A diffusion type three-dimensional bone reconstruction method and system based on image distribution enhancement

Abstract

The application provides a diffusion type three-dimensional bone reconstruction method and system based on image distribution enhancement, and belongs to the technical field of medical image processing, and comprises the following steps: acquiring DRR image and X-ray image data and performing data preprocessing; asymmetric gray scale remapping and KAN network driven geometric deformation are respectively applied to the DRR image and the X-ray image, different disturbance coefficients are given to the geometrically deformed images by adopting an asymmetric disturbance strategy to obtain enhanced DRR images and X-ray images; the enhanced DRR images and X-ray images are subjected to double-domain multi-scale feature fusion to obtain conditional embedded features; and based on the conditional embedded features, an initial noise volume is denoised to a bone CT volume by adopting a conditional diffusion model, so that three-dimensional bone reconstruction is completed. The application realizes the enhancement of bone features and the suppression of soft tissue noise in three-dimensional reconstruction.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of medical image processing, and particularly relates to a diffusion type three-dimensional bone reconstruction method and system based on image distribution enhancement. BACKGROUND

[0002] The statements in this section merely provide background information related to the present application and do not necessarily constitute prior art.

[0003] Pediatric craniospinal irradiation (CSI) is a complex and highly precise radiotherapy technique, mainly used for the treatment of primary central nervous system tumors such as medulloblastoma and ependymoma. The entire course usually lasts for several days, and precise anatomical alignment is required every day to ensure accurate irradiation range and avoid damage to critical organs. Although a positioning CT is usually obtained on the first day of treatment, repeated CT scanning during treatment is not practical due to radiation dose and practical operation limitations, especially for pediatric patients. Therefore, daily positioning can only rely on low-dose orthogonal two-dimensional X-rays. However, X-ray images have limited structural details and cannot present complete three-dimensional anatomical information, which poses a great challenge to adaptive radiotherapy procedures that require volume information for accurate alignment.

[0004] The prior art uses a scheme to reconstruct CT-like three-dimensional bone anatomy from daily X-ray images, however, in pediatric CSI, 2D-to-3D reconstruction based on deep learning is hindered by the lack of real X-ray and CT paired training data. For this reason, most existing methods resort to using digitally reconstructed radiographs (DRRs) generated from CT volumes as a substitute. Although DRRs are strictly aligned with CT in structure, there are significant differences in noise characteristics, soft tissue visibility, and imaging artifacts, especially the differences in soft tissue performance, which further weaken the clinical generalization ability of the model trained with DRRs. SUMMARY

[0005] To overcome the shortcomings of the prior art, the present application proposes a diffusion type three-dimensional bone reconstruction method and system based on image distribution enhancement, which can learn under the mixed conditions of large-scale simulated DRRs and a small amount of real clinical data. The present application makes it possible to train a clinically usable model with limited real data, providing a feasible alternative solution for pediatric radiotherapy with data scarcity.

[0006] To achieve the above-mentioned purpose, one or more embodiments of the present application provide the following technical solutions:

[0007] In a first aspect, the present application discloses a diffusion type three-dimensional bone reconstruction method based on image distribution enhancement, comprising:

[0008] acquire DRR image and X-ray image data and perform data preprocessing;

[0009] respectively apply asymmetric gray scale remapping and KAN network driven geometric deformation to the DRR image and the X-ray image, and give different perturbation coefficients to the images after geometric deformation by using an asymmetric perturbation strategy to obtain enhanced DRR image and X-ray image;

[0010] perform double-domain multi-scale feature fusion on the enhanced DRR image and X-ray image to obtain conditional embedding features;

[0011] based on the conditional embedding features, use a conditional diffusion model to denoise an initial noise volume to a skeletal CT volume to complete three-dimensional skeletal reconstruction.

[0012] In a second aspect, the present application discloses a diffusion type three-dimensional skeletal reconstruction system based on image distribution enhancement, comprising:

[0013] a data acquisition module for acquiring DRR image and X-ray image data and performing data preprocessing;

[0014] a domain distribution perturbation module for respectively applying asymmetric gray scale remapping and KAN network driven geometric deformation to the DRR image and the X-ray image, and giving different perturbation coefficients to the images after geometric deformation by using an asymmetric perturbation strategy to obtain enhanced DRR image and X-ray image;

[0015] a double-domain multi-scale feature fusion module for performing double-domain multi-scale feature fusion on the enhanced DRR image and X-ray image to obtain conditional embedding features;

[0016] a three-dimensional reconstruction module for based on the conditional embedding features, using a conditional diffusion model to denoise an initial noise volume to a skeletal CT volume to complete three-dimensional skeletal reconstruction.

