CBCT high-quality CT image synthesis method based on structure prior guidance

Abstract

The invention discloses a CBCT (Cone Beam Computed Tomography) high-quality CT (Computed Tomography) image synthesis method based on structure prior guidance, which comprises the following construction steps of patient data collection and arrangement, space guidance deformable attention convolution, structure prior guidance type conditional diffusion image synthesis and reasoning process and quality evaluation. The method specifically comprises the steps of technical design of a space guiding module, technical design of a deformable sampling kernel generator, technical design of an attention weight modulator, contour extraction operation, diffusion modeling operation, conditional denoising network design and target training. The method provided by the invention has the advantages of structural perception, high spatial selectivity, high modulability and the like; and the reduction precision of the model on the key anatomical region is enhanced by taking a contour map as a structure priori condition. The whole diffusion generation network realizes unification of local detail reservation and global structure stability while maintaining noise robustness, and gradually generates a high-quality pseudo CT image which is close to a real CT modal and has excellent structure consistency.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of medical diagnosis and treatment, and particularly relates to a method for synthesizing high-quality CT images based on structure prior guidance of CBCT. BACKGROUND

[0002] In modern medical diagnosis and treatment, medical imaging technology plays a crucial role. Computerized tomography (CT) and cone beam computed tomography (CBCT) are two important imaging methods, each with unique advantages and application scenarios. CT imaging technology, with its outstanding performance, plays a key role in clinical practice. It can provide high-resolution images that clearly show the fine internal anatomical structures of the human body, providing strong support for doctors to accurately diagnose diseases. Its low artifact feature minimizes image interference, allowing doctors to accurately identify diseased areas and reduce the risk of misdiagnosis and missed diagnosis. Excellent soft tissue contrast allows doctors to clearly distinguish between different tissues, especially in tumor diagnosis, neurological disease examination, and other aspects, helping doctors to comprehensively assess the disease and develop precise treatment plans. However, CT technology has the significant drawback of high radiation dose, and long-term or frequent CT scanning may pose potential radiation hazards to patients, which to some extent limits its application in certain scenarios. CBCT technology has developed rapidly in recent years and has shown unique value in specific clinical fields. It has low radiation dose, effectively reducing the risk of radiation exposure for patients, especially for patients sensitive to radiation or those requiring multiple imaging. Its fast imaging speed allows it to obtain images in a short time, meeting the needs of rapid clinical diagnosis, and it has obvious advantages in emergency and intraoperative imaging. In radiotherapy, it helps doctors accurately determine the treatment target and improve treatment effectiveness. However, it is important to note that the CBCT imaging process is susceptible to scattering and noise, resulting in poor reconstructed image quality, with many artifacts and high noise levels, which severely affects image readability and diagnostic accuracy, greatly limiting its widespread application in clinical environments with extremely high image quality requirements. Given the characteristics of CT and CBCT, combining the advantages of both through technical means has become an important research direction in the field of medical imaging.

[0003] It has great clinical significance and application prospect to realize the synthesis of CBCT medical images into CT images by using deep learning technology. It is expected to reduce the radiation dose received by patients while ensuring the accuracy of diagnosis, and to obtain images of similar CT quality, making up for the shortcomings of CBCT image quality, and providing better and reliable image support for clinical diagnosis and treatment. However, there are still many limitations in the application of current existing deep learning methods in this field, such as lack of universality and generalization ability, difficulty in adapting to complex and variable clinical data and individual differences of different patients, complex model training process, high requirement for computing resources, low training efficiency, and the quality of generated synthetic images still needs to be further improved when dealing with image details and complex structures. Although the existing mainstream unsupervised image conversion methods such as CycleGAN and CUT can perform style transfer between unpaired images, there is often a non-negligible structural offset between different modalities in actual medical images, making it difficult to preserve the original anatomical structure of the generated image. The core of the medical image translation task is the fidelity of the structure, so how to introduce structural constraints without the help of paired supervision becomes the key to this task. SUMMARY

[0004] (1) Technical problems to be solved

[0005] The purpose of the present application is to realize the synthesis of CBCT medical images into CT images by using deep learning technology, aiming to obtain CT quality images with lower radiation dose, and to solve the limitations of current deep learning methods in this field, such as: existing methods generally lack sufficient universality and generalization ability, making it difficult to adapt to complex and variable clinical data and individual differences of different patients, limiting their clinical applicability; the model training process is usually complex and requires high computing resources, with low training efficiency, raising the threshold for technical application; the quality of the generated synthetic images by existing methods still has shortcomings when dealing with image details and complex structures, making it difficult to meet the stringent requirements of anatomical structure fidelity for clinical diagnosis. Thus ensuring the accuracy of the generated image in anatomical structure, solving the problems existing in the prior art.

