of "pose illusion" and "detail blurring" in a single translation task.

[0136] The technical effects of the embodiments of this application will be verified and explained below.

[0137] This embodiment conducted comparative experiments on multiple standard datasets and real clinical cases. For quantitative performance evaluation, Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity Index (SSIM) were used as objective evaluation metrics. Table 1 shows the comparison results between this invention and several mainstream medical image translation methods (all calculated within the ROI region), where PSNR is in dB and SSIM is a dimensionless index with a value range of [0,1].

[0138] like Figure 5 As shown, this is a performance comparison of different algorithms on the pelvic and spine datasets.

[0139] Experimental results show that this application achieves a PSNR improvement of approximately 3.94 dB over the state-of-the-art diffusion model method (TPDM) and significantly outperforms GAN-based methods (Pix2Pix, etc.) on SSIM. This demonstrates the strong constraint effect of the CNCSN module on anatomical structures and the superior performance of alternating fusion sampling in texture restoration.

[0140] In downstream clinical validation, the enhanced images were used to perform automated organ segmentation experiments. Table 2 compares the segmentation accuracy (Dice coefficient) of the original weight-bearing CBCT images with the enhanced images of this invention. The Dice coefficient is a dimensionless index with a value range of [0,1].

[0141] like Figure 6 The table shown is a comparison of Dice coefficients for the main anatomical regions.

[0142] Especially in the cardiac and femoral regions, due to the extremely severe scattering artifacts in the original weight-bearing CBCT, conventional algorithms are completely unable to achieve effective segmentation (Dice is close to 0). However, this application achieves anatomical and functional reconstruction of these regions by introducing spiral CT priors.

[0143] In terms of computational efficiency, this application reduces the inference time to 0.32 seconds per frame, which is approximately 10 times faster than similar diffusion model methods (TPDM's 3.22 seconds). The video memory usage is only about 3461 MB, and the number of parameters is 54.98 MB. Table 3 compares the inference overhead of different methods, where inference time is in seconds, parameter count is in MB, and video memory usage is in MB.

[0144] like Figure 7 The table shown is a comparison of the computational complexity of different methods.

[0145] like Figure 3 As shown, this embodiment extracts the HU value intensity distribution curve for the lumbar facet joint region and performs quantitative comparative analysis between the original weight-bearing CBCT image and the enhanced image of this application. Figure 3 In the diagram, the blue curve (Ours) represents the HU value distribution of the image enhanced by the method of this application, and the orange curve (WBCBCT) represents the HU value distribution of the original negative-weight CBCT image.

[0146] Figure 3 (a) shows the HU value intensity distribution curve along the horizontal direction of the lumbar spine cross section. Figure 3 (a) The small image in the upper right corner shows the specific location of the sampling line, which spans the entire lumbar spine slice, covering soft tissue, bone, and intervertebral structures. The curve comparison shows that the original weight-bearing CBCT (orange curve) exhibits a severe HU value shift in the soft tissue region (light red shaded area in the image, approximately 50-150 pixels), showing a significant negative deviation, with drastic curve fluctuations and extremely low contrast. However, after enhancement using the method described in this application (blue curve), the HU value in the soft tissue region recovers to a normal level close to 0 HU, and the curve becomes smooth and stable. In the bone region (peak areas approximately 0-50 and 200-250 pixels), the enhanced HU value peaks are sharper and clearer, and the contrast of the bone cortex boundary is significantly improved.

[0147] Figure 3 (b) Shows the local magnified HU value distribution curve of the lumbar facet joint region. Figure 3 (b) The small image in the upper right corner shows the specific location of this local area, focusing on the bilateral facet joints. The curve comparison shows that the original weight-bearing CBCT (orange curve) has a HU value of only about 0-100 HU in the joint space region (light green shaded area in the image, approximately 22-30 pixels), and the boundaries of the facet joints are blurred. However, after enhancement using the method of this application (blue curve), the peak HU value of the facet joint region reaches 600-700 HU, approaching the HU value level of standard spiral CT for bone. The contrast between the joint space and bone is improved by about 2.5 times, and the edge sharpness of the facet joints is significantly enhanced, allowing for clear differentiation of the fine structures of the articular surfaces.