[0017] In a third aspect, the present application discloses an electronic device comprising a memory and a processor, and computer instructions stored in the memory and running on the processor, when the computer instructions are run by the processor, the steps of the above-mentioned diffusion type three-dimensional skeletal reconstruction method based on image distribution enhancement are completed.

[0018] In a fourth aspect, the present application discloses a computer readable storage medium for storing computer instructions, when the computer instructions are executed by the processor, the steps of the above-mentioned diffusion type three-dimensional skeletal reconstruction method based on image distribution enhancement are completed.

[0019] Compared with the prior art, the present application has the following beneficial effects:

[0020] The application proposes an asymmetric domain perturbation method to bridge the domain gap, enhance the bone features and suppress the soft tissue noise. By applying asymmetric gray scale remapping and KAN network driven geometric deformation to DRR and X-ray respectively, the domain gap between the synthetic domain and the real domain is narrowed, and the bone signal is strengthened and the soft tissue noise is suppressed. The network accepts various distortion scenarios in the training stage, and remains robust on real low dose images, significantly reducing the performance gap between the laboratory and the clinic.

[0021] The application constructs a dual-domain multi-scale feature fusion module, which realizes structure alignment and cross-modal semantic fusion, enabling the model to learn domain-invariant bone representations by focusing on local anatomical details and global context; cross-modal semantic feature alignment provides robust three-dimensional guidance under domain shift, enabling the prior knowledge of DRR and the appearance information of real X-ray to be deeply coupled in the same semantic space, ensuring that the reconstruction result has both anatomical consistency and clinical interpretability.

[0022] The application provides a clinically valuable real X-ray and CT paired dataset, which supports model training and verification closer to the clinical scene through real paired data, and helps the clinical effectiveness and reliability of the model.

[0023] The advantages of the additional aspects of the application will be partially given in the following description, partially become obvious from the following description, or be known by the practice of the application. BRIEF DESCRIPTION OF DRAWINGS

[0024] The drawings accompanying the specification of the application form part of the specification and serve to further illustrate the illustrative embodiments of the application and to explain the principles of the application, without in any way presenting an improper limitation of the application.

[0025] Figure 1 The flowchart of the diffusion type three-dimensional bone reconstruction method based on image distribution enhancement described in embodiment one of the application.

[0026] Figure 2 The architecture diagram of the diffusion type three-dimensional bone reconstruction method based on image distribution enhancement described in embodiment one of the application.

[0027] Figure 3 The training flowchart of the diffusion type three-dimensional bone reconstruction method based on image distribution enhancement described in embodiment one of the application.

[0028] Figure 4 The comparison diagram of the model head reconstruction result under low dose dual-view X-ray input and real CT described in embodiment one of the application; wherein, figure (a) is the head reconstruction result generated by the model prediction, and figure (b) is the real CT of the head.

[0029] Figure 5 Figure 6 is a comparison chart of the model chest reconstruction results and the real CT under the low-dose dual-view X-ray input described in embodiment one of the present application; wherein figure (a) is the chest reconstruction result generated by the model prediction, and figure (b) is the real CT of the chest.

[0030] Figure 6 Figure 7 is a comparison chart of the model waist reconstruction results and the real CT under the low-dose dual-view X-ray input described in embodiment one of the present application; wherein figure (a) is the waist reconstruction result generated by the model prediction, and figure (b) is the real CT of the waist. DETAILED DESCRIPTION

[0031] It should be noted that the following detailed description is exemplary in nature and is intended to provide further description of the present application. Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs.

[0032] It should be noted that the terms used herein are only for the purpose of describing specific embodiments and are not intended to limit the exemplary embodiments according to the present application.

[0033] In the case of no conflict, the embodiments in the present application and the features in the embodiments can be combined with each other.