[0006] (2) Technical solutions

[0007] In order to solve the above technical problems, the present application discloses a CBCT synthesis high-quality CT image method based on structure prior guidance, comprising the following steps:

[0008] I. Patient data collection and sorting:

[0009] Data of 200 patients were collected, including different stages, pathological types, tumor locations and age groups. The data covered CBCT images and non-paired CT images of different patients. The data was expanded to 3 times the original size using a Gaussian noise data enhancement method to improve the model generalization ability. Finally, the training set, test set and validation set were divided according to the ratio of 7:1:2, and the image database was constructed.

[0010] II. Structure prior guided conditional diffusion image synthesis

[0011] 1) Contour extraction operation:

[0012] The Canny edge detection operator was used to process the input CBCT image to extract the main boundary information of the tissue and organs, and generate a structure contour map C.

[0013] 2) Diffusion modeling operation:

[0014] In the structure guided image generation framework, the image generation process is based on the diffusion probability mechanism of the symmetric modeling of forward diffusion-reverse denoising. Through a Markov chain, the Gaussian noise is gradually converted into the target image distribution in the latent space, wherein the forward process is a pre-defined random process, and the core goal is to convert any given original image into a noise image close to Gaussian distribution , which is the starting point of sampling. At each step t, a small amount of Gaussian noise is added to the current image, gradually destroying its structure and texture, and finally approaching pure noise. Its probability transition function is defined as follows:

[0015]

[0016] wherein, denotes the cumulative noise attenuation coefficient;

[0017] 3) Conditional denoising network design:

[0018] In each reverse sampling step of the diffusion modeling operation, a conditional denoiser is designed to predict the noise component and gradually restore the target image. The conditional denoising network fuses the following three types of information for joint modeling:

[0019] ① Time step information t: The dynamic regulation ability of the diffusion stage is introduced through the embedding mechanism, so that the network has stage perception ability at different denoising depths;

[0020] ② Structure contour prior : The contour map extracted from the input CBCT image is used as a structure guided condition to guide the network to focus on key anatomical areas and improve the geometric consistency of the restoration;

[0021] 4) Training objectives:

[0022] Optimize mean squared error loss:

[0023] L diff = E X 0 , c , t , ε ~ N (0, I ) [ ε θ X t , t , C -ε 2 2 ]

[0024] Through this mechanism, each diffusion sampling step is guided by a contour map, so that the generated image maintains the consistency of anatomical structure to the greatest extent during modality transformation;

[0025] III. Spatial Guided Deformable Convolution:

[0026] 1) Technical design of the space guidance module:

[0027] From the input feature map The system actively extracts spatial prior information related to image structure to generate a high-resolution guidance map, which is used to accurately identify spatial locations in the image where there are drastic structural changes, significant edges, or prominent anatomical boundaries.

[0028] 2) Technical design of deformable sampling kernel generator:

[0029] Based on the structural response information provided by the guidance map obtained in step 1), the actual receptive position of each convolution sampling point in space is dynamically adjusted to construct an adaptive deformation spatial sampling mechanism; the receptive field of this spatial sampling mechanism has position sensitivity and geometric matching capability in space.

[0030] An adaptive displacement vector field is generated based on the structural response map generated by the spatial guidance module to precisely control the sampling position of each convolution point, enabling the convolution kernel to automatically adapt to different structural geometries. Specifically, this includes the following steps:

[0031] ① Input guide diagram The structural response map generated by the spatial guidance module serves as the basis for offset prediction;

[0032] ② Migration field prediction network:

[0033] Constructing small CNN subnetworks ,right Perform mapping to output the offset vector field:

[0034]

[0035] Where K represents the number of sampling points, such as K=9 corresponding to a 3×3 convolution, with K two-dimensional offset vectors at each position;

[0036] ③Sampling location reconstruction:

[0037] For each basic sampling point The final sampling location is:

[0038]

[0039] This operation ensures that the convolution kernel has a unique sampling path at each pixel, enabling the construction of non-uniform adaptive receptive fields based on the image structure;

[0040] ④ Bilinear interpolation:

[0041] Since the offset can be a non-integer coordinate, bilinear interpolation is used to sample features at the misaligned positions:

[0042]

[0043] in Represents the nearest integer grid point. For interpolation weights;

[0044] Through the above mechanism, the model can dynamically reconstruct the receptive field based on structural information, achieving geometric alignment perception and boundary-sensitive modeling.