[0148] The quantitative analysis results of the above HU value intensity distribution curves show that the method of this application can effectively suppress scattering artifacts in heavy-duty CBCT images, restore the true HU value distribution of soft tissue and bone regions, and significantly improve image contrast and the recognizability of anatomical structures.

[0149] like Figure 4 As shown, this embodiment verifies the necessity of each core module through ablation experiments. Figure 4 The images, from left to right, show the effects of: original weight-bearing CBCT input (WBCBCT), baseline method, using only CNCSN (W / CNCSN), using CNCSN and MDS (W / CNCSN, MDS), and the complete solution (All modules). The green ROI boxes focus on skeletal structures, and the yellow ROI boxes focus on soft tissue structures. The display window is [-1000, 1000] HU. Experimental results show that without CNCSN constraints, the generated images tend to produce pose illusions, mistaking the compressed state of small joints in a weight-bearing position for a relaxed state in a supine position; without masked diffusion sampling, numerous radial stripe artifacts and noise patches appear on the outer edges of the image; without alternating fusion sampling, obvious artifacts and ghosting appear at the bone edges, and high-frequency details such as trabeculae become blurred.

[0150] Furthermore, this embodiment compares the differences between injecting control information during the encoding stage and injecting it during the decoding stage. Experiments show that injecting control signals during the decoding stage (at the skip connection) better protects image contrast and effectively avoids the potential damage to low-level features that might result from injection during the encoding stage. This verifies the rationality of the design choice in this application to inject pose features into the backbone decoder.

[0151] This application also provides a computing device, including: a memory for storing computer-executable instructions; and a processor for executing the computer-executable instructions to implement the steps of the above-described method.

[0152] This application also provides a weighted cone-beam CT enhancement device, including:

[0153] The image acquisition module is used to acquire the negative-position CBCT source image to be processed;

[0154] The edge detection module is used to perform edge detection on the CBCT source images under load to obtain contour feature maps; the contour feature maps represent the patient's posture information under load.

[0155] The forward diffusion module is used to perform forward diffusion processing on negative CBCT source images to obtain noisy images;

[0156] The channel stitching module is used to stitch the contour feature map with the noisy image through channels to obtain the stitching tensor.

[0157] The attitude feature extraction module is used to input the spliced ​​tensor into the control encoder to obtain attitude features;

[0158] The feature fusion module is used to inject pose features into the backbone decoder of the pre-trained diffusion model via skip connections through zero convolutional layers to obtain fused features; wherein, the pre-trained diffusion model is trained based on the supine position paired dataset.

[0159] The image generation module is used to perform reverse diffusion sampling based on fusion features to obtain enhanced negative-weight CT images.

[0160] Furthermore, the pre-trained diffusion model is a denoising diffusion generative adversarial network, including a conditional generator and a discriminator;

[0161] The condition generator models the back transition probability implicitly, and its expression is:

[0162]

[0163] in, For the noisy image at step t, This is the image at step t-1, where t is the diffusion time step. For contour feature map, The image is a negative-weight CBCT source image. For latent variables that follow a standard normal distribution, for The prior distribution, For condition generator, For a clear image of the prediction;

[0164] The condition generator is trained using a variance-preserving marginal distribution that satisfies:

[0165]

[0166] in, For the clear image Noisy image at step t The conditional probability distribution, Indicates a Gaussian distribution. This is the cumulative signal preservation factor. It is the identity matrix. =exp(- ), variance scheduling function >0.

[0167] Furthermore, the pre-trained diffusion model also includes a backbone encoder corresponding to the backbone decoder; the feature fusion module is specifically used to: establish multi-scale skip connections between the control encoder and the backbone decoder.

[0168] Set up zero convolutional layers on each skip connection path; where the weights and biases of the zero convolutional layers are initialized to zero;

[0169] The fusion features at the i-th resolution scale are obtained using the following formula:

[0170]

[0171] in, To fuse feature maps, This is the locked feature map of the backbone encoder at the i-th scale. To control the encoder's pose feature map at the i-th scale, For zero convolution operators, These are the parameters for a zero convolutional layer.