[0034] Term explanation:

[0035] Three-dimensional reconstruction: Three-dimensional reconstruction is a multi-disciplinary technology that integrates computer vision, graphics and image processing. It processes data from two-dimensional images, laser scanning point clouds, CT / MRI tomographic images and other data from different perspectives through feature extraction, camera calibration, point cloud registration and fusion, surface reconstruction and texture mapping algorithms to recover the geometric structure, texture information and spatial position relationship of the object or scene in three-dimensional space, and finally generate a digital three-dimensional model with interactive and visual characteristics.

[0036] Diffusion model: Diffusion model is a deep learning technology based on probabilistic generative model. The core principle is to simulate the "diffusion" and "inverse diffusion" processes to learn the true distribution of data and generate high-quality new data. Its workflow consists of two steps: first, the forward diffusion process, in which small Gaussian noise is gradually added to the original data to make it evolve into a completely random noise distribution; then the inverse diffusion process, in which a neural network is trained to learn the rules of removing noise step by step from random noise, and finally the noise is restored to new samples with the same distribution as the original data. This model has precise capturing ability for data details and is widely used in image generation, medical image synthesis and other fields, and is one of the core technologies in the current generative artificial intelligence field.

[0037] Embodiment One

[0038] In one or more embodiments, a diffusion-based three-dimensional bone reconstruction method based on image distribution enhancement is disclosed, as shown, comprising the following steps: Figures 1-2

[0039] Step S1, obtain clinical low-dose anteroposterior X-ray and DRR image data and perform data preprocessing; scale the X-ray and DRR images to the same size and perform gray scale normalization.

[0040] In this embodiment, the input image size is set to 256x256, which is not limited, and the size of the reconstructed data and the training data is consistent.

[0041] Step S2, construct a domain distribution perturbation module, apply asymmetric gray scale remapping and KAN network driven geometric deformation to the DRR image and X-ray image respectively, and use an asymmetric perturbation strategy to give different perturbation coefficients to the geometrically deformed images to obtain enhanced DRR images and X-ray images, thereby narrowing the domain difference between the synthetic domain and the real domain, and strengthening the bone signal and suppressing soft tissue noise, to obtain

[0042] Step S2-1, use a non-parametric gray scale remapping module SR to randomly sample N control points to define a piecewise linear mapping on the normalized intensity value [-1, 1]:

[0043] (1)

[0044] wherein, is a piecewise linear mapping; is a gray scale remapping function; is an input image; N is the number of control points; is the i-th pixel point in the image; is an indicator function, which takes the value 1 when the input x belongs to the interval , otherwise 0.

[0045] This transformation compresses the soft tissue intensity range while enhancing the contrast of the bone structure. It introduces photometric diversity (e.g. detector sensitivity, exposure conditions) that reflects the variability of clinical acquisition, and guides the model to prefer high-contrast regions, suppress the influence of soft tissue, and enhance the saliency of bone.

[0046] Step S2-2, use a spatial transformation module that combines global affine and local non-rigid deformation to simulate the variability of bone geometric anatomy.

[0047] wherein the global affine part is defined as:​

[0048] (2)

[0049] where, are the coordinates after global affine; A and b represent the transformation matrix and translation parameters, covering the overall body difference.

[0050] Local non-rigid deformation learns a set of differentiable basis functions through KAN network to generate continuous deformation field:

[0051] (3)

[0052] where, is the continuous deformation field; is the learnable nonlinear basis function with adaptive smoothness; is the learnable weight coefficient for controlling the contribution of the ith basis function in the overall deformation field; is the bias term for shifting the input of the nonlinear basis function, enhancing the function expression ability; is the direction vector of input mapping, which determines the projection direction of input data in the ith basis function; T is the transpose; k is the number of basis functions, which determines the complexity of the deformation field. More basis functions mean that the model can express more complex and detailed non-rigid deformation; is the differentiable basis function.

[0053] The final composite transformation is obtained:

[0054] (4)

[0055] where, is the geometric deformation result; λ is the weight of adjusting the deformation. KAN can achieve flexible and anatomically reasonable perturbation while avoiding excessive deformation of non-skeletal regions.