[0045] 3) Pay attention to the technical design of the weighted modulator:

[0046] A spatial-channel attention mechanism is introduced at each deformable sampling point to further control the feature response intensity at each location. Specifically, this includes the following steps:

[0047] ① Contextual semantic compression:

[0048] For the input feature map Perform global average pooling (GAP) to obtain the semantic description vector:

[0049]

[0050] ②Attention-generating network:

[0051] Construct two layers of nonlinear mapping to generate attention vectors for sampling points:

[0052]

[0053] in, For learnable parameters, Activated for ReLU For the Sigmoid function, note the vector. Each component represents the weighting coefficient of the corresponding variable sampling point;

[0054] ③ Deformable convolution fusion output:

[0055] Final output features:

[0056]

[0057] Each sampling point is weighted and then used to calculate the output features, thus forming a "structure-oriented response regulation".

[0058] IV. Reasoning Process and Quality Assessment

[0059] Using CBCT images as input, contour maps extracted using the Canny operator are used as a guide, and then the SCG-CT model is progressively sampled to generate pseudo-CT images. Evaluation metrics include:

[0060] Mean absolute error, root mean square error, structural similarity, and peak signal-to-noise ratio;

[0061] Then, the synthetic images were compared with real CT scans to verify the model's performance in terms of image fidelity and structural consistency.

[0062] Preferably, the technical design approach for the spatial guidance module is as follows:

[0063] ① Extract semantic saliency cues from the original feature map:

[0064] First, a spatial attention path is constructed through a series of lightweight convolutional operations. This path is dedicated to mining information such as texture abrupt changes, edge intensity gradients and context-aware differences in the current input feature map.

[0065] ② Construct a semantic structure response graph:

[0066] The output is a single-channel feature map. The response value of each pixel represents the importance of that location in terms of spatial structure to subsequent modeling, or its "structural sensitivity". The larger the value, the more complex the structure of the region or the more obvious the geometric abrupt change.

[0067] As a deformable kernel and a signal source for attention weight modulation:

[0068] The generated guidance map is not only used to predict the offset coordinates of subsequent convolutional sampling kernels, i.e., the source of the "bias field" of deformable convolution, but also serves as the core reference for adjusting the response intensity in the attention mechanism, ensuring that the model learns key structural regions with high accuracy.

[0069] Preferably, a conditional embedding residual block is added to the spatially guided deformable attentional convolution in step two to generate the basic structure of the image model. Temporal embedding and structural and semantic information of the image need to be processed. Image information is processed by connecting spatially guided deformable attentional convolution and regular convolution.

[0070] Preferably, the conditional denoising network design also includes step ③, which is a feature extraction network. The feature extraction network is composed of spatially guided deformable attentional convolution and conditional embedding residual blocks from step 2. It has dynamic sampling and regional response modulation capabilities and can adaptively capture structural boundaries and semantic key regions.

[0071] Preferably, an explicit structural constraint mechanism is introduced into the diffusion generation mechanism. In each diffusion iteration step, the contour information extracted from the input CBCT image is used as the structural prior condition. As the image is gradually restored from Gaussian noise to the target CT modality, the generation path is continuously guided so that it always fits the real organ boundary and tissue morphology, preventing geometric distortion or anatomical distortion caused by modal drift, texture replacement, etc.

[0072] Preferably, the forward process includes a closed analytical solution, allowing for direct sampling from a single weighted sample without explicitly sampling each step during the training phase. Get any .

[0073] (3) Beneficial effects

[0074] To improve the structural fidelity and clinical applicability of CBCT images in the process of synthesizing pseudo-CT images, this invention proposes a structure prior-guided diffusion generation method (SCG-CT) that integrates an adaptive perception mechanism of spatially guided deformable attentional convolution (SGDAC). This method is built on the framework of conditional diffusion model, and jointly models explicit structural priors and content response modulation. Without relying on image pairing, it effectively alleviates the deformation distortion and anatomical loss problems that are prone to occur in existing unsupervised image conversion methods when dealing with structural differences between modalities.

[0075] Specifically, the spatially guided deformable attentional convolution module, by introducing a spatial guidance mechanism and deformable sampling kernels, not only enhances the model's adaptive modeling ability for complex structural regions in images, but also significantly improves its understanding of semantic changes between heterogeneous modalities. Introducing this mechanism during the decoding stage allows for dynamic adjustment of the feature upsampling path, ensuring that the generated image maintains boundary coherence and structural continuity during spatial reconstruction, thereby guaranteeing anatomical consistency during modality transfer.

[0076] Compared to traditional U-Net or fixed convolutional structures, the spatially guided deformable attentional convolutional component proposed in this invention possesses advantages such as structure awareness, strong spatial selectivity, and high moduliability. Combining contour maps as structural priors further enhances the model's accuracy in reconstructing key anatomical regions. The entire diffusion-generative network maintains noise robustness while achieving a balance between local detail preservation and global structural stability, providing a more stable, reliable, and medically interpretable solution for CBCT pseudo-CT synthesis tasks. Attached Figure Description

[0077] Appendix Fig. 1 This is a schematic diagram of the spatially guided deformable attentional convolution and extended residual block of the present invention;

[0078] Appendix Fig. 2 This is a diagram of the structure-prior guided conditional diffusion image synthesis method and image generation network structure of the present invention. Detailed Implementation

[0079] The present invention will be described in detail below using different manufacturing processes. The examples listed will enable those skilled in the art to better understand the present invention, but will not limit the present invention in any way. For example, the present invention can be extended to the clinical diagnosis and treatment applications of CBCT to replace CT quality images in order to make up for the deficiencies in CBCT image quality.