[0172] Furthermore, during the training process of the pre-trained diffusion model, the loss function includes adversarial loss and cycle consistency loss:

[0173]

[0174] in, For the total loss, To combat the losses, For cycle consistency loss, These are the weighting coefficients.

[0175] Furthermore, the edge detection module is specifically used for:

[0176] The Canny edge detection operator is used to process the negative CBCT source image. Based on the preset high and low thresholds, the edges are connected by the hysteresis thresholding method to obtain a binarized contour feature map.

[0177] Furthermore, the image generation module is also used to perform mask diffusion sampling: adaptive thresholding is performed on the negative-weight CBCT source image to obtain an anatomical mask; wherein, the pixel value of the tissue region in the anatomical mask is 1, and the pixel value of the background region is 0;

[0178] In each iteration of the backdiffusion sampling, the predicted image is projected according to the following formula:

[0179]

[0180] in, The projected image. For dissection masking, For the image predicted by the model, For background pixels, For Hadamard products.

[0181] Furthermore, the image generation module is also used to perform alternating fusion sampling:

[0182] Parallel evolution of structure-preserving trajectories and bone-enhancing trajectories; wherein, the structure-preserving trajectory is used to preserve anatomical geometric boundaries, and the bone-enhancing trajectory is used to enhance bone texture details;

[0183] When performing structure-preserving trajectory execution, the contour feature map is used as a conditional input;

[0184] When executing the skeletal enhancement trajectory, set the conditional input to empty;

[0185] Threshold segmentation is performed on the output image of the skeleton enhancement trajectory to obtain the skeleton mask;

[0186] The outputs of the structure-preserving trajectory and the bone enhancement trajectory are semantically fused according to the following formula to obtain the final enhanced CT image:

[0187]

[0188] in, To ultimately enhance the image, For skeletal mask, The output image for the skeleton augmentation trajectory. Output image that preserves the trajectory of the structure. For Hadamard products.

[0189] Furthermore, the device also includes a preprocessing module for processing the negative-position CBCT source images after acquisition:

[0190] The CT value is truncated in the negative CBCT source image to limit the pixel value to a preset display window range;

[0191] The truncated image is normalized.

[0192] It should be noted that the weight-bearing cone-beam CT enhancement device provided in this application embodiment is used to perform the above method steps.

[0193] This application also provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the steps of the above-described method.

[0194] Finally, it should be noted that the above descriptions are merely preferred embodiments of this application and are not intended to limit this application. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. A method for stress cone beam CT enhancement, applied to an image processing device, comprising: include: ​ Acquire the negative-weight CBCT source image to be processed; Edge detection is performed on the CBCT source image under load to obtain a contour feature map; wherein, the contour feature map represents the patient's posture information under load. The negative-weight CBCT source image is subjected to forward diffusion processing to obtain a noisy image; The contour feature map is concatenated with the noisy image through channels to obtain a concatenated tensor. The spliced ​​tensor is input into the control encoder to obtain attitude features; The pose features are injected into the backbone decoder of a pre-trained diffusion model via skip connections through zero convolutional layers to obtain fused features; wherein the pre-trained diffusion model is trained on a supine paired dataset. Based on the fusion features, reverse diffusion sampling is performed to obtain enhanced negative-weight CT images.

2. The method of claim 1, wherein, The pre-trained diffusion model is a denoised diffusion generative adversarial network, which includes a conditional generator and a discriminator; The condition generator models the back transition probability implicitly, and its expression is: wherein, is the noisy image for the t-th step, is the image for the t-1-th step, t is the diffusion time step, is the contour feature map, is the negative weight CBCT source image, is the latent variable following a standard normal distribution, is the prior distribution of is the conditional generator, is the posterior distribution, is the predicted clean image; The condition generator is trained using a variance-preserving marginal distribution that satisfies: in, For the clear image Noisy image at step t The conditional probability distribution, Indicates a Gaussian distribution. This is the cumulative signal preservation factor. It is the identity matrix. =exp(- ), variance scheduling function >

0.