[0056] Step S2-3, the strategy of asymmetric perturbation intensity setting is adopted, respectively, to assign different perturbation coefficients to the DRR image and the X-ray image, to apply stronger perturbation to the DRR image, and to perform more conservative enhancement to the real X-ray image, to maintain anatomical fidelity. The final perturbed image is obtained by first applying gray scale remapping, and then performing the composite geometric transformation of step 2-2 to distort the image:

[0057] （ 5)

[0058] (6)

[0059] where, is the image after asymmetric perturbation enhancement; for spatial warping of remapped image; x is the original input image; for perturbation intensity coefficient; for perturbation intensity coefficient applied on DRR image, aiming to introduce more variation and weaken the domain difference with X-ray; for perturbation intensity coefficient applied on Xray, aiming to keep the authenticity of anatomical structure and avoid over deformation, greater than ; for DRR image; for Xray image; and > , ensuring that the enhancement of DRR is stronger than that of X-ray.

[0060] Step S3, fusing the enhanced DRR image and the X-ray image through a dual-domain multi-scale feature fusion module to obtain conditional embedding features.

[0061] A dual-domain multi-scale feature fusion module is constructed. In order to bridge the semantic gap between DRR and X-ray input, a dual-domain multi-scale feature fusion module is constructed, which performs structure alignment from fine to coarse and cross-modal semantic fusion, specifically including:

[0062] Step S3-1, inputting the encoder to extract the preliminary feature map respectively, and dividing the preliminary feature map into non-overlapping small blocks at multiple scales to obtain local features at different levels.

[0063] For each input image (enhanced DRR or X-ray), a shared shallow encoder is used to extract a preliminary feature map. These feature maps are divided into non-overlapping small blocks at multiple scales to form hierarchical local features, which can be expressed as:

[0064] (7)

[0065] (8)

[0066] In the formula, is the feature representation of the enhanced DRR (source domain) at the scale; is the feature representation of the Xray (target domain) at the scale; E is a shallow encoder; and represent the enhanced DRR image and the X-ray image respectively; represents the feature scale, and , (1 / 2, 1 / 4, 1 / 8 resolution can be taken), is the total number of multi-scale resolutions, which determines the depth of the feature hierarchy.

[0067] Preferably, in order to enhance the granularity and discriminability of the extracted features, different scale pooling operations are added behind the convolutional layers of the encoder, and for each patch, multi-scale feature representations are generated by using different scale pooling operations. Different scale pooling operations include average pooling and max pooling; a multi-resolution, multi-view feature concatenation set is obtained for each patch, where average pooling provides global structure information and max pooling highlights key edge information, thereby jointly improving the granularity and discriminability of the skeletal features. The features captured by this operation range from fine skeletal edges to coarse anatomical layouts.

[0068] Step S3-2, the DRR domain and X-ray domain extracted feature blocks are spliced at each scale, which can be expressed as:

[0069] (9)

[0070] In the formula, is the DRR domain feature; is the Xray domain feature; is the splicing operation; is the cross-domain fusion feature at scale . is the multi-scale level index.

[0071] Step S3-3, the spliced features are transmitted through a transformer-based interaction module, which is used to capture cross-domain and location long-range dependencies, to obtain cross-domain fusion features, which can be expressed as:

[0072] (10)

[0073] In the formula, is the enhanced cross-domain feature at scale . is the cross-domain fusion feature at scale . is layer normalization; is the attention interaction module for global dependency modeling.

[0074] This interaction step is repeated at different scales, so that the model can align the homologous skeletal regions even under intensity or geometric displacement.

[0075] Step S3-4, the L different scale cross-domain fusion features are learnable weighted summed and projected into a compact conditional embedding feature, which is used to guide the latent diffusion model subsequently. The process can be expressed as:

[0076] (11)

[0077] wherein, is a conditional embedding feature; denotes the weight of each scale, which is a learnable parameter, and the sum of all weights is 1.

[0078] The embodiment enables the model to learn domain-invariant skeletal representations by jointly focusing on local anatomical details and global context through the dual-domain multi-scale feature fusion module.