[0080] Example 1:

[0081] A method for synthesizing high-quality CT images using CBCT based on structural prior guidance includes the following steps:

[0082] To address the challenge of synthesizing pseudo-CT images from CBCT images, a novel method for synthesizing high-quality CT images based on structural prior guidance, abbreviated as SCG-CT (Structure-Constrained Generation for CT), is proposed. This method integrates contour structure prior constraints with a spatially guided deformable attentional convolution mechanism (SGDAC) and embeds it into a conditional diffusion model architecture. Under unpaired supervision, it achieves the transformation of CBCT images into high-fidelity CT images, balancing anatomical consistency and texture realism.

[0083] Step 1: Patient Data Collection and Organization

[0084] Data from 200 patients were collected, including different stages, pathological types, tumor locations, and age groups. The data encompassed CBCT images and unpaired CT images from various patients. Gaussian noise data augmentation was used to augment the data to three times the original size to improve the model's generalization ability. Finally, the image database was constructed by dividing the dataset into training, test, and validation sets in a 7:1:2 ratio.

[0085] Step 2: Spatially Guided Deformable Attention Convolution (SGDAC)

[0086] To enhance the ability of neural networks to model structural details in image synthesis tasks, this invention proposes a Spatially-Guided Deformable Attention Convolution (SGDAC) module, which aims to combine global guidance information with local structural changes to achieve spatially adaptive feature response modulation. The SGDAC module consists of three core substructures:

[0087] A. Spatial Guidance Module

[0088] This module is designed to extract features from the input feature map. The system actively extracts spatial prior information related to image structure to generate a high-resolution guidance map. This map is used to accurately identify spatial regions in the image where there are drastic structural changes, significant edges, or prominent anatomical boundaries. It should be noted that, as shown in the attached diagram... Fig. 2 As shown, the Spatial Guided Deformable Attention Convolution (SGDAC) module operates at multiple locations in the encoder-decoder network of the diffusion model in the method. The input features refer to the features input to this module in general, and the symbolic representation of X can be removed.

[0089] In practical medical image synthesis tasks, the geometric complexity and structural sensitivity of different regions in an image vary significantly. For example, there are often obvious intensity transitions and texture changes at organ edges, lesion boundaries, or cavity-tissue junctions. If these regions are not handled properly during synthesis, structural distortions or blurring artifacts can easily occur. Conventional convolution operations, due to their smooth local perception mechanism, struggle to explicitly identify these regions, thus affecting the anatomical fidelity of the generated image.

[0090] To overcome this problem, this module is designed based on the following technical approach:

[0091] 1. Extract semantic saliency cues from the original feature map:

[0092] First, a spatial attention path is constructed through a series of lightweight convolutional operations (such as 3×3 convolution + BN + ReLU + 1×1 compressed convolution). This path aims to extract information such as texture abrupt changes, edge intensity gradients, and context-aware differences in the current input feature map.

[0093] 2. Construct a semantic structure response map (Guidance Map):

[0094] The output is a single-channel feature map. The response value of each pixel represents the importance of that location in terms of spatial structure for subsequent modeling, or its "structural sensitivity". The larger the value, the more complex the structure of the region or the more obvious the geometric abrupt change.

[0095] 3. As a deformable kernel and a signal source for attention weight modulation:

[0096] This guiding graph is not only used to predict the offset coordinates of subsequent convolutional sampling kernels (i.e., the source of the "bias field" of deformable convolution), but also serves as a core reference for adjusting the response intensity in the attention mechanism, ensuring that the model learns key structural regions with high accuracy.

[0097] In this way, the spatial guidance module enables the construction of a "structural response map" from the original feature map, providing external control variables for the spatial geometry of the entire SGDAC module. This effectively overcomes the problem of insufficient modeling of structural changes by traditional static convolutional kernels, providing a clear, continuous, and differentiated spatial basis for subsequent convolutional sampling and attention weight allocation. Furthermore, this guidance map can be cascaded, fused, or residual-propagated across multiple network layers, enabling the accumulation of multi-layer perceptual structural discriminative power and achieving global constraints on multi-scale anatomical structures.