3. The method according to claim 1, characterized in that, The pre-trained diffusion model also includes a backbone encoder corresponding to the backbone decoder; The process of injecting the pose features into the backbone decoder of a pre-trained diffusion model via skip connections through zero convolutional layers to obtain fused features includes: Establish a multi-scale jump connection between the control encoder and the backbone decoder; A zero-convolutional layer is set on each of the paths of the skip connections; wherein the weights and biases of the zero-convolutional layer are initialized to zero; The fusion features at the i-th resolution scale are obtained using the following formula: in, To fuse feature maps, This is the locked feature map of the backbone encoder at the i-th scale. To control the encoder's pose feature map at the i-th scale, For zero convolution operators, These are the parameters for a zero convolutional layer.

4. The method according to claim 2, characterized in that, During the training of the pre-trained diffusion model, the loss function includes adversarial loss and cycle consistency loss: in, For the total loss, To combat the losses, For cycle consistency loss, These are the weighting coefficients.

5. The method according to claim 1, characterized in that, The step of performing edge detection on the negative-position CBCT source image to obtain a contour feature map includes: The negative CBCT source image is processed using the Canny edge detection operator. Based on preset high and low thresholds, the edges are connected using the hysteresis thresholding method to obtain a binarized contour feature map.

6. The method according to claim 1, characterized in that, The reverse diffusion sampling based on the fused features includes performing mask diffusion sampling: Adaptive threshold segmentation is performed on the negative-position CBCT source image to obtain an anatomical mask; wherein, in the anatomical mask, the pixel value of the tissue region is 1, and the pixel value of the background region is 0; In each iteration of the backdiffusion sampling, the predicted image is projected according to the following formula: in, The projected image. This is the anatomical mask. For the image predicted by the model, For background pixels, For Hadamard products.

7. The method according to claim 1, characterized in that, The reverse diffusion sampling based on the fusion features includes performing alternating fusion sampling: Parallel evolution of structure-preserving trajectories and bone-enhancing trajectories; wherein the structure-preserving trajectory is used to preserve anatomical geometric boundaries, and the bone-enhancing trajectory is used to enhance bone texture details; When executing the structure-preserving trajectory, the contour feature map is used as a conditional input; When executing the bone enhancement trajectory, set the conditional input to empty; Threshold segmentation is performed on the output image of the bone enhancement trajectory to obtain the bone mask; The outputs of the structure-preserving trajectory and the bone enhancement trajectory are semantically fused according to the following formula to obtain the final enhanced CT image: in, To ultimately enhance the image, For the skeleton mask, The output image of the bone enhancement trajectory. The output image that preserves the trajectory of the structure. For Hadamard products.

8. The method according to claim 1, characterized in that, After acquiring the negative-weight CBCT source image to be processed, the method further includes: The CT value is truncated in the negative CBCT source image to limit the pixel value to a preset display window range; The truncated image is normalized.

9. A weighted cone-beam CT enhancement device, characterized in that, include: The image acquisition module is used to acquire the negative-position CBCT source image to be processed; An edge detection module is used to perform edge detection on the weight-bearing CBCT source image to obtain a contour feature map; wherein, the contour feature map represents the patient's posture information under weight-bearing conditions; The forward diffusion module is used to perform forward diffusion processing on the negative-position CBCT source image to obtain a noisy image; The channel stitching module is used to perform channel stitching between the contour feature map and the noisy image to obtain a stitching tensor. The attitude feature extraction module is used to input the spliced ​​tensor into the control encoder to obtain attitude features; The feature fusion module is used to inject the pose features into the backbone decoder of a pre-trained diffusion model via a skip connection through a zero convolutional layer to obtain fused features; wherein, the pre-trained diffusion model is trained based on a supine paired dataset; The image generation module is used to perform reverse diffusion sampling based on the fusion features to obtain enhanced negative-weight CT images.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method as described in any one of claims 1 to 8.

Images