[0079] Steps S2 and S3 are connected and gain from each other before and after the dual-domain multi-scale feature fusion module of the asymmetric domain distribution perturbation. Specifically, the asymmetric domain distribution perturbation applies strong perturbation to the synthetic DRR and slight perturbation to the X-ray through gray scale remapping and geometric deformation, thereby actively creating domain differences and simultaneously amplifying skeletal signals and suppressing soft tissue noise at the data level, providing the subsequent network with an input distribution with "bone priority" and a larger domain span; the dual-domain multi-scale feature fusion takes the Transformer cross-domain attention as the core, and aligns the perturbed DRR and X-ray at the patch-token level in both directions, so that the anatomical prior of the DRR and the clinical appearance of the X-ray complement and correct each other in the same latent space, generating a highly robust conditional embedding. After the two are connected in series, the diversity brought by the perturbation is fully utilized by the fusion module, significantly improving the domain generalization ability and structural fidelity of three-dimensional skeletal reconstruction in the pediatric low-dose scenario.

[0080] Step S4, based on the conditional embedding feature, uses a conditional diffusion model to denoise the initial noise volume to a skeletal CT volume to complete three-dimensional skeletal reconstruction.

[0081] Step S4-1, a pre-trained autoencoder ( , ) is used to compress the real CT volume into a latent representation, which is expressed as:

[0082] (12)

[0083] (13)

[0084] In the formula, denotes the encoder, denotes the decoder, denotes the compressed latent representation of the CT image, is the CT image, denotes the reconstructed CT image, is the completely noisy latent variable (the noisiest state).

[0085] Step S4-2, a standard denoising diffusion probability model DDPM is used in this latent space. In the forward diffusion process, the latent variable is updated as follows: The noise is added step by step, which is represented by the formula:

[0086] (14)

[0087] where, represents the variance table; is the identity matrix; is the Gaussian distribution; is the latent variable of the th step.

[0088] The reverse denoising process learns a series of conditional distributions , which gradually reconstruct the clean latent representation.

[0089] In order to inject semantic priors in the generative process, the fused conditional embeddings are integrated into the denoising network through a hierarchical cross-attention mechanism. At each time step t , the latent representation is projected into the query matrix, while the conditional context is projected into the key and value matrices, which are represented by the formula:

[0090] (15)

[0091] where Q is the query matrix; is the key matrix; is the value matrix; ) is the latent representation; is the conditional embedding; , and are trainable linear transformations used to generate Query, Key and Value in the attention mechanism, respectively. They are standard components of cross-attention modules.

[0092] The attention output is calculated as:

[0093] (16)

[0094] where, is the attention output; is the attention mechanism; σ is the softmax function, and d is the query key dimension. This mechanism enables the network to adaptively focus on anatomically relevant features during the denoising process, guided by the 2D input.

[0095] Step S4-3, integrate KAN into the diffusion backbone network based on 3D U-Net. Specifically, at the encoder-decoder junction of the deepest layer of U-Net, a KAN-MLP layer (feature projection + basis function expansion) is added. KAN provides a learnable activation function and functional basis expansion, further enhancing the representation ability of the denoising network, significantly improving the model's ability to capture anatomical continuity, geometric symmetry, and local curvature. This leads to finer bone structure preservation and smoother reconstruction of areas such as spinous processes and joint cavities.

[0096] Step S4-4, after T-step denoising, the final latent code The full-resolution CT image is obtained by decoding the decoder. It is expressed as:

[0097] (17)

[0098] In the formula, represents the reconstructed CT image, is the CT latent representation obtained by sampling and denoising the diffusion model, used to decode the final CT volume.

[0099] Further, the training process of the above model, as shown in Figure 3 , includes: step (1) data preprocessing and input, step (2) constructing domain distribution disturbance module, step (3) constructing dual-domain multi-scale feature fusion module, step (4) using conditional diffusion model for three-dimensional reconstruction, step (5) designing loss function to optimize the model.

[0100] The construction of the training set includes: constructing and preprocessing the two data sources required for training - large-scale simulated DRR and scarce clinical real X-ray / CT paired data, laying a data foundation for subsequent cross-domain reconstruction.

[0101] The embodiment designs the loss function. When the latent diffusion process learns to generate a reliable bone structure, a semantic alignment constraint is further introduced to ensure structural consistency between domains and enhance the anatomical fidelity of the reconstructed CT volume. Therefore, the total training objective includes three components: voxel-level reconstruction loss, structural feature alignment loss, and channel-level consistency loss. The specific implementation process is as follows:

[0102] The mean square error (MSE) is used to penalize the difference between the reconstructed CT volume and the real CT. Its definition is as follows:

[0103] (18)

[0104] In the formula, is the voxel-level reconstruction loss, is the real CT image, is the square of the Euclidean norm.