[0098] B. Deformable Offset Generator

[0099] This module is responsible for dynamically adjusting the actual receptive position of each convolutional sampling point in space based on the structural response information provided by the guidance diagram, thereby constructing an adaptive deformation-adaptive spatial sampling mechanism. This mechanism allows the convolutional kernel to break through the limitations of traditional mesh sampling, enabling its receptive field to have spatial position sensitivity and geometric matching capabilities. It is the core component of the entire SGDAC module in realizing "structurally adaptive convolution".

[0100] Traditional convolution operations use a fixed sampling template (e.g., 3×3) at each location, ignoring the actual geometric differences in the image content. This is especially problematic when processing edges, curved structures, or complex textures, easily leading to feature aliasing or local structural damage. Particularly in medical images, structural boundaries exhibit significant spatial variability and directional sensitivity, which fixed convolution receptive fields cannot capture.

[0101] Therefore, this module aims to generate an adaptive displacement vector field (offset map) based on the structural response map generated by the spatial guidance module, in order to precisely control the sampling position of each convolution point, so that the convolution kernel can automatically adapt to different structural geometries.

[0102] The specific implementation is as follows:

[0103] 1. Input guide map The structural response map generated by the spatial guidance module serves as the basis for offset prediction.

[0104] 2. Migration Field Prediction Network:

[0105] Constructing small CNN subnetworks ,right Perform mapping to output the offset vector field:

[0106]

[0107] Where K represents the number of sampling points (e.g., K=9 corresponds to a 3×3 convolution), and each position has K two-dimensional offset vectors.

[0108] 3. Sampling location reconstruction:

[0109] For each basic sampling point The final sampling location is:

[0110]

[0111] This operation ensures that the convolution kernel has a unique sampling path at each pixel, enabling the construction of a non-uniform adaptive receptive field based on the image structure.

[0112] 4. Bilinear interpolation:

[0113] Since the offset can be a non-integer coordinate, bilinear interpolation is used to sample features at the misaligned positions:

[0114]

[0115] in Represents the nearest integer grid point. These are the interpolation weights.

[0116] Through the above mechanism, the model can dynamically reconstruct the receptive field based on structural information, realize geometric alignment perception and boundary-sensitive modeling, which is a key means to ensure structural consistency in the CBCT synthetic CT process.

[0117] C. Note the Attention Weight Modulator

[0118] This module aims to introduce a spatial-channel attention mechanism at each deformable sampling point to further control the feature response intensity at each location. Its goal is to enhance the response of regions highly correlated with structure or semantics, while suppressing interference from background, low-correlation regions, or artifact regions, thus achieving a structure-prioritized, semantically focused response modulation mechanism.

[0119] In CBCT image synthesis tasks, the diagnostic value of different image regions varies significantly, with structures such as tissue boundaries, bone contours, and tumor regions playing a decisive role in the quality of the synthesis. Therefore, the model needs to be able to "consciously" increase the attention weight of these regions while weakening the processing of irrelevant or featureless background regions, thereby improving the local fidelity and global coherence of the generated image.

[0120] This module provides a content-aware weighting mechanism by modeling the response intensity of offset sampling points, thereby regulating the explicit fusion weights at different structural locations.

[0121] Specific implementation:

[0122] 1. Contextual semantic compression:

[0123] For the input feature map Perform global average pooling (GAP) to obtain the semantic description vector:

[0124]

[0125] 2. Attention Generation Network:

[0126] Construct two layers of nonlinear mapping to generate attention vectors for sampling points:

[0127]

[0128] in, For learnable parameters, Activated for ReLU This is the Sigmoid function. Note the vector. Each component represents the weighting coefficient of the corresponding variable sampling point.

[0129] 3. Deformable convolutional fusion output

[0130] Final output features:

[0131]

[0132] Each sampling point is weighted and then used to calculate the output features, thus forming a "structure-oriented response regulation".

[0133] Through this module, SGDAC can not only adapt to structural deformations in different regions, but also "focus" on key anatomical locations, improving the model's selective expressive ability at the feature level and significantly enhancing the local contrast and boundary clarity of the generated images.

[0134] D. Spatial Guided Deformable Attention Convolution - Conditional Embedding Residual Block

[0135] Spatial guided deformable attention-based conditional embedding residual blocks are a fundamental model structure for image generation, requiring the handling of temporal embeddings and processing of the image's structural and semantic information. They utilize fully connected networks for temporal embeddings and spatial guided deformable attention-based convolutions and regular convolutions to process image information. The residual structure effectively achieves embedding and suppresses gradient vanishing.

[0136] Step 3: Structure-Prior Guided Conditional Diffusion Image Synthesis

[0137] To ensure that the images generated during modality conversion faithfully preserve the anatomical layout of the original CBCT images in terms of spatial structure, this invention introduces an explicit structural constraint mechanism into the diffusion generation mechanism to improve the structural preservation capability and clinical usability of the synthesized images.