[0105] To enhance the semantic consistency between the original view and the augmented view, cosine similarity is used to align the structure embedding extracted from the encoder, which is defined as follows:

[0106] (19)

[0107] where, is the structure feature alignment loss, F and F' are the global feature vectors from the original image and the distribution perturbed image, respectively, is the Euclidean norm.

[0108] Considering that specific anatomical features often occupy different channels in the learned representation, a per-channel consistency loss is imposed on the aligned feature maps, which is defined as follows:

[0109] (20)

[0110] where, is the channel-level consistency loss; is the feature map of the DRR image in the c-th channel; is the feature map of the real X-ray image in the c-th channel; is the L1 norm, i.e., the sum of absolute differences at all pixel positions.

[0111] The total loss is calculated by weighting the voxel-level reconstruction loss, the structure feature alignment loss, and the channel-level consistency loss. The calculation formula of the total loss is as follows:

[0112] (21)

[0113] where, is the total loss function, and are hyperparameters set by grid search. In this embodiment = 0.4, = 0.1.

[0114] To verify the effectiveness of the present application, as shown in Table 1, the experimental results of the method proposed in this embodiment and other baseline models in the reconstruction cases of the head and torso are shown. The method of this embodiment is always superior to other methods in multiple evaluation indicators such as peak signal-to-noise ratio, structural similarity, and perceptual loss index score, indicating that it has better image quality and structural fidelity.

[0115] Table 1 Quantitative comparison results of different methods for head and torso

[0116]

[0117] For the structural similarity index SSIM, the method of the embodiment improves 3.29% and 5.05% compared with the baseline models X2CT and DISPR in the head region, and also has similar improvement (3.38%) in the trunk region, which shows that the overall and local skeletal morphology is better preserved. In terms of pixel-level accuracy, the method of the embodiment achieves the lowest mean absolute error MAE, which is reduced by 11.13%-16.19% compared with the baseline model X2CT and by 19.30% compared with the baseline model DISPR, which is especially beneficial to the complex anatomy of the head and trunk. The method of the embodiment also greatly improves the signal-to-noise ratio PSNR, and the signal-to-noise ratio of the head region is 4.28% and 8.71% higher than the baseline models X2CT and DISPR, respectively, and the signal-to-noise ratio of the trunk region is 6.47% and 4.96% higher than the baseline models X2CT and DISPR, respectively, which reflects effective texture preservation and noise suppression. The perceived quality measured by the LPIPS index further confirms the advantages of the method proposed in the embodiment, and the perceived quality of the head region is reduced by 24.48% and 41.58% compared with the baseline models X2CT and DISPR, respectively, and the perceived quality of the trunk region is increased by 30.46% and 15.38%, respectively, which shows stronger semantic consistency and visual preservation. Overall, although the baseline methods provide reasonable performance, they are limited in detail fidelity and robustness. In contrast, the method of the application integrates distribution enhancement, dual-domain fusion and alignment constraint, which significantly improves the reconstruction quality compared with other models.

[0118] As shown in Figures 4-6 , the three-dimensional skeletal volume reconstructed by the model in the head, chest and waist under low-dose dual-view X-ray input is respectively shown, and the comparison results with the corresponding real CT are shown. Among them, as shown in Figure 4 (a), the head reconstruction result generated by the model prediction is shown, as shown in Figure 2 (b), the head real CT is shown; as shown in Figure 5 (a), the chest reconstruction result generated by the model prediction is shown, as shown in Figure 5 (b), the chest real CT is shown; as shown in Figure 6 (a), the waist reconstruction result generated by the model prediction is shown, as shown in Figure 6 (b), the waist real CT is shown. As can be seen from the comparison chart, the cranial vault curvature, sagittal and coronal suture direction reconstructed by the method proposed in the application coincide with the real CT, the vertebral body height, spinous process and lamina edge in the spinal region are clear and continuous, the posterior rib and the rib-vertebral joint are complete and the thickness is uniform; the overall structure is smooth without artifacts, and the contrast levels of high-density cortical bone and low-density marrow cavity are still maintained under low-dose X-ray input, and the anatomical morphology is highly consistent with the gold standard.