[0138] The core idea of ​​this structural constraint mechanism is to use the contour information extracted from the input CBCT image as a structural prior condition in each diffusion iteration step. As the image is gradually restored from Gaussian noise to the target CT modality, the generation path is continuously guided so that it always fits the real organ boundary and tissue morphology, preventing geometric distortion or anatomical distortion caused by modal drift, texture replacement, etc.

[0139] This strategy not only breaks through the limitation of traditional diffusion models relying on paired data to learn image structure, but also improves the structural consistency and model generalization ability under unsupervised generation conditions. It is a key optimization of diffusion models in the field of medical imaging.

[0140] The specific implementation process is as follows:

[0141] 1. Contour Extraction

[0142] To introduce anatomical-level prior constraints and enhance the spatial structure preservation capability of the diffusion model during image reconstruction, this invention introduces an explicit structural contour map as a prior condition in the image generation process. Specifically, the classic Canny edge detection operator is used to process the input CBCT image, extracting the main boundary information of its tissues and organs to generate a structural contour map C, which guides the gradual reconstruction process of the image at each step of the diffusion model.

[0143] The Canny operator boasts advantages such as high boundary localization accuracy and strong robustness to noise, making it particularly suitable for clearly extracting edge information of structures such as bones, soft tissues, and cavities in medical images. The resulting contour map is highly consistent with the anatomical structure in space, providing a clear and well-defined structural framework for the denoising network. It serves as a spatial geometric reference for the model's generated images, thereby effectively avoiding structural deformation and misanatomical phenomena while maintaining texture transformation capabilities.

[0144] 2. Diffusion Modeling

[0145] In the structure-guided image generation framework adopted in this invention, the image generation process is based on the diffusion probability mechanism of forward diffusion-backward denoising symmetric modeling. The basic idea of ​​the diffusion model is to model image generation as a step-by-step "denoising" process, that is, to gradually transform Gaussian noise into the target image distribution in the latent space through a Markov chain.

[0146] The forward process is a predefined stochastic process whose core objective is to transform any given original image... Convert into a noisy image that approximates a Gaussian distribution. This serves as the starting point for sampling. At each step t, a small amount of Gaussian noise is added to the current image, gradually disrupting its structure and texture until it approaches pure noise. Its probability transfer function is defined as follows:

[0147]

[0148] in, This represents the cumulative noise attenuation coefficient.

[0149] This forward process has a closed analytical solution, allowing for direct sampling from each step without explicit sampling during the training phase, using only a single weighted sampling. Get any This greatly simplifies the training process.

[0150] 3. Conditional Denoising Network Design

[0151] To achieve high-quality structure preservation and modal reproduction, this invention designs a conditional denoiser. This network is used to predict noise components and progressively reconstruct the target image at each backsampling step of the diffusion model. It integrates the following three types of information for joint modeling:

[0152] 1) Time step information t: The dynamic control capability of the diffusion stage is introduced through the embedding mechanism, so that the network has the ability to perceive the stage at different denoising depths;

[0153] 2) Structural profile prior The contour map extracted from the input CBCT image serves as a structural guidance condition, guiding the network to focus on key anatomical regions and improving the geometric consistency of the reconstruction.

[0154] 3) Feature extraction network: It is composed of spatially guided deformable attentional convolution-conditional embedding residual blocks proposed in this invention. It has dynamic sampling and regional response modulation capabilities and can adaptively capture structural boundaries and semantic key regions.

[0155] This conditional denoiser effectively maintains the consistency of anatomical structures while gradually restoring image details in the diffusion chain, making it a key module for ensuring image quality and structural fidelity.

[0156] 4. Training Objectives

[0157] Optimize mean squared error loss:

[0158] L diff = E X 0 , c , t , ε ~ N (0, I ) [ ε θ X t , t , C - ε 2 2 ]

[0159] This mechanism guides each diffusion sampling step with a contour map, ensuring that the generated image maintains anatomical consistency to the greatest extent possible during modality transformation.

[0160] Step 4: Reasoning Process and Quality Assessment

[0161] During the inference phase, CBCT images are used as input, and contour maps extracted using the Canny operator are used as guidance to gradually sample and generate pseudo-CT images by inputting them into the SCG-CT model.

[0162] Evaluation metrics include: Mean Absolute Error (MAE), Root Mean Square Error (RMSE), Structural Similarity (SSIM), and Peak Signal-to-Noise Ratio (PSNR).

[0163] By comparing the synthesized images with real CT scans, the performance of the model in terms of image fidelity and structural consistency is verified.