[0119] In the skull region, the method of the present application accurately preserves the skull curvature and suture details while maintaining the integrity of the overall structure. In the spine, the method consistently reconstructs the vertebral body arrangement, spinous processes, and intervertebral boundaries with high fidelity, providing sufficient anatomical clarity for clinical interpretation and surgical planning. For the thoracic cavity, the method achieves superior rib continuity, accurate intercostal spacing, and good costovertebral joints, especially in the posterior ribs that are often obscured in standard X-ray images. Compared with traditional methods, the method of the present application is robust in restoring fine-grained skeletal morphology under low-dose imaging conditions. Overall, the method proposed in the present application provides anatomically consistent and clinically useful three-dimensional reconstruction, showing strong application potential in orthopedic diagnosis, radiotherapy positioning, and treatment planning.

[0120] Embodiment Two

[0121] In one or more embodiments, a diffusion-based three-dimensional bone reconstruction system based on image distribution enhancement is disclosed, specifically comprising:

[0122] A data acquisition module for acquiring DRR images and X-ray image data and performing data preprocessing;

[0123] A domain distribution perturbation module for applying asymmetric gray scale remapping and KAN network driven geometric deformation to the DRR images and X-ray images, respectively, and using an asymmetric perturbation strategy to assign different perturbation coefficients to the geometrically deformed images to obtain enhanced DRR images and X-ray images;

[0124] A dual-domain multi-scale feature fusion module for dual-domain multi-scale feature fusion of the enhanced DRR images and X-ray images to obtain conditional embedding features;

[0125] A three-dimensional reconstruction module for reducing noise in an initial noise volume to a skeletal CT volume based on the conditional embedding features using a conditional diffusion model to complete three-dimensional bone reconstruction.

[0126] Embodiment Three

[0127] The present embodiment provides an electronic device comprising a memory and a processor, and computer instructions stored on the memory and running on the processor, when the computer instructions are run by the processor, the steps of the above-mentioned diffusion-based three-dimensional bone reconstruction method based on image distribution enhancement are completed.

[0128] Embodiment Four

[0129] The present embodiment provides a computer readable storage medium for storing computer instructions, when the computer instructions are executed by the processor, the steps of the above-mentioned diffusion-based three-dimensional bone reconstruction method based on image distribution enhancement are completed.

[0130] The computer program instructions can also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer-implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flowchart block or blocks. Figure 1 one or more flowcharts and / or blocks Figure 1 means for functionally implementing the steps listed in the flowchart block or blocks.

[0131] These computer program instructions can also be stored in a computer readable memory that can direct a computer or other programmable data processing apparatus to function in a particular manner, such that the instructions stored in the computer readable memory produce an article of manufacture including instructions which implement the function specified in the flowchart block or blocks. Figure 1 one or more flowcharts and / or blocks Figure 1 means for functionally implementing the steps listed in the flowchart block or blocks.

[0132] The computer program instructions can also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer-implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flowchart block or blocks. Figure 1 one or more flowcharts and / or blocks Figure 1 means for functionally implementing the steps listed in the flowchart block or blocks.

[0133] The above description of the various embodiments can have emphasized certain aspects of the various embodiments, which description is not to be taken to mean that other aspects of the various embodiments are not equally valid and important. Those skilled in the art will readily observe that numerous modifications and alterations of the device and process can be made without departing from the scope of the present application.

[0134] The specific embodiments described herein are illustrative and not restrictive. Many variations of the described embodiments will become apparent to those of skill in the art upon reviewing the above description. The scope of the application should, therefore, be determined not with reference to the above description, but instead with reference to the appended claims, along with their full scope of equivalents.

Claims

1. A diffusion-based three-dimensional bone reconstruction method based on image distribution enhancement, characterized in that, The method comprises the following steps: acquire DRR image and X-ray image data and perform data preprocessing; apply asymmetric gray scale remapping and KAN network driven geometric deformation to the DRR image and X-ray image respectively, and give different perturbation coefficients to the geometrically deformed images by using an asymmetric perturbation strategy to obtain enhanced DRR image and X-ray image; perform double-domain multi-scale feature fusion on the enhanced DRR image and X-ray image to obtain conditional embedding features; based on the conditional embedding features, use a conditional diffusion model to denoise an initial noise volume to a skeletal CT volume to complete three-dimensional skeletal reconstruction.