[0164] Specifically, CBCT and unpaired CT images from 200 patients at the same hospital were collected. Data from 140 patients was used as the training set to train the proposed structure-guided diffusion generation model; data from 20 patients was used as the validation set for optimal model selection during training; and data from the remaining 40 patients was used as the test set, which, after rigid and deformation registration, was used to quantitatively evaluate model performance. Model training was conducted on a server with an NVIDIA A6000 graphics card with 48GB of VRAM, with a maximum of 200 iterations and a learning rate of 0.0001. The model consists of two core parts: first, a U-Net structure incorporating Spatially Guided Deformable Attention Convolutional (SGDAC) for extracting semantic and structurally sensitive features; and second, a contour-prior-guided conditional diffusion module built upon this U-Net, which constrains and guides the generation process through edge structure maps.

[0165] During the inference phase, CBCT images of any patient are input. First, their structural contour maps are extracted as conditional inputs. Then, high-quality pseudo-CT images that are close to the real CT modality and have excellent structural consistency are generated step by step through the trained SCG-CT model. These images can be used for diagnostic reference and subsequent radiotherapy planning.

[0166] Table 1 shows a comparison of the original CBCT images and the pseudo-CT images generated after applying this method.

[0167] evaluation index original CBCT image pseudo-CT image formed after applying the present scheme mean absolute error 94.68 67.48 root mean square error 130.30 112.32 structural similarity 34.40 35.66 peak signal-to-noise ratio 0.9116 0.9560

[0168] The comparison of the above experimental data shows that the pseudo-CT images obtained using deep learning based on this invention can reduce the average absolute error between the pseudo-CT images and real CT scans by more than 28%. This indicates that the proposed solution has the potential to reduce the radiation dose received by patients while ensuring diagnostic accuracy, and simultaneously obtain images of similar quality to CT scans, thus compensating for the shortcomings of CBCT image quality and providing higher quality and more reliable image support for clinical diagnosis and treatment.

[0169] The above description is only a preferred embodiment of the present invention. All equivalent changes and modifications made within the scope of the claims of the present invention should be included in the scope of the present invention.

Claims

1. A method for synthesizing high-quality CT images using CBCT based on structural prior guidance, characterized in that: It includes the following steps: I. Patient Data Collection and Organization: Data from 200 patients were collected, including different stages, pathological types, tumor locations, and age groups. The data covered CBCT images and unpaired CT images from different patients. Gaussian noise data augmentation was used to expand the data to 3 times the original size to improve the model's generalization ability. Finally, the training set, test set, and validation set were divided in a 7:1:2 ratio to complete the construction of the image database. II. Structure-Prior Guided Conditional Diffusion Image Synthesis: 1) Contour extraction operation: The Canny edge detection operator is used to process the input CBCT image to extract the main boundary information of tissues and organs and generate structural contour map C; 2) Diffusion modeling operations: In the structure-guided image generation framework, the image generation process is based on a diffusion probability mechanism of forward diffusion-backward denoising symmetric modeling. A Markov chain progressively transforms Gaussian noise into the target image distribution in the latent space. The forward process is a predefined stochastic process, and its core objective is to transform any given original image... Convert into a noisy image that approximates a Gaussian distribution. As the starting point for sampling, this process adds a small amount of Gaussian noise to the current image at each step t, thereby gradually destroying its structure and texture, eventually approaching pure noise. Its probability transfer function is defined as follows: ； in, This represents the cumulative noise attenuation coefficient; 3) Conditional denoising network design: In each step of the diffusion modeling operation, a conditional denoising mechanism is designed. The conditional denoising network predicts noise components and gradually reconstructs the target image. It integrates the following three types of information for joint modeling: ①Time step information t: The dynamic control capability of the diffusion stage is introduced through the embedding mechanism, so that the network has the ability to perceive the stage at different denoising depths. ②Structural profile prior The contour map extracted from the input CBCT image serves as a structural guidance condition, guiding the network to focus on key anatomical regions and improving the geometric consistency of the reconstruction. 4) Training objectives: Optimize mean squared error loss: ； Through this mechanism, each diffusion sampling step is guided by a contour map, so that the generated image maintains the consistency of anatomical structure to the greatest extent during modality transformation; III. Spatial Guided Deformable Convolution: 1) Technical design of the space guidance module: From multiple locations in the encoder-decoder network of the diffusion model, the input features, i.e., the input feature map, are obtained. The system actively extracts spatial prior information related to image structure to generate a high-resolution guidance map, which is used to accurately identify spatial locations in the image where there are drastic structural changes, significant edges, or prominent anatomical boundaries. 2) Technical design of deformable sampling kernel generator: Based on the structural response information provided by the guidance map obtained in step 1), the actual receptive position of each convolution sampling point in space is dynamically adjusted to construct an adaptive deformation spatial sampling mechanism; the receptive field of this spatial sampling mechanism has position sensitivity and geometric matching capability in space. An adaptive displacement vector field is generated based on the structural response map generated by the spatial guidance module to precisely control the sampling position of each convolution point, enabling the convolution kernel to automatically adapt to different structural geometries. Specifically, this includes the following steps: ① Input guide diagram The structural response map generated by the spatial guidance module serves as the basis for offset prediction; ② Migration field prediction network: Building small CNN subnetworks ,right Perform mapping to output the offset vector field: ； Where K represents the number of sampling points, such as K=9 corresponding to a 3×3 convolution, with K two-dimensional offset vectors at each position; ③Sampling location reconstruction: For each basic sampling point The final sampling location is: ； This operation ensures that the convolution kernel has a unique sampling path at each pixel, enabling the construction of non-uniform adaptive receptive fields based on the image structure; ④ Bilinear interpolation: Since the offset can be a non-integer coordinate, bilinear interpolation is used to sample features at the misaligned positions: ； in Represents the nearest integer grid point. For interpolation weights; Through the above mechanism, the model can dynamically reconstruct the receptive field based on structural information, achieving geometric alignment perception and boundary-sensitive modeling. 3) Pay attention to the technical design of the weighted modulator: A spatial-channel attention mechanism is introduced at each deformable sampling point to further control the feature response intensity at each location. Specifically, this includes the following steps: ① Contextual semantic compression: For the input feature map Perform global average pooling (GAP) to obtain the semantic description vector: ； ②Attention-generating network: Construct two layers of nonlinear mapping to generate attention vectors for sampling points: ； in, For learnable parameters, Activated for ReLU For the Sigmoid function, note the vector. Each component represents the weighting coefficient of the corresponding variable sampling point; ③ Deformable convolution fusion output: Final output features: ； Each sampling point is weighted and then used to calculate the output features, thus forming a "structure-oriented response regulation". IV. Reasoning Process and Quality Assessment Using CBCT images as input, contour maps extracted using the Canny operator are used as a guide, and then the SCG-CT model is progressively sampled to generate pseudo-CT images. Evaluation metrics include: Mean absolute error, root mean square error, structural similarity, and peak signal-to-noise ratio; Then, the synthetic images were compared with real CT scans to verify the model's performance in terms of image fidelity and structural consistency.

2. The method for synthesizing high-quality CT images based on structure prior guidance in CBCT according to claim 1, characterized in that: The specific technical design approach for the space guidance module is as follows: ① Extract semantic saliency cues from the original feature map: First, a spatial attention path is constructed through a series of lightweight convolution operations (such as 3×3 convolution + BN + ReLU + 1×1 compressed convolution). This path is dedicated to mining information such as texture abrupt changes, edge intensity gradients and context-aware differences in the current input feature map. ② Construct a semantic structure response graph: The output is a single-channel feature map. The response value of each pixel represents the importance or "structural sensitivity" of that location in terms of spatial structure for subsequent modeling. The larger the value, the more complex the structure of that region or the more obvious the geometric abrupt change. ③ As a deformable kernel and a signal source for attention weight modulation: The generated guidance map is not only used to predict the offset coordinates of subsequent convolutional sampling kernels, i.e., the source of the "bias field" of deformable convolution, but also serves as the core reference for adjusting the response intensity in the attention mechanism, ensuring that the model learns key structural regions with high accuracy.

3. The method for synthesizing high-quality CT images based on structure prior guidance in CBCT according to claim 2, characterized in that: In step two, a conditional embedding residual block is added to the spatially guided deformable attentional convolution to generate the basic structure of the image model. Temporal embedding and structural and semantic information of the image need to be processed. Image information is processed by connecting the spatially guided deformable attentional convolution and regular convolution.

4. The method for synthesizing high-quality CT images based on structure prior guidance in CBCT according to claim 3, characterized in that: The conditional denoising network design also includes step ③, which is a feature extraction network. The feature extraction network consists of spatially guided deformable attentional convolution and conditional embedding residual blocks from step 2. It has dynamic sampling and regional response modulation capabilities and can adaptively capture structural boundaries and semantic key regions.

5. The method for synthesizing high-quality CT images based on structure prior guidance in CBCT according to claim 1, characterized in that: An explicit structural constraint mechanism is introduced into the diffusion generation mechanism. In each diffusion iteration step, the contour information extracted from the input CBCT image is used as the structural prior condition. As the image is gradually restored from Gaussian noise to the target CT modality, the generation path is continuously guided so that it always fits the real organ boundary and tissue morphology, preventing geometric distortion or anatomical distortion caused by modal drift, texture replacement, etc.

6. The method for synthesizing high-quality CT images based on structure prior guidance in CBCT according to claim 1, characterized in that: The forward process has a closed analytical solution, allowing for direct sampling from each step without explicit sampling during the training phase, using only a single weighted sampling. Get any .

Images