2. The diffusion-based three-dimensional bone reconstruction method based on image distribution enhancement of claim 1, wherein, The gray scale remapping randomly samples N control points to perform piecewise linear mapping on normalized intensity values: wherein, is a piecewise linear mapping; is a gray scale remapping function; N is the number of control points; is a control point; is an input image; is an indicator function; is the i-th pixel in the image.

3. The diffusion-based three-dimensional bone reconstruction method using image distribution enhancement of claim 1, wherein, The KAN network driven geometric deformation uses a spatial transformation module combining global affine and local non-rigid deformation: In the formula, is the result of the geometric deformation; λ is the weight that regulates the deformation; A and b represent the transformation matrix and the translation parameters; are differentiable basis functions; are piecewise linear maps; is the input image.

4. The diffusion-based three-dimensional bone reconstruction method based on image distribution enhancement of claim 1, wherein, The asymmetric perturbation strategy gives different perturbation coefficients to DRRs and X-rays respectively, which is expressed as: wherein, is the image enhanced by asymmetric perturbation; is the spatial warping of the remapped image; x is the original input image; is the perturbation intensity coefficient; is the geometric deformation result; is the gray level remapping function; is the perturbation intensity coefficient applied on the DRR image; is the perturbation intensity coefficient applied on the Xray; is the DRR image; is the Xray image; and > .

5. The diffusion-based three-dimensional bone reconstruction method based on image distribution enhancement of claim 1, wherein, The double-domain multi-scale feature fusion comprises: input the encoder to extract preliminary feature maps, and divide the preliminary feature maps into non-overlapping small blocks at multiple scales to obtain local features at different levels, average pooling and maximum pooling operations are performed on each group of small blocks at multiple resolutions to obtain feature blocks; the feature blocks extracted from the DRR domain and the X-ray domain are spliced at each scale; the spliced features are transmitted through a transformer-based interaction module to obtain cross-domain fusion features; the cross-domain fusion features at different scales are weighted and summed to obtain conditional embedding features.

6. The diffusion-based three-dimensional bone reconstruction method based on image distribution enhancement of claim 5, wherein, The transformer-based interaction module is expressed as: wherein is the enhanced cross-domain feature at scale ; is the cross-domain fused feature at scale ; is layer normalization.

7. The diffusion-based three-dimensional bone reconstruction method using image distribution enhancement of claim 1, wherein, The loss function of the training process is calculated by a voxel-level reconstruction loss, a structure feature alignment loss, and a channel-level consistency loss to obtain a total loss wherein, is the total loss function, and are hyperparameters set by grid search; is the channel-level consistency loss; is the structure feature alignment loss; is the voxel-level reconstruction loss.

8. A diffusion-based three-dimensional bone reconstruction system based on image distribution enhancement, characterized by, The method comprises the following steps: a data acquisition module for acquiring DRR image and X-ray image data and performing data preprocessing; a domain distribution perturbation module for applying asymmetric gray scale remapping and KAN network driven geometric deformation to the DRR image and X-ray image respectively, and giving different perturbation coefficients to the geometrically deformed images by using an asymmetric perturbation strategy to obtain enhanced DRR image and X-ray image; a double-domain multi-scale feature fusion module for performing double-domain multi-scale feature fusion on the enhanced DRR image and X-ray image to obtain conditional embedding features; a three-dimensional reconstruction module for using a conditional diffusion model to denoise an initial noise volume to a skeletal CT volume based on the conditional embedding features to complete three-dimensional skeletal reconstruction.

9. An electronic device, comprising: a memory and a processor, and computer instructions stored in the memory and running on the processor, when the computer instructions are run by the processor, the method of claim 1-7 based on image distribution enhancement for diffusion type three-dimensional skeletal reconstruction is completed.

10. A computer-readable storage medium, characterized in that, a computer instruction storage, when the computer instruction is executed by the processor, the method of claim 1-7 based on image distribution enhancement for diffusion type three-dimensional skeletal reconstruction is completed.

Images